diff --git a/.github/workflows/build_and_publish.yml b/.github/workflows/build_and_publish.yml
index 7af5766..99ac9b3 100644
--- a/.github/workflows/build_and_publish.yml
+++ b/.github/workflows/build_and_publish.yml
@@ -11,7 +11,7 @@ on:
         default: all
 
 env:
-  DOCKERHUB_REPO: docker.io/kyuz0/amd-strix-halo-toolboxes
+  DOCKERHUB_REPO: gitea.wefers.page/julian/amd-strix-halo-toolboxes
   LOCAL_PREFIX: llama
 
 jobs:
@@ -63,11 +63,18 @@ jobs:
       - name: Check out repository
         uses: actions/checkout@v3
 
-      - name: Log in to Docker Hub
-        uses: docker/login-action@v2
+      - name: Cache podman storage for ${{ matrix.backend }}
+        uses: actions/cache@v5
         with:
-          username: ${{ secrets.DOCKERHUB_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_TOKEN }}
+          key: podman-storage-${{ matrix.backend }}
+          restore-keys: |
+            podman-storage-${{ matrix.backend }}
+            podman-storage
+          path: ~/.local/share/containers/storage
+
+      - name: Log in to Docker Hub
+        run: |
+          podman login -u ${{ secrets.DOCKERHUB_USERNAME}} -p ${{ secrets.DOCKERHUB_TOKEN }}
 
       - name: Set build timestamp
         run: echo "BUILD_TS=$(date +%Y%m%dT%H%M%S)" >> $GITHUB_ENV
@@ -86,12 +93,48 @@ jobs:
           CHN="${DOCKERHUB_REPO}:${NAME}"
 
           echo "→ Building ${DF}"
-          docker build --no-cache -t "${LI}" -f "${DF}" .
+
+          # we use buildah to eventually make use of pushing with
+          # zstd:chunked compression, which is much more efficient
+          # than dockers gzip format.
+          # --pull: ensure we use the latest version of the base image
+          # --squash: flatten the final result image into one single layer.
+          #           Avoids large image sizes due to intermediate files
+          #            that are irrelevant for the user
+          # --format oci: use the OCI image format, which allows for pushing with zstd:chunked
+          # --no-cache: Recompute every step in the dockerfile, even if the previous layer
+          #             has not ben invalidated. Needed since we pull from ze internet.
+          # --cache-(to|from): pull/push the intermedia cache layers resulting from
+          #                    --mount options in the Dockerfile
+          # NOTE: we are mounting cache layers for dnf and pushing them. This cache
+          #       layer is shared amongst all Dockerfiles, since they have the identical
+          #       mount parameter. When parallel building with buildah, those cache layers
+          #       compete. In parallel, they all pull the latest fitting cache, then maybe
+          #       add some packages relevant to their specific variant, then afterwards push
+          #       the cache again. When multiple buildahs push the dnf cache, they could invalidate
+          #       the just-pushed cache of another builder instance, so some packages might
+          #       always be missing. SOLUTION: we give each containers dnf cache an individual
+          #       id, thus cache per variant.
+          buildah bud \
+            --pull \
+            --squash \
+            --format oci \
+            --no-cache \
+            -t "${LI}" \
+            -f "${DF}" \
+            .
+
+          echo "→ Running smoke test..."
+          podman run --rm "${LI}" llama version
+          podman run --rm "${LI}" llama-cli --help || { status=$?; echo "llama-cli exited with status $status"; [[ $status -eq 0 || $status -eq 1 || $status -eq 134 ]]; }
+          podman run --rm "${LI}" llama-server --help || { status=$?; echo "llama-server exited with status $status"; [[ $status -eq 0 || $status -eq 1 || $status -eq 134 ]]; }
+
+          # push with zstd:chunked compression, see https://github.com/containers/storage/blob/main/docs/containers-storage-zstd-chunked.md
 
           echo "→ Tag & push immutable → ${IMM}"
-          docker tag "${LI}" "${IMM}"
-          docker push "${IMM}"
+          buildah tag "${LI}" "${IMM}"
+          buildah push --compression-format zstd:chunked "${IMM}"
 
           echo "→ Tag & push channel → ${CHN}"
-          docker tag "${IMM}" "${CHN}"
-          docker push "${CHN}"
+          buildah tag "${IMM}" "${CHN}"
+          buildah push --compression-format zstd:chunked "${CHN}"
diff --git a/.gitignore b/.gitignore
index ed8ebf5..88d0b24 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1 +1,2 @@
-__pycache__
\ No newline at end of file
+__pycache__
+research
\ No newline at end of file
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index 07bbca3..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           pp512 |         72.93 ± 0.06 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           tg128 |          2.95 ± 0.06 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index d3dcb44..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         16.57 ± 0.04 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.08 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index dc35095..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           pp512 |         71.58 ± 0.06 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           tg128 |          2.99 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 914ebcc..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         16.48 ± 0.03 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.07 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index b4213c1..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           pp512 |         79.51 ± 0.07 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           tg128 |          2.72 ± 0.10 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 86f7ce7..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         17.57 ± 0.05 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.15 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 2c8897a..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           pp512 |         79.24 ± 0.10 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           tg128 |          2.67 ± 0.11 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 77126a6..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         17.53 ± 0.03 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.15 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index 8e9d026..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           pp512 |         78.28 ± 0.06 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           tg128 |          2.98 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index b8f691a..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         17.18 ± 0.03 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.06 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index baf867e..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           pp512 |         80.59 ± 0.10 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           tg128 |          2.99 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 108f8a6..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         17.27 ± 0.02 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.07 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 3836dd7..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |           pp512 |         17.65 ± 0.01 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |           tg128 |          3.00 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index b2e6658..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |          4.94 ± 0.01 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          1.69 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index aeb25b4..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |           pp512 |        54.76 ± 11.46 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |           tg128 |          3.00 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 31ef62e..0000000
--- a/benchmark/results/04-02-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |          7.15 ± 0.02 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          2.27 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index 455d013..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |           pp512 |        454.95 ± 1.90 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |           tg128 |         22.26 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 71489ab..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         94.79 ± 0.56 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         16.48 ± 0.09 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index e9729d2..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |           pp512 |        425.21 ± 1.79 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |           tg128 |         23.41 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index cd27b5f..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         93.83 ± 0.40 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         16.55 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index d58dab9..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |           pp512 |        407.15 ± 2.05 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |           tg128 |         21.51 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index aa07720..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        101.09 ± 0.37 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         16.23 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index f69d3f6..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |           pp512 |        414.23 ± 2.09 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |           tg128 |         23.11 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 6b8851c..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        100.06 ± 0.38 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         15.97 ± 0.45 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index 5c3d51b..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |           pp512 |        489.62 ± 3.63 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |           tg128 |         23.40 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 718fbd4..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         92.48 ± 1.13 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         16.50 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 45c08f7..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |           pp512 |        425.86 ± 2.29 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |           tg128 |         23.41 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 9bd643e..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         92.06 ± 0.08 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         16.51 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index f8630c5..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |           pp512 |        106.42 ± 0.08 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |           tg128 |         10.87 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index a025866..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |          6.09 ± 0.00 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          8.28 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index f2c87cb..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |           pp512 |        333.10 ± 6.48 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |           tg128 |          9.51 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index ad53f7f..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |         78.99 ± 0.25 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          8.13 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__fa1.log
deleted file mode 100644
index 01e9227..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |           pp512 |        398.34 ± 1.32 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |           tg128 |         35.94 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index e85162c..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         90.22 ± 4.88 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         22.35 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index da85676..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |           pp512 |        398.87 ± 1.21 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |           tg128 |         36.09 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 40cfc38..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         92.13 ± 0.15 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         21.56 ± 1.34 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index a68e922..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |           pp512 |        947.86 ± 2.03 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |           tg128 |         33.77 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 11651a6..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         84.85 ± 1.04 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         21.89 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index d44f270..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |           pp512 |        952.84 ± 2.21 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |           tg128 |         35.23 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 3627139..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         84.01 ± 0.58 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         21.97 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log
deleted file mode 100644
index 35840f2..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |           pp512 |        983.72 ± 3.21 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |           tg128 |         36.20 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 7c3adf6..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         80.32 ± 1.28 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         22.31 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 3d7dec0..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |           pp512 |        955.10 ± 4.53 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |           tg128 |         36.16 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 8bd5686..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         81.34 ± 1.80 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         22.32 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 5754f88..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |           pp512 |        368.78 ± 0.17 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |           tg128 |         40.80 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 58b0878..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |          6.35 ± 0.00 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         18.75 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index 8c3838a..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |           pp512 |        877.18 ± 8.15 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |           tg128 |         40.07 ± 0.78 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index f135b33..0000000
--- a/benchmark/results/04-02-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |         90.27 ± 0.42 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         23.07 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index e023a01..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           pp512 |         48.83 ± 0.01 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           tg128 |          2.78 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 94fd676..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         29.25 ± 0.17 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.45 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 3a2ff58..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           pp512 |         49.38 ± 0.03 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           tg128 |          2.79 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 0c8dbc3..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         29.17 ± 0.18 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.46 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 238a9e8..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           pp512 |        146.04 ± 0.21 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           tg128 |          2.78 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 14bad93..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         36.22 ± 0.16 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.43 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index b6f9056..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           pp512 |        146.83 ± 0.25 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           tg128 |          2.78 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 0611423..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         36.40 ± 0.23 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.46 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index ac63a96..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           pp512 |        155.06 ± 0.11 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           tg128 |          2.79 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 1986253..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         38.36 ± 0.61 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.46 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index b16ded7..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           pp512 |        151.70 ± 0.21 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           tg128 |          2.78 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 5bd76eb..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         38.35 ± 0.67 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.46 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index a342daf..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |           pp512 |         21.74 ± 0.01 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |           tg128 |          2.81 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 23f10da..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |          8.35 ± 0.01 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          2.36 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 6907745..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |           pp512 |         99.39 ± 0.58 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |           tg128 |          2.76 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 36b4c25..0000000
--- a/benchmark/results/04-02-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |         11.79 ± 0.02 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          2.44 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1.log
deleted file mode 100644
index df4a892..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           pp512 |        800.17 ± 1.72 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           tg128 |          8.49 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index f184a50..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        169.18 ± 1.16 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          7.11 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 32590b1..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           pp512 |        803.22 ± 2.21 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           tg128 |          8.49 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 505df24..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        170.11 ± 0.81 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          7.11 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log
deleted file mode 100644
index 8118e71..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           pp512 |        764.18 ± 1.66 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           tg128 |          8.48 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index f3f55d3..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        166.22 ± 1.20 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          7.10 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 661b27c..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           pp512 |        766.68 ± 1.07 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           tg128 |          8.48 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 6cd94b5..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        164.84 ± 1.99 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          7.10 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log
deleted file mode 100644
index 1ef3f29..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           pp512 |        990.88 ± 3.15 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           tg128 |          8.50 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 5b40f9e..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        172.42 ± 3.61 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          7.10 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 545b269..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           pp512 |        799.71 ± 2.09 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           tg128 |          8.49 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 5974e95..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        170.19 ± 1.69 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          7.10 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 3c69e17..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |           pp512 |         19.70 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |           tg128 |          8.24 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index f5ad8c4..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |         16.69 ± 0.01 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          6.41 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log
deleted file mode 100644
index e04be5a..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |           pp512 |        222.01 ± 0.94 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |           tg128 |          7.59 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index d5938e7..0000000
--- a/benchmark/results/04-02-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |         76.47 ± 0.38 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          6.39 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__fa1.log
deleted file mode 100644
index 2325924..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           pp512 |        393.61 ± 2.94 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           tg128 |         42.58 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 47210cf..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        562.85 ± 0.47 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         40.59 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 266236d..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           pp512 |        388.54 ± 2.76 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           tg128 |         42.61 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 9c300e9..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        564.71 ± 0.81 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         40.60 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index 9da9fbc..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           pp512 |       1026.87 ± 6.06 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           tg128 |         41.90 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 6b04e21..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |       1042.36 ± 2.24 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         40.08 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 6463c3e..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           pp512 |       1027.41 ± 6.28 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           tg128 |         42.05 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 34cd5ba..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |       1038.86 ± 3.17 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         40.04 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log
deleted file mode 100644
index 34472dc..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           pp512 |       1070.15 ± 5.54 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           tg128 |         42.56 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 9a56129..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |       908.79 ± 27.38 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         40.91 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index eefc5b7..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           pp512 |       1038.67 ± 2.82 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           tg128 |         42.57 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 19b5211..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |       821.93 ± 29.40 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         40.92 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 21e6010..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |           pp512 |       676.59 ± 50.83 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |           tg128 |         47.22 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index b7d95f1..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        342.52 ± 0.46 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         35.25 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index 7ba7684..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |           pp512 |       951.76 ± 41.03 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |           tg128 |         46.68 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index c01a4ec..0000000
--- a/benchmark/results/04-02-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        799.39 ± 0.69 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         41.15 ± 0.06 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1.log
deleted file mode 100644
index 682b041..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           pp512 |        202.36 ± 3.50 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           tg128 |         15.80 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 9057c7d..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         41.36 ± 0.87 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          9.65 ± 0.30 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 366d2a3..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           pp512 |        200.10 ± 8.37 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           tg128 |         16.04 ± 0.06 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index cb8f1d5..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         41.53 ± 0.41 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          9.50 ± 0.75 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index a353c9d..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           pp512 |        205.05 ± 3.62 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           tg128 |         14.98 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index e89ce17..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         51.11 ± 0.62 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          9.63 ± 0.07 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 50fa18d..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           pp512 |        203.41 ± 3.52 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           tg128 |         15.00 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index e681c9d..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         51.19 ± 0.64 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          9.58 ± 0.10 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
deleted file mode 100644
index 803ea12..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           pp512 |        200.04 ± 4.11 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           tg128 |         16.19 ± 0.09 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 7bd5d26..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         41.42 ± 0.37 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         10.94 ± 0.42 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 291f2c8..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           pp512 |       197.48 ± 10.80 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           tg128 |         16.20 ± 0.08 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index a79a619..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         41.60 ± 0.36 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         10.81 ± 0.61 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 5264d6e..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |           pp512 |        119.82 ± 3.30 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |           tg128 |         17.75 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 8ecec15..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |         27.41 ± 0.01 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          3.42 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index 67b83cc..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |           pp512 |        133.28 ± 1.45 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |           tg128 |         15.98 ± 0.25 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 7945690..0000000
--- a/benchmark/results/04-02-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |         30.79 ± 0.06 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          6.50 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index f8f2cd5..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |        489.11 ± 2.88 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         27.18 ± 0.16 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index c30a4c0..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        214.97 ± 1.13 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         18.57 ± 1.40 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 3f8e597..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |        492.32 ± 2.55 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         27.23 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 23b0efd..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        207.64 ± 0.55 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         18.84 ± 0.97 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 4bf79a0..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |        576.03 ± 3.01 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         26.12 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index d1e5d2a..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        249.94 ± 1.13 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         19.18 ± 0.33 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 3d12444..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |        569.42 ± 8.52 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         27.07 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index fedbfed..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        250.24 ± 0.88 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         18.77 ± 0.98 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index a6ca809..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |        512.10 ± 4.69 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         27.27 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index c1c2630..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        216.18 ± 0.74 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         19.71 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 7e6c7c6..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |        493.72 ± 3.45 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         27.32 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index aec4b8d..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        209.02 ± 0.16 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         19.67 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 4f9d7a2..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |           pp512 |        168.95 ± 7.69 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |           tg128 |         10.62 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 69c594e..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |         75.04 ± 0.02 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          8.68 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index e9a9b4f..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |           pp512 |        351.97 ± 2.56 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |           tg128 |          9.42 ± 0.21 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index e5c1a75..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        127.67 ± 0.45 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          8.31 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1.log
deleted file mode 100644
index 0dd1297..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |        815.37 ± 5.82 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         58.54 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index cef8f0a..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        170.56 ± 4.38 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         31.30 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index c248812..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |        811.39 ± 6.56 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         58.57 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d87028a..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        171.54 ± 4.45 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         31.29 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index 73135a6..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |      1078.99 ± 11.01 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         56.45 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 0fa6c28..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        217.17 ± 8.71 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         30.94 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index e4c7e21..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |      1080.52 ± 10.73 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         57.49 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index bb64b06..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        218.42 ± 7.66 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         30.96 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1.log
deleted file mode 100644
index 75d0678..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |      1056.78 ± 36.08 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         59.15 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index fef8ac0..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        175.40 ± 4.11 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         31.98 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index e795bb5..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |      1039.16 ± 53.94 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         59.16 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index cacca70..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        174.67 ± 4.22 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         31.98 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 3e8cb37..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |           pp512 |       823.08 ± 48.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |           tg128 |         66.14 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 7e07c93..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        112.99 ± 0.13 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         27.35 ± 0.07 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index 9fd3aa4..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |           pp512 |      1064.73 ± 70.49 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |           tg128 |         68.93 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index f189e87..0000000
--- a/benchmark/results/04-02-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        152.30 ± 3.42 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         34.18 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1.log
deleted file mode 100644
index 33abf49..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |       1209.23 ± 7.46 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         71.48 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 1e0bd42..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        177.01 ± 5.01 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         34.40 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 3c417e5..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |       1207.91 ± 9.78 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         71.48 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d871650..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        175.56 ± 3.86 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         34.37 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log
deleted file mode 100644
index 22e9a0f..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |       1263.87 ± 7.23 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         68.78 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 0afdf48..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        222.20 ± 8.55 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         33.48 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 5e79f7a..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |       1260.69 ± 6.89 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         68.94 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index a96e13b..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        222.34 ± 7.55 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         33.52 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log
deleted file mode 100644
index 3b0cc22..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |      1241.85 ± 15.18 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         72.57 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 5027639..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        176.98 ± 4.28 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         35.43 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 70fc750..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |      1229.55 ± 20.23 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         72.45 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 0b7f117..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        176.47 ± 4.18 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         35.44 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log
deleted file mode 100644
index f27f9db..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |           pp512 |       846.24 ± 47.60 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |           tg128 |         86.32 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index c8693bb..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        114.18 ± 0.07 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         30.07 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log
deleted file mode 100644
index 4e7288a..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |           pp512 |       1005.90 ± 6.10 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |           tg128 |         79.55 ± 6.96 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 15abf33..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        153.83 ± 3.76 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         37.44 ± 0.05 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index f5ae063..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           pp512 |        193.67 ± 2.12 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           tg128 |         28.98 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index d5132fd..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        217.45 ± 0.19 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         23.87 ± 3.54 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index d64554b..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           pp512 |        263.91 ± 3.82 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           tg128 |         28.97 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 206af26..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        296.64 ± 0.50 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         23.44 ± 4.28 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index a915bc9..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           pp512 |        592.54 ± 4.38 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           tg128 |         27.45 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 99b98d0..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        449.68 ± 1.06 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         25.10 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 3d76bbf..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           pp512 |        592.83 ± 4.39 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           tg128 |         27.75 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index c1c013e..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        448.82 ± 1.02 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         25.07 ± 0.35 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index b89bb7b..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           pp512 |        282.60 ± 2.04 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           tg128 |         28.89 ± 0.20 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 5d5c0db..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        238.71 ± 0.62 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         23.93 ± 4.11 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index b8878a5..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           pp512 |        590.03 ± 3.05 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           tg128 |         28.73 ± 0.52 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index be1cd00..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        413.78 ± 0.61 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         23.54 ± 3.30 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 35f7add..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |           pp512 |        426.39 ± 3.26 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |           tg128 |         31.84 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 9ed3aa4..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        228.41 ± 1.50 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         22.47 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 19a2fa7..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |           pp512 |       509.22 ± 20.34 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |           tg128 |         29.92 ± 0.05 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index faa295c..0000000
--- a/benchmark/results/04-02-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |       370.94 ± 32.12 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         26.00 ± 0.20 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1.log
deleted file mode 100644
index a9d4fa7..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           pp512 |        323.33 ± 0.27 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           tg128 |         14.24 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index ee21601..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        232.79 ± 5.34 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         11.65 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 9ed3a0f..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           pp512 |        324.44 ± 0.31 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           tg128 |         14.24 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 45ba7e6..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        229.19 ± 6.79 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         11.66 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index 86e7334..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           pp512 |        936.69 ± 1.33 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           tg128 |         14.23 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 9364a43..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        258.34 ± 1.81 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         11.63 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 2eda1f4..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           pp512 |        935.37 ± 1.09 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           tg128 |         14.20 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 5f3c340..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        261.44 ± 5.27 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         11.62 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log
deleted file mode 100644
index fb87a4a..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           pp512 |        943.63 ± 1.62 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           tg128 |         14.25 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 14934cf..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |       396.59 ± 26.74 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         11.65 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index ba60108..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           pp512 |        942.52 ± 1.34 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           tg128 |         14.25 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 878aaa3..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        376.68 ± 9.34 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         11.65 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 55932ee..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |           pp512 |        125.50 ± 0.06 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |           tg128 |         14.45 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 41a43d8..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        111.11 ± 0.04 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         11.40 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index edac157..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |           pp512 |        687.05 ± 0.75 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |           tg128 |         14.14 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 24d4dab..0000000
--- a/benchmark/results/04-02-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |       376.92 ± 18.46 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         11.72 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index 1ff6f83..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           pp512 |        463.92 ± 1.19 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           tg128 |          4.02 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index b36f842..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        191.32 ± 3.30 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          3.68 ± 0.11 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index c643303..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           pp512 |        528.00 ± 0.44 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           tg128 |          4.02 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index cc34693..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        201.67 ± 1.78 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          3.74 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 00bcee5..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           pp512 |        508.08 ± 0.85 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           tg128 |          4.00 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 12a07a9..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        222.44 ± 2.25 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          3.72 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 97f96ee..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           pp512 |        508.48 ± 0.88 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           tg128 |          4.00 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 93404ec..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        220.03 ± 0.98 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          3.72 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index d8498c8..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           pp512 |        549.57 ± 2.42 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           tg128 |          4.02 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 08d310a..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        215.98 ± 0.94 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          3.73 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 59af47c..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           pp512 |        529.01 ± 0.98 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           tg128 |          4.02 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 359ca8d..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        204.92 ± 2.92 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          3.73 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index db92211..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |           pp512 |          9.32 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |           tg128 |          3.87 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 7a34429..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |          9.20 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          3.60 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 858bf96..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |           pp512 |        123.07 ± 0.27 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |           tg128 |          3.92 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 767256d..0000000
--- a/benchmark/results/04-02-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |         82.96 ± 0.72 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          3.66 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1.log
deleted file mode 100644
index c50aeb9..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           pp512 |      2870.77 ± 12.89 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           tg128 |         84.57 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 861d2dd..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |      1639.03 ± 15.14 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         61.51 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 7ec3e36..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           pp512 |      2807.93 ± 16.33 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           tg128 |         84.66 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index aba12ed..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |      1262.15 ± 24.34 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         61.54 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1.log
deleted file mode 100644
index 6b110c7..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           pp512 |       2891.85 ± 2.60 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           tg128 |         82.18 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 71df791..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |      1600.62 ± 30.98 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         60.21 ± 0.11 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 98f27b4..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           pp512 |       2893.75 ± 3.92 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           tg128 |         82.15 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d1578cd..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |      1419.18 ± 40.21 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         60.24 ± 0.11 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1.log
deleted file mode 100644
index fe1a281..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           pp512 |      2805.65 ± 13.25 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           tg128 |         85.35 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index ef479c4..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |      1215.66 ± 10.33 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         62.02 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index da4aac2..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           pp512 |      2800.57 ± 47.75 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           tg128 |         85.47 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index e240976..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |      1214.20 ± 13.26 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         62.03 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 6ebe735..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |           pp512 |        657.19 ± 0.41 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |           tg128 |         86.55 ± 0.10 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 404a158..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        493.70 ± 0.98 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         58.57 ± 0.11 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1.log
deleted file mode 100644
index 0537089..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |           pp512 |     1977.82 ± 204.87 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |           tg128 |         91.09 ± 3.96 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index e118216..0000000
--- a/benchmark/results/04-02-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |      1149.92 ± 30.21 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         67.86 ± 0.22 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1.log
deleted file mode 100644
index f098ce7..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           pp512 |        181.09 ± 1.36 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           tg128 |         51.77 ± 0.73 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 0d8b56a..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        217.33 ± 0.32 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         36.51 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index d6ccfa5..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           pp512 |        181.23 ± 1.39 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           tg128 |         52.06 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 9bfd24d..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        218.30 ± 0.10 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         34.08 ± 4.33 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index a05b582..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           pp512 |        681.58 ± 4.94 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           tg128 |         50.85 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 5948f34..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |       396.76 ± 35.92 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         40.13 ± 0.05 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index abd20b5..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           pp512 |        678.97 ± 4.29 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           tg128 |         51.88 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index e3aae6c..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |       400.64 ± 35.51 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         37.97 ± 3.34 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log
deleted file mode 100644
index 5365be2..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           pp512 |       649.28 ± 39.54 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           tg128 |         52.00 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index c611dc9..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        259.94 ± 5.81 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         40.30 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 83fdee3..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           pp512 |       666.65 ± 12.50 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           tg128 |         52.05 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 420ec5d..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        261.35 ± 6.47 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         40.36 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 4e6b4ce..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |           pp512 |       643.04 ± 39.69 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |           tg128 |         54.00 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 394417c..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        195.45 ± 2.65 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         37.02 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index 40087c2..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |           pp512 |        597.02 ± 9.82 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |           tg128 |         57.38 ± 0.04 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index e8acd38..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        278.37 ± 7.19 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         42.78 ± 0.09 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1.log
deleted file mode 100644
index 78c56b0..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           pp512 |        547.85 ± 6.58 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           tg128 |         73.52 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index e4ebb3b..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        402.32 ± 0.67 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         52.50 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index bb60d45..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           pp512 |        546.41 ± 6.71 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           tg128 |         73.52 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index f06a549..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        402.65 ± 1.50 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         52.58 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log
deleted file mode 100644
index 77a4b05..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           pp512 |      1779.88 ± 16.15 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           tg128 |         73.26 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 7323d53..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        583.30 ± 9.18 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         57.85 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 537f920..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           pp512 |      1785.44 ± 15.68 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           tg128 |         73.22 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 9599307..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |       600.15 ± 13.61 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         57.78 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log
deleted file mode 100644
index adb683c..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           pp512 |      1742.62 ± 12.05 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           tg128 |         73.51 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 46f28e1..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        428.95 ± 5.63 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         57.82 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index c06bd1f..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           pp512 |       1730.96 ± 9.70 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           tg128 |         73.53 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index e0f62d9..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        425.86 ± 3.58 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         57.90 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log
deleted file mode 100644
index fd7fc59..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           pp512 |      1300.97 ± 78.99 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           tg128 |         77.58 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 92c9d52..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        337.80 ± 4.40 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         53.06 ± 0.10 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log
deleted file mode 100644
index 5534eab..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           pp512 |      1397.71 ± 70.15 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           tg128 |         80.99 ± 0.06 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 9f1beb7..0000000
--- a/benchmark/results/04-02-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        416.91 ± 7.90 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |         60.56 ± 0.77 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm-7_2__fa1.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm-7_2__fa1.log
deleted file mode 100644
index 4aeb30d..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           pp512 |        549.58 ± 0.25 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           tg128 |         50.55 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm-7_2__fa1__longctx32768.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index a671c82..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        144.34 ± 1.03 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          5.60 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 8caec46..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           pp512 |        548.97 ± 0.14 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           tg128 |         50.52 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 259fde9..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        143.70 ± 0.31 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          5.60 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1.log
deleted file mode 100644
index d34295f..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           pp512 |       1597.02 ± 1.89 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           tg128 |         51.01 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 0f41bbb..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        193.62 ± 1.29 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          6.93 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index ed7cd83..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           pp512 |       1598.36 ± 1.08 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           tg128 |         51.01 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index e9332b6..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        191.18 ± 2.26 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          6.94 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log
deleted file mode 100644
index dd9ab75..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           pp512 |       1590.62 ± 1.92 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           tg128 |         51.17 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 8634ecc..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        210.19 ± 4.57 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          5.56 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 08f165c..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           pp512 |       1590.40 ± 2.71 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           tg128 |         51.21 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 355d6a7..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        216.78 ± 3.22 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          5.56 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log
deleted file mode 100644
index b7dc396..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |           pp512 |        349.95 ± 0.30 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |           tg128 |         56.00 ± 0.22 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 34caa8e..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        152.53 ± 0.11 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          9.29 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__vulkan_radv__fa1.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__vulkan_radv__fa1.log
deleted file mode 100644
index e89bb38..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |           pp512 |       1355.55 ± 2.34 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |           tg128 |         55.88 ± 0.13 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log b/benchmark/results/04-02-2026/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 2ca9650..0000000
--- a/benchmark/results/04-02-2026/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 | pp2048 @ d32768 |        246.20 ± 1.24 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |   tg32 @ d32768 |          8.76 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/04-02-2026/system_info.json b/benchmark/results/04-02-2026/system_info.json
deleted file mode 100644
index b7fac37..0000000
--- a/benchmark/results/04-02-2026/system_info.json
+++ /dev/null
@@ -1 +0,0 @@
-{"distro": "Fedora Linux 43 (Workstation Edition)", "kernel": "6.18.5-200.fc43.x86_64", "linux_firmware": "linux-firmware-20260110-1.fc43.noarch", "timestamp": "04 Feb 2026"}
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index 6fe8641..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           pp512 |         66.52 ± 7.27 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           tg128 |          2.76 ± 0.08 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index f2148c0..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         16.57 ± 0.05 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.08 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 648542a..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           pp512 |         71.47 ± 0.20 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |           tg128 |          2.82 ± 0.07 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 2892f38..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         16.47 ± 0.03 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.06 ± 0.02 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 665e557..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         81.03 ± 0.03 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.07 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 19a7c47..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         17.37 ± 0.00 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.15 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 5b9f63c..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         79.43 ± 0.03 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.80 ± 0.05 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 623b913..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         17.42 ± 0.00 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.15 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index a56cbfc..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         73.64 ± 0.05 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.74 ± 0.22 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index d33cd4c..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,18 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:96: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f1a3468b5a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f1a3468b96b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f1a3468baef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4e882) [0x7f1a37496882]
-/usr/local/lib64/libggml-hip.so.0(+0x2d53c4e) [0x7f1a3749bc4e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f1a346a2e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f1a37b75630]
-/usr/local/bin/llama-bench() [0x40ae7c]
-/usr/local/bin/llama-bench() [0x408bd1]
-/lib64/libc.so.6(+0x35b5) [0x7f1a340215b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f1a34021668]
-/usr/local/bin/llama-bench() [0x409cf5]
-✖ ! [rocm-7alpha] Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__fa1 __longctx32768 failed (exit 0)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index d353fd9..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         72.94 ± 2.79 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.16 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index e87a569..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         17.14 ± 0.00 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.07 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1.log
deleted file mode 100644
index 8b12325..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         77.29 ± 5.81 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.83 ± 0.04 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index d908ae5..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         17.54 ± 0.00 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.00 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index b7571ea..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         76.84 ± 4.54 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.84 ± 0.03 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 8d34cbb..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         17.30 ± 0.00 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.00 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 3690a51..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 |           pp512 |         41.19 ± 7.76 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 |           tg128 |          1.87 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index fd57886..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         10.37 ± 0.00 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          1.26 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 67ff2d8..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 |           pp512 |         47.53 ± 0.02 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 |           tg128 |          2.96 ± 0.05 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 4b28cb9..0000000
--- a/benchmark/results/09-01-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         12.50 ± 0.00 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          2.27 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index f3637a5..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |           pp512 |        133.48 ± 0.41 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |           tg128 |         22.52 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index c5e6271..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         34.18 ± 0.22 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          8.12 ± 0.14 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index cb1d652..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |           pp512 |        181.28 ± 1.15 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |           tg128 |         22.65 ± 0.06 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 943b26b..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         35.28 ± 0.24 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          7.97 ± 0.43 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index ba953ff..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        279.35 ± 0.90 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.56 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 9436551..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         37.23 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          9.31 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 82410db..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        307.88 ± 1.76 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.76 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 148c3f1..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         38.53 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          9.32 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index af2dbbf..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        190.14 ± 0.23 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.71 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index c3d2826..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         36.33 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.11 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index d6ec8a8..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        235.84 ± 0.85 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.71 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 4f4003f..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         38.02 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.14 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1.log
deleted file mode 100644
index f9ea3e5..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        279.68 ± 1.30 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.83 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index 91c56ad..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         37.93 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.99 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 083d2b2..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        305.29 ± 1.90 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.83 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index ea1b0e2..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         38.08 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.00 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index b1f3f04..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           pp512 |        320.89 ± 0.75 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |          2.37 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 6adb97a..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         23.20 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          1.80 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 2f9b902..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           pp512 |        281.21 ± 0.80 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         25.02 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 88db148..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         34.18 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.41 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7_2__fa1.log
deleted file mode 100644
index 6d916e4..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |           pp512 |         56.12 ± 0.13 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |           tg128 |         16.60 ± 0.04 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 4c9a183..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         22.87 ± 0.07 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          6.89 ± 0.67 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 79e4936..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |           pp512 |        105.50 ± 0.54 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |           tg128 |         16.65 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 6020dcf..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         32.93 ± 0.13 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          7.24 ± 0.04 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index b038542..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        151.37 ± 0.24 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.55 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index aaba1d0..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         31.21 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.25 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 2352a80..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        195.36 ± 1.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.61 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index cf289f9..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         36.34 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.25 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
deleted file mode 100644
index 71b4a38..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |         88.58 ± 0.11 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.68 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 9601a3b..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         30.36 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.24 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index c43eada..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        152.92 ± 0.24 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.69 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index cd24ef1..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         35.26 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.23 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__fa1.log
deleted file mode 100644
index a73c174..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        146.64 ± 0.35 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.60 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index 603c9c5..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         34.21 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.19 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 6c10da2..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        188.13 ± 0.15 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.61 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index ec666d0..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         36.05 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.20 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index b242880..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           pp512 |        290.49 ± 0.30 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         17.74 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 078fc28..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         22.76 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          5.26 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index ec49e8c..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           pp512 |        261.76 ± 0.99 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         17.93 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index ae3dc57..0000000
--- a/benchmark/results/09-01-2026/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         33.30 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          8.69 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index eb98cfd..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           pp512 |         49.50 ± 0.01 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           tg128 |          2.78 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 6c2824c..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         27.87 ± 0.75 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.42 ± 0.06 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 1533579..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           pp512 |         49.27 ± 0.02 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |           tg128 |          2.79 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index c15e681..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         28.46 ± 0.36 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          2.44 ± 0.04 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index ce0f9af..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        148.26 ± 0.07 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.78 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 7b5211d..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         34.54 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.46 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 44dfe3c..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        147.21 ± 0.14 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.78 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 035ac3b..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         34.82 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.46 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index c7be012..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |         67.05 ± 0.01 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index e74c380..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         26.90 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.46 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 05a9c21..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |         66.64 ± 0.02 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 4a65754..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         26.86 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.46 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1.log
deleted file mode 100644
index 426d6b7..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        148.44 ± 0.08 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index 086c92f..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         32.66 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.46 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 6898f44..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        146.61 ± 0.04 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index a60c480..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         32.99 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.46 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 00c2046..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 |           pp512 |        100.89 ± 0.24 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 |           tg128 |          2.81 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index a1e97ab..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         18.12 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          2.16 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 1f40e8f..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 |           pp512 |         87.66 ± 0.55 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 |           tg128 |          2.78 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 4832afb..0000000
--- a/benchmark/results/09-01-2026/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         21.96 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          2.39 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1.log
deleted file mode 100644
index 921c345..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           pp512 |        797.13 ± 2.39 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           tg128 |          8.49 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 9dfec45..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        146.47 ± 5.52 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          7.11 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 0832956..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           pp512 |        803.39 ± 2.22 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |           tg128 |          8.49 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 91e85e1..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        152.56 ± 6.51 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          7.11 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log
deleted file mode 100644
index c87eda5..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        801.73 ± 2.77 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.49 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index d252ff5..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        163.31 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.10 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 0a3cf91..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        805.52 ± 3.18 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.48 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 1f87ce0..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        164.32 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.10 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log
deleted file mode 100644
index b03e29d..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        798.60 ± 3.84 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.49 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 9a0fea9..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        153.77 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.10 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 09d46a4..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        799.84 ± 4.89 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.49 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 6ff28f3..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        159.82 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.11 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index 90224a5..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        156.32 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.11 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 612fcc2..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        803.71 ± 3.13 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.50 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 3c53158..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        163.31 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.11 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log
deleted file mode 100644
index e5b1efe..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           pp512 |       187.83 ± 22.96 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           tg128 |          8.19 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index d2ca7d7..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         64.52 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          5.69 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log
deleted file mode 100644
index 784b23b..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           pp512 |        195.84 ± 0.06 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           tg128 |          7.56 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 5e0335b..0000000
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         75.42 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          6.23 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__fa1.log
deleted file mode 100644
index a25472a..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           pp512 |        393.57 ± 2.37 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           tg128 |         41.69 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 7ccfdf7..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        559.16 ± 0.99 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         39.74 ± 0.02 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index eaea7c6..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           pp512 |        389.10 ± 3.02 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |           tg128 |         41.68 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d0ae54b..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        558.52 ± 1.33 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         39.73 ± 0.02 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index 4aa26ac..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1065.39 ± 1.75 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.12 ± 0.02 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 504249f..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        823.17 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         38.90 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 50abf7d..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1034.18 ± 3.12 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.08 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index fab837f..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        896.75 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         38.88 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log
deleted file mode 100644
index 12cfa4e..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |        567.35 ± 4.92 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.67 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index ec34bc3..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        660.41 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         39.42 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 2433767..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |        560.67 ± 3.15 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.63 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 9e1315b..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        663.35 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         39.44 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__fa1.log
deleted file mode 100644
index f038039..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |      1051.12 ± 10.25 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.47 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index 39204d6..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        704.97 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         39.02 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index da08d9b..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |      1028.01 ± 11.24 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.37 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 3b1d0a7..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        743.16 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         39.05 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 64ecbc8..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 |           pp512 |      1253.52 ± 10.26 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 |           tg128 |         47.03 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index bae8a21..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        408.37 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         34.93 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index aa1388b..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 |           pp512 |      1016.39 ± 35.31 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 |           tg128 |         46.53 ± 0.03 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 4426e2b..0000000
--- a/benchmark/results/09-01-2026/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        403.09 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         40.91 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1.log
deleted file mode 100644
index 8ce6864..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           pp512 |       178.32 ± 26.83 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           tg128 |         12.97 ± 0.98 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 9777216..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         38.11 ± 0.36 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          9.90 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index ce3e316..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           pp512 |       167.63 ± 28.76 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |           tg128 |         12.73 ± 0.74 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 929e46f..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |         37.92 ± 0.33 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          9.78 ± 0.20 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index 2d23b75..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        206.60 ± 0.55 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.93 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index a8041b1..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         47.83 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          9.71 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 3e46d51..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |       197.06 ± 14.56 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.02 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 8658106..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         47.67 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          9.70 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
deleted file mode 100644
index d0be143..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        202.07 ± 3.84 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.09 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 6ef3440..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         38.42 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         10.00 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 7607a8d..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |       188.66 ± 20.66 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         13.61 ± 1.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 8a32625..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         38.43 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         10.01 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__fa1.log
deleted file mode 100644
index 771dfa7..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |       187.93 ± 19.38 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         12.70 ± 1.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index d353746..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,6 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-Hip error: 'unspecified launch failure'(719) at /longer_pathname_so_that_rpms_can_support_packaging_the_debug_info_for_all_os_profiles/src/rocm-libraries/projects/hipblaslt/library/src/amd_detail/hipblaslt.cpp:148
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-✖ ! [rocm7.1.1] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__fa1 __longctx32768 failed (exit 0)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 3e315d4..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,7 +0,0 @@
-ggml_cuda_init: failed to initialize ROCm: no ROCm-capable device is detected
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |         17.92 ± 2.98 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |          7.99 ± 0.19 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 7ec3d7e..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         57.28 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          9.55 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 74cc3d7..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |           pp512 |        168.14 ± 0.52 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |           tg128 |          2.08 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index e62c799..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         17.62 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          1.39 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index 382b3ea..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |           pp512 |        146.89 ± 0.98 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |           tg128 |         18.09 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 23324eb..0000000
--- a/benchmark/results/09-01-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         22.66 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          6.52 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index 3aa8493..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |       484.09 ± 10.61 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         27.12 ± 0.16 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 4f16aaf..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        162.38 ± 4.20 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         18.71 ± 1.16 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 50fd7e6..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |        492.54 ± 2.48 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         27.09 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 3bcfa1d..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        161.72 ± 4.90 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         19.36 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 96acea1..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        576.12 ± 2.09 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.83 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index c06be6a..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        254.34 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         19.13 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 7809d51..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        585.67 ± 2.54 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.83 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index b898478..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        244.89 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         19.18 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index 050cd71..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        492.51 ± 1.28 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.04 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 80b86d3..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        203.91 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         19.28 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index acf2f5f..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        494.46 ± 2.69 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.13 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 736f285..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        173.11 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         19.24 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__fa1.log
deleted file mode 100644
index 2b9fd5f..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        469.57 ± 3.31 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.97 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index bc82b1f..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        173.18 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         19.22 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index e4e32a8..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        471.12 ± 8.43 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.02 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 699cf1d..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        172.54 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         19.16 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 5c4f9fe..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |        424.44 ± 1.61 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         10.62 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 3d61772..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         65.51 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          8.05 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 33f57b0..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |        364.62 ± 2.62 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |          9.49 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 9b6a21f..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         93.65 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          8.14 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1.log
deleted file mode 100644
index a326a57..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |        813.78 ± 5.52 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         58.57 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index a38fcee..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        154.84 ± 3.34 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         31.32 ± 0.02 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index dba8fa9..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |       789.10 ± 47.98 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         58.51 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 49dce2c..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        155.23 ± 3.28 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         31.27 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index 6b3826e..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1073.10 ± 11.76 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.78 ± 0.02 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index c8a6856..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        206.02 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         31.04 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 3e19bb9..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1075.09 ± 15.15 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.72 ± 0.02 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d3cb0bc..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        204.43 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         31.02 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1.log
deleted file mode 100644
index efb6752..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       904.96 ± 12.42 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         58.50 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 1446b4a..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        158.93 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         31.07 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 2b1dc75..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        906.79 ± 8.48 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         58.55 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 6e97883..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        158.87 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         31.05 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__fa1.log
deleted file mode 100644
index 70d5d5b..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       999.37 ± 14.29 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.24 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index e11c921..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        166.76 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         30.57 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index f3ee322..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1004.04 ± 12.55 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.32 ± 0.04 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index fa388a8..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        168.58 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         30.68 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 01df500..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1180.84 ± 8.60 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         66.24 ± 0.04 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 00d42b3..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         71.45 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         21.82 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index 7e7199a..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1046.73 ± 6.25 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         68.71 ± 0.14 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index edb34ce..0000000
--- a/benchmark/results/09-01-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        109.86 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         30.94 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1.log
deleted file mode 100644
index 06add3d..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |       1207.32 ± 7.42 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         71.51 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 3022aa0..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        161.97 ± 3.98 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         34.41 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 101bd0a..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           pp512 |     1096.05 ± 129.46 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |           tg128 |         71.57 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 54589c8..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        160.95 ± 3.41 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         34.37 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log
deleted file mode 100644
index a7ccbd8..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1240.19 ± 1.93 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         69.06 ± 0.02 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 4178bdc..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        286.57 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         33.33 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 096ba0e..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1246.06 ± 12.57 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         68.95 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index ae4c172..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        211.86 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         33.37 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log
deleted file mode 100644
index 1f7c7dd..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1225.75 ± 5.62 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         71.54 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index ac7f41b..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        163.98 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         34.13 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 243d803..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1228.38 ± 14.75 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         71.53 ± 0.03 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 097c8cf..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        165.67 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         34.01 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__fa1.log
deleted file mode 100644
index 4051986..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1224.80 ± 12.64 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         69.07 ± 0.04 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index 23ad554..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        171.30 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         33.02 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index ede00e6..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1228.44 ± 13.05 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         68.99 ± 0.03 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index e3dd511..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        172.85 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         33.05 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log
deleted file mode 100644
index df1dc18..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |     1072.21 ± 149.58 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |          1.52 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 8a11d0b..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         71.87 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          1.45 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log
deleted file mode 100644
index 2539f39..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |      1075.31 ± 42.44 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         86.69 ± 0.11 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 6079ff7..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        111.06 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         33.81 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index fd2dcb3..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           pp512 |        211.96 ± 2.48 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           tg128 |         28.74 ± 0.49 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index e9297f0..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        239.82 ± 0.51 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         24.76 ± 1.87 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index fe3532d..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           pp512 |        262.57 ± 3.77 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |           tg128 |         28.94 ± 0.05 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d6b3f2b..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        295.41 ± 0.37 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         23.85 ± 3.58 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 1e7106e..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        587.41 ± 3.59 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         28.12 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index f0ae3e2..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        421.06 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         25.55 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 41593f2..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        610.91 ± 4.82 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         28.22 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 27251b7..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        432.47 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         25.56 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index 0b88bef..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        354.31 ± 5.52 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         29.40 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 2b3ed00..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        327.85 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         26.06 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 3ea4bf5..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        364.74 ± 5.05 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         29.38 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index f3c68f2..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        340.53 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         26.05 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1.log
deleted file mode 100644
index e3d43ae..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        586.56 ± 7.78 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         28.80 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index 2d23fe7..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        403.28 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         25.53 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 058efe5..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        609.81 ± 7.04 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         28.84 ± 0.02 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index ae25dd6..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        411.40 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         25.54 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index c0635f7..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |        634.07 ± 4.20 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         33.94 ± 0.02 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 9e31a3e..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        121.89 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         22.94 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index a95a432..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |       538.47 ± 29.53 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         31.56 ± 0.11 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index c355291..0000000
--- a/benchmark/results/09-01-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        211.76 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         27.44 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1.log
deleted file mode 100644
index a2749f2..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           pp512 |        323.22 ± 0.21 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           tg128 |         14.24 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 9f14bc9..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        231.95 ± 3.74 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         11.66 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 5cc6d4b..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           pp512 |        324.04 ± 0.14 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |           tg128 |         14.25 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d9d363a..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        230.54 ± 3.36 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         11.66 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index ca8d844..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        931.79 ± 1.30 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.20 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index f18870e..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        247.33 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         11.61 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 7f55d20..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        936.67 ± 1.30 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.20 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 3566962..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        259.06 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         11.63 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log
deleted file mode 100644
index c78e668..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        428.84 ± 1.18 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.24 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 5337922..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        275.24 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         11.63 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 7e1d970..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        431.06 ± 0.61 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.24 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 989d957..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        283.40 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         11.63 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__fa1.log
deleted file mode 100644
index 8c947c9..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        926.89 ± 0.25 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.25 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index a688c0f..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        262.69 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         11.66 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index a9ef8e3..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        934.30 ± 0.96 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.26 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 3a0e24a..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        254.81 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         11.67 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 1c332ff..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |           pp512 |       602.68 ± 80.42 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |           tg128 |         14.56 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index aa3f7c3..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         21.40 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         11.97 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index 8bb191e..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |           pp512 |        548.18 ± 1.59 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |           tg128 |         13.94 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 7cec059..0000000
--- a/benchmark/results/09-01-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        231.70 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.11 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index e6e1d62..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           pp512 |        430.73 ± 1.02 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           tg128 |          3.86 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index f7bcf1c..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        193.46 ± 0.50 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          3.69 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 02273a2..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           pp512 |        525.55 ± 1.67 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |           tg128 |          4.00 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 76656be..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        202.49 ± 3.52 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          3.72 ± 0.02 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 24ad07c..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        472.15 ± 0.56 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.00 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index a72de4d..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        188.56 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.72 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 302fb73..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        531.41 ± 1.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.00 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d754100..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        214.27 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.72 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index 1c6842e..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        470.21 ± 1.24 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.01 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index c14e74b..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        179.14 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.73 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index efdae4b..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        526.32 ± 1.23 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.02 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index a4539a4..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        193.22 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.73 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__fa1.log
deleted file mode 100644
index e9c1e83..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        467.63 ± 1.25 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.04 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index d71116c..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        182.27 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.74 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index ffbc638..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        527.37 ± 1.47 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.03 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d9ea532..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        201.00 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.74 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__fa1.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__fa1.log
deleted file mode 100644
index 68f29ba..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,2 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index ef8b29d..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 |           pp512 |       111.81 ± 20.34 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 |           tg128 |          3.85 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index e6b8559..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         73.77 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          3.40 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index e1a2ff4..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 |           pp512 |        107.40 ± 0.65 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 |           tg128 |          3.92 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index f1e6c00..0000000
--- a/benchmark/results/09-01-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         64.09 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          3.67 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1.log
deleted file mode 100644
index 83da529..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           pp512 |      2858.32 ± 17.99 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           tg128 |         84.57 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 5def9e8..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |      1182.57 ± 31.53 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         61.59 ± 0.02 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 455b875..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           pp512 |     2679.11 ± 228.92 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |           tg128 |         70.08 ± 4.20 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index ebaaee2..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |      1138.90 ± 19.11 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         61.59 ± 0.04 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1.log
deleted file mode 100644
index c102746..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2884.56 ± 5.24 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         80.80 ± 0.03 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index eeeb59c..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1446.85 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         59.42 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 2a53d30..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2874.72 ± 3.55 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         80.97 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index a78c350..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1258.46 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         59.59 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1.log
deleted file mode 100644
index e5cb703..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |      2829.05 ± 14.01 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         82.17 ± 4.20 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 421c142..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1118.35 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         61.04 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index e8c43cb..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |      2843.10 ± 21.02 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         84.76 ± 0.02 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 143a699..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1123.24 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         61.04 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7.1.1__fa1.log
deleted file mode 100644
index 8431abc..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |     2657.05 ± 331.16 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         62.00 ± 3.43 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index 9fb395a..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1124.76 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         59.78 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index dbf2273..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |      2818.77 ± 65.80 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         65.67 ± 4.85 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index cc2947e..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1136.22 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         59.85 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 6dcd777..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |           pp512 |     1514.96 ± 340.21 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |           tg128 |         81.61 ± 2.29 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index a3d57b9..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        188.74 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         66.83 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1.log
deleted file mode 100644
index ac9a8fa..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |           pp512 |     1235.50 ± 244.41 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |           tg128 |         81.02 ± 2.09 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 56ae086..0000000
--- a/benchmark/results/09-01-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        694.43 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         49.80 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1.log
deleted file mode 100644
index 2299f09..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           pp512 |        181.34 ± 1.51 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           tg128 |         52.25 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 6275774..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        217.76 ± 0.23 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         36.60 ± 0.02 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index c90c1af..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           pp512 |        181.33 ± 1.42 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           tg128 |         52.27 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index c7a144b..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        215.89 ± 4.16 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         36.64 ± 0.02 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index 3553f04..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        683.09 ± 7.89 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.50 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 64eaca1..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        334.72 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         40.07 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 78c2f6b..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        684.21 ± 8.30 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.99 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 6b62e44..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        333.73 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         40.14 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log
deleted file mode 100644
index a0cb708..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           pp512 |        674.34 ± 3.87 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           tg128 |         52.10 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 5cd16ba..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        352.35 ± 0.56 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         40.44 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 8898523..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           pp512 |        679.62 ± 5.41 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |           tg128 |         52.11 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 75ba15a..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        352.62 ± 1.18 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         40.46 ± 0.03 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__fa1.log
deleted file mode 100644
index b917599..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        675.95 ± 6.90 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.15 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index 17ea910..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        257.37 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         38.80 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 83c349e..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        667.48 ± 9.09 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.15 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index f3f74b8..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        255.32 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         38.87 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 5b42185..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           pp512 |        862.37 ± 1.02 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           tg128 |          5.59 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 2ba5783..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        183.43 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          5.21 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index 9676853..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           pp512 |       585.93 ± 27.57 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           tg128 |         57.18 ± 0.07 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index de517e6..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        160.38 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         38.70 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1.log
deleted file mode 100644
index 4b54d27..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           pp512 |        548.07 ± 6.87 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           tg128 |         73.66 ± 0.02 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index ae41eeb..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        337.62 ± 3.36 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         52.54 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 9aea9e3..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           pp512 |        546.89 ± 7.03 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           tg128 |         73.57 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 9029d6a..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        337.07 ± 4.50 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         52.54 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log
deleted file mode 100644
index 872dcfb..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1776.02 ± 15.96 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.43 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 409ef11..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        596.09 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         57.20 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index a92dd8c..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1777.68 ± 17.27 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.45 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 3d98873..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        546.37 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         57.26 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log
deleted file mode 100644
index e0b5cee..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           pp512 |      1734.91 ± 19.75 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           tg128 |         73.48 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index e5b1d94..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        557.68 ± 0.88 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         57.80 ± 0.01 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 5330505..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           pp512 |      1709.32 ± 28.51 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |           tg128 |         73.53 ± 0.00 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 52dd5fe..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        541.93 ± 1.58 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |         57.84 ± 0.02 |
-
-build: e0c93af2a (7938)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7.1.1__fa1.log
deleted file mode 100644
index ecfd557..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1710.76 ± 45.80 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.34 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index c21ca46..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        420.14 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         55.52 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 4a79534..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1732.05 ± 23.11 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.34 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d853d82..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        417.44 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         55.34 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log
deleted file mode 100644
index ca81fc1..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           pp512 |     1771.72 ± 240.97 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           tg128 |          7.95 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index a764fdc..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        294.23 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          7.42 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log
deleted file mode 100644
index 65dad4f..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           pp512 |      1429.10 ± 24.10 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           tg128 |         80.56 ± 0.18 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index b18d7b4..0000000
--- a/benchmark/results/09-01-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        284.79 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         56.04 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm-7_2__fa1.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm-7_2__fa1.log
deleted file mode 100644
index 4834219..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           pp512 |        549.23 ± 0.33 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           tg128 |         50.62 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm-7_2__fa1__longctx32768.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index fa012c7..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        139.39 ± 0.75 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          5.59 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 58b1ece..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           pp512 |        549.22 ± 0.08 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |           tg128 |         50.67 ± 0.01 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 1f2b6d2..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 | pp2048 @ d32768 |        137.30 ± 1.39 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |   tg32 @ d32768 |          5.60 ± 0.00 |
-
-build: a14b960bc (7816)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1.log
deleted file mode 100644
index 28fa54f..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1595.85 ± 4.24 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.05 ± 0.02 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index eabd52b..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        188.69 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.93 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 97ed387..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1599.42 ± 4.42 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.06 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 8b942a9..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        187.77 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.93 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log
deleted file mode 100644
index 0008484..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |        671.91 ± 0.33 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.72 ± 0.02 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 724fd52..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        147.45 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.59 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 55d9338..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |        671.59 ± 0.34 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.64 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d7a864a..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        145.15 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.59 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7.1.1__fa1.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7.1.1__fa1.log
deleted file mode 100644
index acd5e4e..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1578.33 ± 5.18 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.91 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7.1.1__fa1__longctx32768.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7.1.1__fa1__longctx32768.log
deleted file mode 100644
index 2bd4e89..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7.1.1__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        165.75 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.79 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7.1.1__hblt0__fa1.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 724b4d9..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1587.41 ± 1.76 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.89 ± 0.01 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7.1.1__hblt0__fa1__longctx32768.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7.1.1__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 92c1781..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__rocm7.1.1__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        165.58 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.78 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 7655db3..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           pp512 |     1261.90 ± 215.63 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           tg128 |          6.26 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 5a169c2..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        104.76 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          4.00 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__vulkan_radv__fa1.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__vulkan_radv__fa1.log
deleted file mode 100644
index 5548853..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1126.16 ± 2.34 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           tg128 |         55.42 ± 0.08 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log b/benchmark/results/09-01-2026/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 697ab8d..0000000
--- a/benchmark/results/09-01-2026/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        162.11 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          8.74 ± 0.00 |
-
-build: 9c142e3a2 (7670)
diff --git a/benchmark/results/09-01-2026/system_info.json b/benchmark/results/09-01-2026/system_info.json
deleted file mode 100644
index d29bc7f..0000000
--- a/benchmark/results/09-01-2026/system_info.json
+++ /dev/null
@@ -1,6 +0,0 @@
-{
-    "distro": "Fedora Linux 42 (Workstation Edition)",
-    "kernel": "6.18.3-100.fc42.x86_64",
-    "linux_firmware": "linux-firmware-20251111-1.fc42.noarch",
-    "timestamp": "09 Jan 2026"
-}
\ No newline at end of file
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index 3cebd78..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        247.81 ± 0.75 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.45 ± 0.27 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 37f9fc9..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         37.61 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.66 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 7ac8d22..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        246.64 ± 0.87 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.63 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 285f8a6..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         37.54 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7fbe5a1d45a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fbe5a1d496b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7fbe5a1d4aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2cb1972) [0x7fbe5cf42972]
-/usr/local/lib64/libggml-hip.so.0(+0x2cb6b0e) [0x7fbe5cf47b0e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7fbe5a1ebe5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7fbe5d63eab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7fbe59b6a5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fbe59b6a668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm-7alpha-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__fa1.log
deleted file mode 100644
index b28b43c..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        250.33 ± 0.67 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.70 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index 133d093..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         84.50 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7fa4112eb5a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fa4112eb96b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7fa4112ebaef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5a8e2) [0x7fa4141028e2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5fa7e) [0x7fa414107a7e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7fa411302e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7fa4147d3ab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7fa410c815b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fa410c81668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm-7alpha] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index f6b1cab..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        250.13 ± 0.62 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.71 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 0e4b151..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f3b59a565a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f3b59a5696b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f3b59a56aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5a8e2) [0x7f3b5c86d8e2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5fa7e) [0x7f3b5c872a7e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f3b59a6de5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f3b5cf3eab0]
-/usr/local/bin/llama-bench() [0x40adbc]
-/usr/local/bin/llama-bench() [0x4088ac]
-/lib64/libc.so.6(+0x35b5) [0x7f3b593ec5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f3b593ec668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm-7alpha] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index b9717bd..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        330.74 ± 2.03 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.74 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index f90aec0..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,9 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         33.80 ± 0.00 |
-HW Exception by GPU node-1 (Agent handle: 0x107a8d10) reason :GPU Hang
-✖ ! [rocm6_4_4-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index a4a0e10..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        330.13 ± 0.85 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.73 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index e66c0c4..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,9 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         33.91 ± 0.00 |
-HW Exception by GPU node-1 (Agent handle: 0x1f16bd10) reason :GPU Hang
-✖ ! [rocm6_4_4-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index f8bb9fc..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        333.45 ± 1.70 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.33 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index a9adb30..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         98.64 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         13.16 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 1052ad9..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        336.20 ± 2.04 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.77 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 8226b0c..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         98.44 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.88 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index c2a507e..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        323.36 ± 0.16 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.68 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 54e1c24..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         47.07 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f6af45f15a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f6af45f196b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f6af45f1aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca0682) [0x7f6af734e682]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca585e) [0x7f6af735385e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f6af4608e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f6af7a23ab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7f6af3f875b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f6af3f87668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 63f3a20..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        323.91 ± 1.10 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.68 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 02f852c..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         46.62 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f95789005a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f957890096b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f9578900aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca0682) [0x7f957b65d682]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca585e) [0x7f957b66285e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f9578917e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f957bd32ab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7f95782965b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f9578296668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1.log
deleted file mode 100644
index 555b577..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        330.90 ± 1.42 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.83 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index e9e7670..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-Hip error: 'an illegal memory access was encountered'(700) at /longer_pathname_so_that_rpms_can_support_packaging_the_debug_info_for_all_os_profiles/src/rocm-libraries/projects/hipblaslt/library/src/amd_detail/hipblaslt.cpp:147
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-✖ ! [rocm7.1.1] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 7a33710..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        329.23 ± 1.32 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.83 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 849d02b..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7fb26dd2a5a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fb26dd2a96b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7fb26dd2aaef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4b5f2) [0x7fb270b325f2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d507ce) [0x7fb270b377ce]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7fb26dd41e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7fb271232ab0]
-/usr/local/bin/llama-bench() [0x40adbc]
-/usr/local/bin/llama-bench() [0x4088ac]
-/lib64/libc.so.6(+0x35b5) [0x7fb26d6c05b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fb26d6c0668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index 92f1c15..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        323.77 ± 1.72 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.70 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 3a80892..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         46.38 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7ffa533f15a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7ffa533f196b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7ffa533f1aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca0682) [0x7ffa5614e682]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca585e) [0x7ffa5615385e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7ffa53408e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7ffa56823ab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7ffa52d875b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7ffa52d87668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7_rc-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 1d3fff3..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        323.19 ± 0.84 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.69 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 568e806..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         46.51 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7ff4771b65a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7ff4771b696b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7ff4771b6aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca0682) [0x7ff479f13682]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca585e) [0x7ff479f1885e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7ff4771cde5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7ff47a5e8ab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7ff476b4c5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7ff476b4c668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7_rc-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__fa1.log
deleted file mode 100644
index 22ca86f..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        330.87 ± 0.79 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.59 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 508ce88..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,17 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-:0:rocdevice.cpp            :3582: 48997963017 us:  Callback: Queue 0x7ff041800000 aborting with error : HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION: The agent attempted to access memory beyond the largest legal address. code: 0x29
-Hip error: 'an illegal memory access was encountered'(700) at /therock/src/rocm-libraries/projects/hipblaslt/library/src/amd_detail/hipblaslt.cpp:147
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-Kernel Name: _ZL15flash_attn_tileILi128ELi128ELi16ELi4ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil
-VGPU=0xe715690 SWq=0x7ff143a14000, HWq=0x7ff041800000, id=3
-	Dispatch Header =0xb02 (type=2, barrier=1, acquire=1, release=1), setup=0
-	grid=[4096, 8, 24], workgroup=[32, 8, 1]
-	private_seg_size=0, group_seg_size=33792
-	kernel_obj=0x7fdfbe030100, kernarg_address=0x0x7ff040801600
-	completion_signal=0x0, correlation_id=0
-	rptr=15, wptr=47
- ✖ ! [rocm7_rc] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index 4a13b64..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        330.19 ± 0.73 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.82 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index bf772f6..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f126437b5a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f126437b96b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f126437baef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4b5f2) [0x7f12671835f2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d507ce) [0x7f12671887ce]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f1264392e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f1267858ab0]
-/usr/local/bin/llama-bench() [0x40adbc]
-/usr/local/bin/llama-bench() [0x4088ac]
-/lib64/libc.so.6(+0x35b5) [0x7f1263d115b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f1263d11668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7_rc] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 504b7db..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           pp512 |        228.89 ± 0.52 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         24.48 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index 0483c52..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |         40.49 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |          9.30 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 3653d58..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           pp512 |        243.57 ± 0.43 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         24.54 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index d1590a2..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |         52.62 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         14.35 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index 3c85cf5..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        194.43 ± 0.27 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.65 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index bd8de34..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         36.61 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f4fa9af05a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f4fa9af096b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f4fa9af0aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2cb1972) [0x7f4fac85e972]
-/usr/local/lib64/libggml-hip.so.0(+0x2cb6b0e) [0x7f4fac863b0e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f4fa9b07e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f4facf5aab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7f4fa94865b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f4fa9486668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm-7alpha-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index c437cd4..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        195.23 ± 0.26 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.64 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 2e0cf96..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         36.83 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.40 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__fa1.log
deleted file mode 100644
index fae0037..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        195.45 ± 0.65 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.50 ± 0.31 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index 3ebab6a..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,4 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index c56811e..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        195.71 ± 0.70 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.69 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 3ebab6a..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,4 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index f984551..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        275.04 ± 0.75 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.57 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 942f525..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,9 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         33.70 ± 0.00 |
-HW Exception by GPU node-1 (Agent handle: 0x9cb5d10) reason :GPU Hang
-✖ ! [rocm6_4_4-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 8422388..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        272.75 ± 1.25 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.56 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index b9c5f49..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,9 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         33.85 ± 0.00 |
-HW Exception by GPU node-1 (Agent handle: 0x2738fd10) reason :GPU Hang
-✖ ! [rocm6_4_4-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index 45ffea3..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        277.38 ± 0.34 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.52 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index 6b49d27..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         92.73 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         11.12 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index c29adb3..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        277.33 ± 0.75 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.62 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 226091a..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,9 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         92.73 ± 0.00 |
-HW Exception by GPU node-1 (Agent handle: 0x3c5c0d10) reason :GPU Hang
-✖ ! [rocm6_4_4] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index ab6d7f5..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        254.32 ± 0.84 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.51 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 594c729..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         46.17 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f75321e15a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f75321e196b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f75321e1aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca0682) [0x7f7534f3e682]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca585e) [0x7f7534f4385e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f75321f8e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f7535613ab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7f7531b775b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f7531b77668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index bb9476b..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        253.04 ± 1.12 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.50 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 2312744..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         46.53 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f042c8285a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f042c82896b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f042c828aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca0682) [0x7f042f585682]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca585e) [0x7f042f58a85e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f042c83fe5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f042fc5aab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7f042c1be5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f042c1be668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__fa1.log
deleted file mode 100644
index ded8528..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        257.70 ± 0.50 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.59 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index d557084..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,23 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7fca6c0ef5a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fca6c0ef96b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7fca6c0efaef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4b5f2) [0x7fca6eef75f2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d507ce) [0x7fca6eefc7ce]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7dd) [0x7fca6c10a46d]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fca6f5f87e0]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe2) [0x7fca6f5fa2b2]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x3bf) [0x7fca6f5ff6ff]
-/usr/local/lib64/libllama.so.0(llama_decode+0xe) [0x7fca6f6004fe]
-/usr/local/bin/llama-bench() [0x40ad9b]
-/usr/local/bin/llama-bench() [0x408a57]
-/lib64/libc.so.6(+0x35b5) [0x7fca6ba855b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fca6ba85668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 499a6d6..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        259.40 ± 0.46 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.61 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 4e96acb..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f4b572795a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f4b5727996b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f4b57279aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4b5f2) [0x7f4b5a0815f2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d507ce) [0x7f4b5a0867ce]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f4b57290e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f4b5a781ab0]
-/usr/local/bin/llama-bench() [0x40adbc]
-/usr/local/bin/llama-bench() [0x4088ac]
-/lib64/libc.so.6(+0x35b5) [0x7f4b56c0f5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f4b56c0f668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index 080c5ce..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        254.22 ± 1.28 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.50 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 4e835eb..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         45.90 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7fc72ee915a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fc72ee9196b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7fc72ee91aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca0682) [0x7fc731bee682]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca585e) [0x7fc731bf385e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7fc72eea8e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7fc7322c3ab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7fc72e8275b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fc72e827668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7_rc-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 65abdf8..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        253.25 ± 1.33 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.53 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 34c164b..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         45.93 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7fc83c7145a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fc83c71496b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7fc83c714aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca0682) [0x7fc83f471682]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca585e) [0x7fc83f47685e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7fc83c72be5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7fc83fb46ab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7fc83c0aa5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fc83c0aa668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7_rc-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__fa1.log
deleted file mode 100644
index 522f67c..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           pp512 |        258.89 ± 0.25 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.54 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 32379f8..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         79.91 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         10.08 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index c5c6987..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,27 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f48fca035a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f48fca0396b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f48fca03aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4b5f2) [0x7f48ff80b5f2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5fe47) [0x7f48ff81fe47]
-/usr/local/lib64/libggml-hip.so.0(_Z19ggml_cuda_mul_mat_qR25ggml_backend_cuda_contextPK11ggml_tensorS3_S3_PS1_+0x7d3) [0x7f48ff98aba3]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5802c) [0x7f48ff81802c]
-/usr/local/lib64/libggml-hip.so.0(+0x2d53e28) [0x7f48ff813e28]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5083f) [0x7f48ff81083f]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7f48fca1e483]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f48ffee17e0]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe2) [0x7f48ffee32b2]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x3bf) [0x7f48ffee86ff]
-/usr/local/lib64/libllama.so.0(llama_decode+0xe) [0x7f48ffee94fe]
-/usr/local/bin/llama-bench() [0x40ad9b]
-/usr/local/bin/llama-bench() [0x4088ac]
-/lib64/libc.so.6(+0x35b5) [0x7f48fc3995b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f48fc399668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7_rc] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__hblt0__fa1 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 786a2a8..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7fb299bf75a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fb299bf796b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7fb299bf7aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4b5f2) [0x7fb29c9ff5f2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d507ce) [0x7fb29ca047ce]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7fb299c0ee5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7fb29d0d4ab0]
-/usr/local/bin/llama-bench() [0x40adbc]
-/usr/local/bin/llama-bench() [0x4088ac]
-/lib64/libc.so.6(+0x35b5) [0x7fb29958d5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fb29958d668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7_rc] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index da28589..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           pp512 |        279.25 ± 0.28 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         17.61 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index e54f589..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |         42.15 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |          7.96 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index 8a162f3..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           pp512 |        244.36 ± 0.45 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         17.73 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index 512278a..0000000
--- a/benchmark/results/20-12-2025/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |         54.92 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         11.62 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index d927b34..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |         65.74 ± 0.01 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 811a47d..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         23.88 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          1.52 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index c040916..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |         65.41 ± 0.02 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 2466e1d..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         24.05 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          1.52 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1.log
deleted file mode 100644
index 4abdd5e..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |         65.85 ± 0.01 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index 8007afd..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         37.47 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          2.61 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index 19f1cc0..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |         65.38 ± 0.05 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 33e4c02..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         37.86 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          2.51 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index de26d53..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        145.84 ± 0.07 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.78 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 8b47c40..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         26.43 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          1.85 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index bb0fc76..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        144.36 ± 0.18 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.78 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 1022078..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,9 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         26.46 ± 0.00 |
-HW Exception by GPU node-1 (Agent handle: 0x31c9cd10) reason :GPU Hang
-✖ ! [rocm6_4_4-rocwmma] Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 03f0f5b..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        145.01 ± 0.05 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.77 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index 3d5ade1..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         56.24 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          2.61 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 0ff9545..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        146.28 ± 0.12 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.78 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 9e418ec..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         56.12 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          2.60 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index 6224321..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        146.01 ± 0.05 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 57c1117..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         35.23 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f2519e175a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f2519e1796b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f2519e17aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca0682) [0x7f251cb74682]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca585e) [0x7f251cb7985e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f2519e2ee5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f251d249ab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7f25197ad5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f25197ad668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1-rocwmma] Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index f6db3a2..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        143.94 ± 0.16 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index abd30e6..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         34.82 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          1.86 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1.log
deleted file mode 100644
index cc7a19c..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        147.07 ± 0.01 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index 2ac9624..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f3d030795a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f3d0307996b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f3d03079aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4b5f2) [0x7f3d05e815f2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d507ce) [0x7f3d05e867ce]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f3d03090e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f3d06581ab0]
-/usr/local/bin/llama-bench() [0x40adbc]
-/usr/local/bin/llama-bench() [0x408b3d]
-/lib64/libc.so.6(+0x35b5) [0x7f3d02a0f5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f3d02a0f668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1] Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 9b4a794..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        145.12 ± 0.04 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 22906b2..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         52.68 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          2.50 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index a0078fb..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        145.84 ± 0.08 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index f6357ab..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         35.12 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          1.85 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 94c795b..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        143.47 ± 0.06 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index a58a6dd..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         35.06 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          1.82 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__fa1.log
deleted file mode 100644
index 4598534..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        145.96 ± 0.08 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.78 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 25b568d..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,28 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f357c5865a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f357c58696b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f357c586aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4b5f2) [0x7f357f38e5f2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5fe47) [0x7f357f3a2e47]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5e41c) [0x7f357f3a141c]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5d5e7) [0x7f357f3a05e7]
-/usr/local/lib64/libggml-hip.so.0(+0x2d580cb) [0x7f357f39b0cb]
-/usr/local/lib64/libggml-hip.so.0(+0x2d53e28) [0x7f357f396e28]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5083f) [0x7f357f39383f]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7f357c5a1483]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f357fa647e0]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe2) [0x7f357fa662b2]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x3bf) [0x7f357fa6b6ff]
-/usr/local/lib64/libllama.so.0(llama_decode+0xe) [0x7f357fa6c4fe]
-/usr/local/bin/llama-bench() [0x40ad9b]
-/usr/local/bin/llama-bench() [0x4088ac]
-/lib64/libc.so.6(+0x35b5) [0x7f357bf1c5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f357bf1c668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7_rc] Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index 4d9e8ad..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        145.24 ± 0.16 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 4b7c8ce..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         52.70 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          2.57 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 3c6c07c..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 |           pp512 |        100.73 ± 0.26 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 |           tg128 |          2.80 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index 22f5b61..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |         30.75 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |          2.46 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 022269b..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 |           pp512 |         88.29 ± 0.76 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 |           tg128 |          2.77 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index fafeab9..0000000
--- a/benchmark/results/20-12-2025/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |         31.75 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |          2.56 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index 0a7e1b1..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        788.05 ± 2.61 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.48 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index ac3d66f..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        139.54 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          5.56 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 2a91c3d..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        787.49 ± 3.69 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.48 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 0ae20f9..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        136.23 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          5.56 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha__fa1.log
deleted file mode 100644
index f2bbffd..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        799.32 ± 2.30 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.49 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index 99f7301..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        273.53 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          7.73 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index 5146b25..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        804.49 ± 4.05 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.49 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 91170dc..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        282.94 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          7.73 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index dcb50e9..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        792.57 ± 2.08 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.47 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 6d3e55b..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        121.02 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          6.74 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 8a899fd..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        794.96 ± 3.40 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.47 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 65b21b6..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        118.46 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          6.74 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log
deleted file mode 100644
index 176af12..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        802.78 ± 0.92 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.49 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index 758716e..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        289.71 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          7.73 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 8016de9..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        810.15 ± 2.26 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.49 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index e1046e5..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        316.13 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          7.73 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index 2eb1c65..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        781.49 ± 1.79 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.49 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 4914b28..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        176.40 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          6.84 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 916ac50..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        795.45 ± 1.95 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.49 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 47f5829..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        179.22 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          6.84 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__fa1.log
deleted file mode 100644
index 23a623e..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        797.09 ± 3.94 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.50 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index 4cb4f21..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        261.91 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          7.74 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 036342e..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        800.44 ± 2.67 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.50 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 65bfaf3..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        282.67 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          7.74 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index 5b64204..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        794.13 ± 2.11 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.49 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index a3fd4c8..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        177.29 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          6.84 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index c886bbe..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        757.71 ± 0.34 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.49 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index f50d2fb..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        175.04 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          6.83 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc__fa1.log
deleted file mode 100644
index 7395dcc..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        803.93 ± 1.68 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.50 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 1ba6a51..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        266.48 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          7.74 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index e0dc575..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        767.62 ± 3.68 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.50 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index ba3c89c..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        280.71 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          7.74 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 8803b27..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           pp512 |        228.96 ± 0.25 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           tg128 |          8.20 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index 7d406bf..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        101.86 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |          6.76 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log
deleted file mode 100644
index 7712766..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           pp512 |        198.30 ± 1.06 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           tg128 |          7.57 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index b4d60c2..0000000
--- a/benchmark/results/20-12-2025/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |         98.61 ± 0.00 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |          6.84 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index b935b7e..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |        575.90 ± 0.60 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.67 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 04b7b08..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        815.96 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         37.94 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 59f12dc..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |        566.25 ± 5.65 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.75 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index f2ecfc7..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        818.18 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         37.96 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha__fa1.log
deleted file mode 100644
index 583fbf1..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |        573.96 ± 1.31 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.73 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index 93a9c1f..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        840.39 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         40.57 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index 38ae77f..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |        568.52 ± 6.46 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.75 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 69cc8c3..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        842.90 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         40.55 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index 5e2df6e..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1111.52 ± 3.84 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.23 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 1b85141..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1161.59 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         38.22 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index f7bf3a0..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1075.82 ± 2.72 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.19 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 69b8f82..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1159.76 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         38.28 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index 5ce08c0..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1111.90 ± 4.65 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.13 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index 857e033..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1083.84 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.84 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index e818cc0..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1074.40 ± 7.61 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.17 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 00c9009..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1126.17 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.77 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index 248c243..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1110.04 ± 2.67 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.43 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 1b323f7..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1034.85 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         38.58 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index c6ab5d1..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1073.92 ± 6.88 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.52 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 04dbbae..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1041.85 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         38.59 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__fa1.log
deleted file mode 100644
index 31d6990..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1119.24 ± 8.14 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.47 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index 749e116..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        964.26 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         40.21 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 6178218..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1086.57 ± 5.04 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.44 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 61cd367..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        972.22 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         40.23 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index 8e0530d..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1100.11 ± 6.03 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.51 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 0d652a5..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1036.83 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         38.63 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 20d0bbe..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1066.41 ± 6.56 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.45 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 6a5e27e..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1030.17 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         38.56 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc__fa1.log
deleted file mode 100644
index 6d6d161..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1110.36 ± 6.45 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.23 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index d566191..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1042.68 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         40.00 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index 9edc424..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           pp512 |       1074.73 ± 4.78 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.51 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index a140be0..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1001.21 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         40.21 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index dbd7200..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1224.54 ± 7.71 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 |           tg128 |         46.56 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index 31888ea..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        569.03 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         39.60 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index 08568b4..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 |           pp512 |        940.69 ± 5.60 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 |           tg128 |         45.38 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index f7023ef..0000000
--- a/benchmark/results/20-12-2025/Nemotron-3-Nano-30B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        489.74 ± 0.00 |
-| nemotron_h_moe 31B.A3.5B Q8_0  |  37.66 GiB |    31.58 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         42.30 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index 1c6bafc..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        195.33 ± 3.19 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.05 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index a42d7d6..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         73.22 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f12702305a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f127023096b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f1270230aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2cb1972) [0x7f1272f9e972]
-/usr/local/lib64/libggml-hip.so.0(+0x2cb6b0e) [0x7f1272fa3b0e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f1270247e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f127369aab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7f126fbc65b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f126fbc6668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm-7alpha-rocwmma] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 04baa81..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        195.28 ± 1.09 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.09 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 871b575..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         72.86 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7fd4172615a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fd41726196b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7fd417261aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2cb1972) [0x7fd419fcf972]
-/usr/local/lib64/libggml-hip.so.0(+0x2cb6b0e) [0x7fd419fd4b0e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7fd417278e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7fd41a6cbab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7fd416bf75b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fd416bf7668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm-7alpha-rocwmma] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha__fa1.log
deleted file mode 100644
index 57846b1..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        194.82 ± 1.18 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.06 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index c5d2182..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         63.02 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         10.73 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index 466d7a6..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        196.80 ± 1.26 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.01 ± 0.07 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 3220457..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         65.07 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         11.83 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index 3d604c2..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        198.21 ± 1.42 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.09 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index b607ba2..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,9 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         66.78 ± 0.00 |
-HW Exception by GPU node-1 (Agent handle: 0x3bf1cd10) reason :GPU Hang
-✖ ! [rocm6_4_4-rocwmma] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 934a790..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        202.49 ± 1.92 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.09 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index a8ac7a1..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         69.68 ± 0.00 |
-Memory access fault by GPU node-1 (Agent handle: 0x168dcd10) on address 0x7fc71a789000. Reason: Page not present or supervisor privilege.
-:0:rocdevice.cpp            :2992: 91563147890 us:  Callback: Queue 0x7fe05ae00000 aborting with error : HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION: The agent attempted to access memory beyond the largest legal address. code: 0x29
-✖ ! [rocm6_4_4-rocwmma] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index bb2c5e9..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        205.39 ± 1.95 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.83 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index 8924df7..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,6 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-HW Exception by GPU node-1 (Agent handle: 0x35d3ed10) reason :GPU Hang
-✖ ! [rocm6_4_4] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 287edc5..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        201.21 ± 1.57 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.97 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 164727d..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,6 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-Memory access fault by GPU node-1 (Agent handle: 0x1c519d10) on address 0x7eff38303000. Reason: Page not present or supervisor privilege.
-✖ ! [rocm6_4_4] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index 90e4554..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        197.07 ± 0.88 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.09 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 564eb79..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         65.17 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          5.26 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 3d9aad0..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        195.14 ± 1.06 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.08 ± 0.05 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 960c254..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         64.99 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7fc8c7bfb5a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fc8c7bfb96b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7fc8c7bfbaef]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca0682) [0x7fc8ca958682]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca585e) [0x7fc8ca95d85e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7fc8c7c12e5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7fc8cb02dab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7fc8c75915b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fc8c7591668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1-rocwmma] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__fa1.log
deleted file mode 100644
index aee2b6f..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        199.48 ± 2.40 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.91 ± 0.14 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index 6d4a794..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,27 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7ff4276395a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7ff42763996b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7ff427639aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4b5f2) [0x7ff42a4415f2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5f0bb) [0x7ff42a4550bb]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5d5e7) [0x7ff42a4535e7]
-/usr/local/lib64/libggml-hip.so.0(+0x2d580cb) [0x7ff42a44e0cb]
-/usr/local/lib64/libggml-hip.so.0(+0x2d53e28) [0x7ff42a449e28]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5083f) [0x7ff42a44683f]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7ff427654483]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7ff42ab427e0]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe2) [0x7ff42ab442b2]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x3bf) [0x7ff42ab496ff]
-/usr/local/lib64/libllama.so.0(llama_decode+0xe) [0x7ff42ab4a4fe]
-/usr/local/bin/llama-bench() [0x40ad9b]
-/usr/local/bin/llama-bench() [0x4088ac]
-/lib64/libc.so.6(+0x35b5) [0x7ff426fcf5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7ff426fcf668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 59896b1..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,24 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f75f10bf5a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f75f10bf96b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f75f10bfaef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4b5f2) [0x7f75f3ec75f2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d55ff5) [0x7f75f3ed1ff5]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5083f) [0x7f75f3ecc83f]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7f75f10da483]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f75f45c87e0]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe2) [0x7f75f45ca2b2]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x3bf) [0x7f75f45cf6ff]
-/usr/local/lib64/libllama.so.0(llama_decode+0xe) [0x7f75f45d04fe]
-/usr/local/bin/llama-bench() [0x40ad9b]
-/usr/local/bin/llama-bench() [0x4088ac]
-/lib64/libc.so.6(+0x35b5) [0x7f75f0a555b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f75f0a55668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__hblt0__fa1 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 3ebab6a..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,4 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index 04307a9..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        197.33 ± 0.72 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.13 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 977deb3..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,20 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         62.78 ± 0.00 |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7efc563255a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7efc5632596b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7efc56325aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca0682) [0x7efc59082682]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca585e) [0x7efc5908785e]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7efc5633ce5e]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7efc59757ab0]
-/usr/local/bin/llama-bench() [0x408c12]
-/lib64/libc.so.6(+0x35b5) [0x7efc55cbb5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7efc55cbb668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7_rc-rocwmma] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 2fd430d..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        199.34 ± 1.19 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.09 ± 0.06 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index c1ee622..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         62.68 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          5.35 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc__fa1.log
deleted file mode 100644
index bb619c0..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        201.37 ± 1.76 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.91 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 0e71630..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,24 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7fb6064785a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fb60647896b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7fb606478aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4b5f2) [0x7fb6092805f2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d55ff5) [0x7fb60928aff5]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5083f) [0x7fb60928583f]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7fb606493483]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fb6099567e0]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe2) [0x7fb6099582b2]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x3bf) [0x7fb60995d6ff]
-/usr/local/lib64/libllama.so.0(llama_decode+0xe) [0x7fb60995e4fe]
-/usr/local/bin/llama-bench() [0x40ad9b]
-/usr/local/bin/llama-bench() [0x4088ac]
-/lib64/libc.so.6(+0x35b5) [0x7fb605e0e5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fb605e0e668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7_rc] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index ba426aa..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        199.51 ± 1.70 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.98 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 35b0598..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,24 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-/opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7ff5be8415a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7ff5be84196b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7ff5be841aef]
-/usr/local/lib64/libggml-hip.so.0(+0x2d4b5f2) [0x7ff5c16495f2]
-/usr/local/lib64/libggml-hip.so.0(+0x2d55ff5) [0x7ff5c1653ff5]
-/usr/local/lib64/libggml-hip.so.0(+0x2d5083f) [0x7ff5c164e83f]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7ff5be85c483]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7ff5c1d1f7e0]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe2) [0x7ff5c1d212b2]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x3bf) [0x7ff5c1d266ff]
-/usr/local/lib64/libllama.so.0(llama_decode+0xe) [0x7ff5c1d274fe]
-/usr/local/bin/llama-bench() [0x40ad9b]
-/usr/local/bin/llama-bench() [0x4088ac]
-/lib64/libc.so.6(+0x35b5) [0x7ff5be1d75b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7ff5be1d7668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7_rc] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 9cfba84..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |           pp512 |        145.16 ± 0.17 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |           tg128 |         17.77 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index 8419851..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |         31.17 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |          5.72 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index 93bee5e..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |           pp512 |        131.53 ± 1.13 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |           tg128 |         18.08 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index e397722..0000000
--- a/benchmark/results/20-12-2025/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |         34.22 ± 0.00 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |          9.66 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index 5730ae3..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        489.10 ± 3.76 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.10 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 4d63c40..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        282.60 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         13.97 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 0f8a52e..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        491.47 ± 1.44 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.14 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 95b1cb9..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        271.39 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         14.61 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha__fa1.log
deleted file mode 100644
index 71be6fa..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        491.95 ± 0.75 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.10 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index 8240393..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        336.95 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         22.45 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index 86257e9..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        492.47 ± 3.40 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.15 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 2f690ca..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        318.73 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         22.44 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index 911023a..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        567.78 ± 2.40 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.92 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 3aa46ab..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        268.54 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         18.12 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 2b2c080..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        575.99 ± 6.42 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.90 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index bbeaf1a..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        255.99 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         18.07 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index b35dbc0..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        576.13 ± 3.25 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.13 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index d6e3121..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,6 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-Memory access fault by GPU node-1 (Agent handle: 0x3b630d10) on address (nil). Reason: Page not present or supervisor privilege.
-✖ ! [rocm6_4_4] Qwen3-30B-A3B-BF16-00001-of-00002__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index f3428c4..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,6 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-HW Exception by GPU node-1 (Agent handle: 0x3cc03d10) reason :GPU Hang
-✖ ! [rocm6_4_4] Qwen3-30B-A3B-BF16-00001-of-00002__hblt0__fa1 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 35fc41e..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        377.73 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         22.29 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index b8f1d36..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        460.49 ± 1.91 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.12 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index e5274e9..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        234.40 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         17.98 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index ebf1c94..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        467.86 ± 1.23 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.11 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 04c685f..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        224.91 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         18.00 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__fa1.log
deleted file mode 100644
index 7edcf87..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        468.47 ± 2.10 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.08 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index eeddc58..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        360.38 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         22.31 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index d7db6d1..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        470.04 ± 3.69 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.02 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index d59ba93..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        337.92 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         22.32 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index 8e01011..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        461.77 ± 2.14 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.10 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 3203344..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        234.95 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         18.08 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 363c223..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        463.39 ± 4.60 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.15 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 769aa5f..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        223.07 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         18.04 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc__fa1.log
deleted file mode 100644
index d1196f9..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        466.61 ± 1.68 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.11 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 28bb187..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        319.20 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         22.52 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index 3140710..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        473.68 ± 1.69 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.11 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index bf51587..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        337.79 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         22.34 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 350bb76..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |        194.36 ± 0.12 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |          9.96 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index d329a30..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |         86.76 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |          8.71 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 3fa928a..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |        167.29 ± 0.18 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |          9.36 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index c038d21..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        102.90 ± 0.00 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |          8.57 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index e6192c9..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        883.95 ± 6.89 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         58.73 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 2f33b05..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        323.55 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         20.52 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 6cc0a0b..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        888.32 ± 6.59 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         58.55 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 99edf40..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        323.18 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         20.51 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha__fa1.log
deleted file mode 100644
index ec5dc7e..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        887.98 ± 2.33 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         58.53 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index 486fb05..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        273.12 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         40.25 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index afb9e9d..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       897.81 ± 13.27 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         58.56 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index a6ee119..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        273.07 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         40.23 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index 1081a96..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1045.84 ± 8.87 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         58.04 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index e96d662..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        310.27 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         28.10 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 40a0181..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1046.62 ± 8.31 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.78 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 972826c..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        313.66 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         28.11 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index 96b85d0..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1065.87 ± 15.74 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.71 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index c0d37cd..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        353.38 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.56 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index bc9b8e5..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1056.16 ± 8.88 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.68 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 5639307..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        341.15 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.59 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index 248c0ca..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       961.79 ± 10.60 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.69 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 64780d6..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        263.57 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         27.72 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index a6ced1b..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        964.88 ± 9.02 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.78 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 5318173..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        263.64 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         27.74 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__fa1.log
deleted file mode 100644
index 8acd7e7..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        992.39 ± 4.30 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.48 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index c655de8..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        286.94 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.51 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 42d8f1f..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        984.99 ± 7.73 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.39 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index a0ba594..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        284.86 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.29 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index 37e54c9..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        966.17 ± 7.59 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.85 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 9940008..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        263.45 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         27.74 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index b07a137..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        960.50 ± 8.25 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.91 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index f29e900..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        263.73 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         27.72 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc__fa1.log
deleted file mode 100644
index 0898dc9..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        986.79 ± 6.92 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.42 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 68652c6..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        410.85 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.33 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index 9c58c7b..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        984.07 ± 5.87 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         57.52 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 29f6134..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        282.74 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.35 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 66e05a7..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1129.76 ± 4.79 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         62.27 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index c56e9f5..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        135.73 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         33.09 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index 2b85a69..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |        905.18 ± 4.26 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         66.46 ± 0.05 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index 98a3d4b..0000000
--- a/benchmark/results/20-12-2025/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        172.42 ± 0.00 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         42.04 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index 769297d..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1189.07 ± 7.02 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         71.58 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 99786f1..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        332.99 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         21.83 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 6ef2d48..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1185.59 ± 10.60 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         71.65 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 39fd865..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        338.96 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         21.84 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha__fa1.log
deleted file mode 100644
index edf9feb..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1185.88 ± 9.66 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         71.64 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index eb57e0e..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        295.00 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         45.59 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index 4baec9f..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1189.90 ± 9.82 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         71.61 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 8e7e86b..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        291.08 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         45.74 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index c00de91..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1201.14 ± 12.83 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         68.92 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 4a793ea..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        319.92 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         30.24 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index fd3bded..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1208.02 ± 13.07 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         68.87 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 37260d9..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        321.94 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         30.28 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log
deleted file mode 100644
index ebb7e8e..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1225.68 ± 19.07 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         69.13 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index 5b810ee..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        387.32 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         43.53 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index cffdfb9..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1231.06 ± 2.02 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         69.08 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index cd8e169..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        361.60 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         44.08 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index 50347bc..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1172.29 ± 9.77 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         69.19 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 1300fbd..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        270.50 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         29.99 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index a50a1bd..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1169.43 ± 4.95 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         69.31 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index dafa37b..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        270.83 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         30.04 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__fa1.log
deleted file mode 100644
index 0a23174..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1183.05 ± 9.42 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         69.17 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index 510d0cb..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        292.36 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         43.95 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 0ecb4c7..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1195.38 ± 5.88 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         69.06 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index a0bbd29..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        287.87 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         43.84 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index 71866cb..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1166.57 ± 8.22 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         69.37 ± 0.05 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index ee0dbbb..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        270.55 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         30.01 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 28f33f2..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1167.10 ± 5.32 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         69.33 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 92ee452..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        270.39 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         30.09 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc__fa1.log
deleted file mode 100644
index b6cdbaa..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1194.63 ± 7.87 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         68.98 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 62b9aa2..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        292.47 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         43.87 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index b165442..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1202.35 ± 10.49 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         68.96 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 531a084..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        292.54 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         44.02 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 5d8e040..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |        825.86 ± 2.68 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         80.94 ± 0.04 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index 5b95baa..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        130.78 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         37.08 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log
deleted file mode 100644
index 94bec33..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |        864.66 ± 2.72 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         83.07 ± 0.04 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index 8b6ad3a..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        168.69 ± 0.00 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         47.63 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index 005fd2f..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        345.64 ± 0.84 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.88 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 432cbb5..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        344.40 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         20.67 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 75330e1..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        356.53 ± 3.90 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.26 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 88740e8..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        352.86 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         24.03 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1.log
deleted file mode 100644
index 9b4db1a..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        346.29 ± 1.98 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.92 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index e6be6ea..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        379.41 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         22.12 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index e2108ce..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        360.22 ± 1.39 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         27.35 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 94e037e..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        386.65 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         25.80 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index 0b57963..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        573.57 ± 2.61 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         25.89 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 8c312b8..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,9 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        464.91 ± 0.00 |
-Memory access fault by GPU node-1 (Agent handle: 0x260d7d10) on address 0x7efad124d000. Reason: Page not present or supervisor privilege.
-✖ ! [rocm6_4_4-rocwmma] Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 5dcd687..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        595.88 ± 2.98 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.34 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 10c5d70..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        480.86 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         23.30 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 31e9e1e..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        576.31 ± 0.99 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         25.64 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index 077eb4c..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        489.52 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         24.50 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 7269e2e..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        590.68 ± 0.83 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.38 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 6bfdd77..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        503.34 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         25.09 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index b4f68a7..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        571.05 ± 4.21 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.45 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 5bad71a..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        444.72 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         21.38 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 6a40914..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        594.40 ± 3.02 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.73 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 8a21a7c..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,33 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        451.46 ± 0.00 |
-:0:rocdevice.cpp            :3587: 7778304718 us:  Callback: Queue 0x7f5274500000 aborting with error : HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION: The agent attempted to access memory beyond the largest legal address. code: 0x29
-Kernel Name: _ZL18flash_attn_ext_vecILi256ELi1EL9ggml_type1ELS0_1ELb0EEvPKcS2_S2_S2_S2_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS6_IjLj3EEiiiiiiiiiiiliiliiiiil
-VGPU=0x3fa84a70 SWq=0x7f5276f10000, HWq=0x7f5274500000, id=2
-	Dispatch Header =0xb02 (type=2, barrier=1, acquire=1, release=1), setup=0
-	grid=[32, 60, 16], workgroup=[32, 4, 1]
-	private_seg_size=0, group_seg_size=8448
-	kernel_obj=0x7f5275b4f600, kernarg_address=0x0x7f3d88d43980
-	completion_signal=0x0, correlation_id=0
-	rptr=813327, wptr=816270
- /opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error
-/usr/local/lib64/libggml-base.so.0(+0x35a5) [0x7f52b67fb5a5]
-/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f52b67fb96b]
-/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f52b67fbaef]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca0682) [0x7f52b9558682]
-/usr/local/lib64/libggml-hip.so.0(+0x2cab085) [0x7f52b9563085]
-/usr/local/lib64/libggml-hip.so.0(+0x2ca58cf) [0x7f52b955d8cf]
-/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7f52b6816483]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f52b9c2e7e0]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe2) [0x7f52b9c302b2]
-/usr/local/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x3bf) [0x7f52b9c356ff]
-/usr/local/lib64/libllama.so.0(llama_decode+0xe) [0x7f52b9c364fe]
-/usr/local/bin/llama-bench() [0x408c92]
-/lib64/libc.so.6(+0x35b5) [0x7f52b61915b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f52b6191668]
-/usr/local/bin/llama-bench() [0x409c25]
-✖ ! [rocm7.1.1-rocwmma] Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__hblt0__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1.log
deleted file mode 100644
index 5299717..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        568.38 ± 2.63 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.50 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index 04895c4..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        462.35 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         22.01 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 6183815..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        587.50 ± 6.59 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.81 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 1c1bdde..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        479.80 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         21.62 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index eb49689..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        567.13 ± 2.43 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.44 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 6954010..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        438.42 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         23.36 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index c22f1e0..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |       594.63 ± 11.03 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.76 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 354c810..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        442.41 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         23.66 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__fa1.log
deleted file mode 100644
index f4855bd..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        570.28 ± 3.77 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.05 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 4d8d15d..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        473.92 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         24.80 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index 0d97f79..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        588.80 ± 0.73 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 534ab2b..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        488.38 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         21.96 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index c79c18c..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |        406.08 ± 1.14 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         33.67 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index e8d9bcb..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        194.48 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         27.27 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 45b0ac9..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |        326.83 ± 0.94 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         30.18 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index 320c1bb..0000000
--- a/benchmark/results/20-12-2025/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        220.99 ± 0.00 |
-| qwen3next 80B.A3B Q8_0         |  79.57 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         27.91 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index 361a6d7..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        421.94 ± 0.25 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.17 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index ecd2fdf..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        332.05 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         11.86 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index a340223..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        424.57 ± 0.40 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.17 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 8d852c2..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        325.94 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         11.85 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha__fa1.log
deleted file mode 100644
index d060db9..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        429.89 ± 0.35 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.26 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index ba5f705..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        309.80 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.77 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index 69b7431..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        430.69 ± 0.83 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.25 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 32c78b5..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        338.46 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.76 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index a2843f4..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        897.73 ± 0.63 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.14 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 312778f..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        288.29 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.01 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 5bb70aa..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        904.28 ± 1.51 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.15 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 7ded8ad..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        290.79 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.00 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index 235334a..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        913.75 ± 0.60 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.21 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index db48774..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        425.43 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.72 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 184977a..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        919.42 ± 0.44 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.19 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 616dc46..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        430.76 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.78 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index bc50843..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        886.59 ± 0.64 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.17 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index fc15a41..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        303.06 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.05 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 328096e..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        890.97 ± 0.75 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.17 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 64f9cf7..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        304.61 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.03 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__fa1.log
deleted file mode 100644
index 05f259d..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        910.34 ± 0.61 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.26 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index 4766a18..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        391.58 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.80 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 9cbc2a8..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        917.22 ± 0.99 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.28 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 58aaf7a..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        394.15 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.77 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index c77a98f..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        882.81 ± 0.63 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.16 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 27debaf..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        302.60 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.02 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 8c5636c..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        891.69 ± 0.16 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.16 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 828c786..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        306.94 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.03 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc__fa1.log
deleted file mode 100644
index 522f057..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        913.03 ± 1.04 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.26 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 7f19c44..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        399.02 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.77 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index 19c608d..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        917.06 ± 0.24 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.27 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 5fd8d29..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        391.45 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.78 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 1e29149..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |           pp512 |        690.02 ± 1.72 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |           tg128 |         14.56 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index 411a64f..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |         54.64 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         12.98 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index 649e0ba..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |           pp512 |        549.78 ± 1.79 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |           tg128 |         13.95 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index 266830e..0000000
--- a/benchmark/results/20-12-2025/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        270.10 ± 0.00 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         11.42 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index 6da9bcd..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        468.46 ± 1.81 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.01 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 8996dbe..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        167.60 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.56 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index fa35df2..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        522.97 ± 0.37 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.01 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index d4ae304..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        170.01 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.56 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha__fa1.log
deleted file mode 100644
index ac992c7..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        477.16 ± 1.86 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.02 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index 9520a2f..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        237.04 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.87 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index 12e2b0a..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        527.33 ± 1.41 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.01 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index fdf835b..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        270.71 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.87 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index e164e47..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        468.30 ± 0.54 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.00 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index f9c5383..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        167.49 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.71 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index fef7449..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        525.67 ± 0.68 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.00 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index d70f1e9..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        169.41 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.71 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 08ada47..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        472.62 ± 0.27 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.00 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index 0f58777..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        246.86 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.86 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 59b9670..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        530.96 ± 0.63 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.00 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 63ed4a7..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        285.15 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.86 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index cd81a2b..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        463.74 ± 0.73 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.03 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 79e38a4..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        193.84 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.75 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 3ec22c1..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        522.71 ± 0.55 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.03 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index b029e97..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        205.34 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.76 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__fa1.log
deleted file mode 100644
index 4d7ac4f..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        466.55 ± 0.52 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.03 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index 1566960..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        234.18 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.89 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index d9990e7..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        526.17 ± 0.74 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.03 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index b854410..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        269.04 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.89 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index 8554509..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        462.78 ± 0.04 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.04 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 646709f..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        194.64 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.76 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index a04ab5b..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        499.88 ± 0.71 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.04 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 5e20aed..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        195.11 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.76 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__fa1.log
deleted file mode 100644
index 7e00b6d..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        474.02 ± 0.22 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.03 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 176bc97..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        308.72 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.85 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index 8372beb..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        506.00 ± 0.53 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          4.04 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index ca30c47..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        320.96 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |          3.89 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 7efd9cc..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-ggml_vulkan: Device memory allocation of size 2819260416 failed.
-ggml_vulkan: Requested buffer size exceeds device buffer size limit: ErrorOutOfDeviceMemory
-main: error: failed to load model '/home/kyuz0/models/gemma-3/BF16/gemma-3-27b-it-BF16-00001-of-00002.gguf'
-✖ ! [vulkan_amdvlk] gemma-3-27b-it-BF16-00001-of-00002__fa1 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index fb6fd18..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-ggml_vulkan: Device memory allocation of size 2819260416 failed.
-ggml_vulkan: Requested buffer size exceeds device buffer size limit: ErrorOutOfDeviceMemory
-main: error: failed to load model '/home/kyuz0/models/gemma-3/BF16/gemma-3-27b-it-BF16-00001-of-00002.gguf'
-✖ ! [vulkan_amdvlk] gemma-3-27b-it-BF16-00001-of-00002__fa1 __longctx16384 failed (exit 0)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 6df7e0c..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 |           pp512 |        107.99 ± 1.50 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 |           tg128 |          3.93 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index 8d0bace..0000000
--- a/benchmark/results/20-12-2025/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |         67.01 ± 0.00 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |          3.76 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index af8f3fc..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2767.54 ± 1.34 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         81.17 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 825bba8..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1439.14 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         68.99 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index b5d1673..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2762.69 ± 4.25 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         81.08 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index c143312..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1442.24 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         68.87 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha__fa1.log
deleted file mode 100644
index 6e0df52..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |      2759.74 ± 13.26 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         84.86 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index c304f88..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1331.68 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         71.38 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index fb3a67c..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |      2748.02 ± 15.88 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         84.54 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 54a7f4d..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1341.74 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         71.34 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index 73eee2d..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2907.52 ± 4.15 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         78.61 ± 0.04 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 26bd649..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1365.96 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         66.95 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 069c775..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2902.86 ± 2.84 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         78.68 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 93d8e93..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1384.06 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         66.93 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1.log
deleted file mode 100644
index 8d4790f..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |      2868.25 ± 16.39 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         80.93 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index bc9a6e4..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1413.39 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         67.93 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 3d1a62f..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |      2874.90 ± 17.97 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         81.07 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index f3ad530..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1414.92 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         68.13 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index bd61453..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |      2812.03 ± 15.70 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         78.66 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index f15de69..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1347.24 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         66.93 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 19e21d1..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |      2820.50 ± 10.20 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         78.66 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 45cfba0..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1390.56 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         67.23 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1__fa1.log
deleted file mode 100644
index ae8df7d..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |      2853.13 ± 21.11 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         81.93 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index e604ec0..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1368.50 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         68.97 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 88706b8..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |      2846.23 ± 16.40 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         81.96 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 7f6b35d..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1361.15 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         67.99 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index 988efe2..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2843.92 ± 0.49 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         78.68 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 0c8d637..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1377.32 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         66.63 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 0d50d4b..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2841.34 ± 6.05 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         78.55 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 585ce4f..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1340.85 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         67.21 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc__fa1.log
deleted file mode 100644
index edb6f5d..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2888.98 ± 3.97 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         81.99 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index d5bf76c..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1300.12 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         65.93 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index 6d438b1..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2874.43 ± 2.44 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         81.99 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index be4afd5..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1332.98 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         67.59 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1.log
deleted file mode 100644
index b62a85a..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1798.72 ± 4.50 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |           tg128 |         90.73 ± 0.09 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index 24bb270..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        466.89 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         74.73 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1.log
deleted file mode 100644
index a154a01..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1633.15 ± 4.31 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |           tg128 |         85.91 ± 0.19 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index 42fcd8d..0000000
--- a/benchmark/results/20-12-2025/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        776.82 ± 0.00 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         58.76 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index 2252fb8..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        274.17 ± 2.38 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.85 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index a29ee58..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        303.64 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         29.04 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index ee65108..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        276.82 ± 4.42 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.95 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index fdb5447..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        303.82 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         29.04 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha__fa1.log
deleted file mode 100644
index 2d33a3b..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        274.29 ± 0.47 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         52.18 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index 48bc097..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        301.57 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         42.81 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index b27b11c..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        274.47 ± 1.46 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         52.23 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 46fc918..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        289.74 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         42.82 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index 7a833ac..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        657.63 ± 7.64 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.65 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 5159dd1..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        479.02 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         28.15 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 9ed737f..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |       649.55 ± 10.69 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.68 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index b3e1604..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        480.59 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         28.88 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index 303a0c0..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        654.41 ± 2.17 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.37 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index a44f435..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        470.66 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         45.00 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index aac1650..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        658.64 ± 9.76 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.93 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 4d3497f..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        472.41 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         45.26 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index dfe030e..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |       637.48 ± 24.73 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.66 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 9ce2672..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        412.07 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         28.91 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index a164176..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |       653.60 ± 10.65 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.79 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 315116f..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        409.77 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         28.99 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__fa1.log
deleted file mode 100644
index 79ce413..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        639.37 ± 5.37 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.13 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index 6d3d42a..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        496.89 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         43.73 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 0a274b3..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        646.49 ± 8.17 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.16 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 81ff7b2..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        501.87 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         44.01 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index 30be698..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |       645.91 ± 13.82 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.67 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index cc45722..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        410.91 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         28.97 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index ec1ce05..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        653.08 ± 2.58 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.78 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 43778f2..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        413.56 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         28.99 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc__fa1.log
deleted file mode 100644
index 8113168..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |       637.84 ± 10.76 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.22 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 9c34c8d..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        522.65 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         43.99 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index a93d1d4..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        642.63 ± 7.27 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.17 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 4db2edc..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        532.69 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         44.03 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index a1dd94b..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           pp512 |        792.77 ± 0.78 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           tg128 |         52.34 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index 123da16..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        266.30 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         39.24 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index 9ffcf99..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           pp512 |        543.66 ± 0.88 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           tg128 |         56.37 ± 0.04 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index 3dc745b..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        299.40 ± 0.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         45.88 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index 4326475..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |        791.20 ± 5.92 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.24 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 381be35..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        539.70 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         42.26 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index fc16f14..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |        795.34 ± 8.45 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.28 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 7c903ac..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        538.62 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         42.21 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha__fa1.log
deleted file mode 100644
index 093a57c..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |       788.07 ± 19.38 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.65 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index 5e1b2bb..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        585.34 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         61.13 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index 59cb88e..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |       777.22 ± 12.74 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.68 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 45a310e..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        533.90 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         61.16 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index 5486142..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1752.24 ± 12.09 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.82 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 5552e25..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        796.69 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         42.10 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 86bf607..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1717.36 ± 12.37 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.94 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 6dcbab1..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        796.57 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         42.08 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log
deleted file mode 100644
index 62ba657..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1710.01 ± 23.22 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.48 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index ae120ce..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        808.14 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         64.31 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index e64cdc4..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |       1726.91 ± 4.81 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.44 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index e298cff..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        817.70 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         64.37 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index 7ff138e..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |       1710.28 ± 7.42 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.12 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index e60eec1..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        670.09 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         42.17 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 29b33c0..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1695.75 ± 25.43 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.15 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 556c411..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        668.65 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         42.04 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1__fa1.log
deleted file mode 100644
index 1530264..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1670.49 ± 30.36 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.45 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index e0cabe3..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        658.07 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         62.69 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 00e6a9e..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |       1661.92 ± 6.16 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.41 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index fd78ca6..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        654.36 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         62.59 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index 2763926..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |       1692.63 ± 8.33 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.09 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index a63519e..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        668.58 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         42.12 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index 49df202..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1721.79 ± 15.21 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.01 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 1b523b7..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        665.87 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         42.13 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc__fa1.log
deleted file mode 100644
index fd42fa3..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1668.78 ± 30.58 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.33 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 6c3bf6a..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        633.71 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         62.64 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index 67ac385..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1674.84 ± 21.95 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.29 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index fc067d5..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        644.71 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         62.66 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log
deleted file mode 100644
index eff8190..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           pp512 |      1965.23 ± 21.66 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           tg128 |         75.24 ± 0.04 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index de60c57..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        512.34 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         56.91 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log
deleted file mode 100644
index 297b848..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           pp512 |      1395.08 ± 16.05 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           tg128 |         79.60 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index 214c4a8..0000000
--- a/benchmark/results/20-12-2025/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        467.70 ± 0.00 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         65.33 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha-rocwmma__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha-rocwmma__fa1.log
deleted file mode 100644
index 71b27ca..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1556.97 ± 0.78 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.65 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 8b6d29a..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         94.98 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         10.03 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha-rocwmma__hblt0__fa1.log
deleted file mode 100644
index c8ebbe3..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1561.25 ± 2.77 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.59 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index c0ef94b..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |         94.41 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         10.04 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha__fa1.log
deleted file mode 100644
index 32f721d..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1572.74 ± 2.60 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.65 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha__fa1__longctx16384.log
deleted file mode 100644
index a5326b0..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        351.26 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         10.04 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha__hblt0__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha__hblt0__fa1.log
deleted file mode 100644
index 3bbc163..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1572.88 ± 2.44 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.64 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 05bda17..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm-7alpha__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        346.35 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         10.03 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4-rocwmma__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4-rocwmma__fa1.log
deleted file mode 100644
index 8bea5c1..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1525.39 ± 0.85 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.05 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index eb58868..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        104.91 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.08 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4-rocwmma__hblt0__fa1.log
deleted file mode 100644
index adc47d5..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1524.22 ± 2.19 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.02 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index b41ed3a..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        106.82 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.07 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4__fa1.log
deleted file mode 100644
index 8074b6d..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1539.28 ± 0.84 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.02 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx16384.log
deleted file mode 100644
index 5adfe23..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        384.07 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.06 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 47a5836..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1538.89 ± 3.35 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.07 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 52dabe0..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        382.87 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         12.07 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1-rocwmma__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1-rocwmma__fa1.log
deleted file mode 100644
index e534712..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1492.67 ± 1.40 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.89 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index be3abc4..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        141.66 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         13.37 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1-rocwmma__hblt0__fa1.log
deleted file mode 100644
index d3c244e..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1495.58 ± 2.18 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.97 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index f7d0ab4..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        141.05 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         13.35 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1__fa1.log
deleted file mode 100644
index 0d4f6c6..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1526.32 ± 2.10 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.96 ± 0.03 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1__fa1__longctx16384.log
deleted file mode 100644
index a4690c7..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        348.77 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         13.34 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1__hblt0__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1__hblt0__fa1.log
deleted file mode 100644
index 5226879..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1530.07 ± 0.42 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.01 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1__hblt0__fa1__longctx16384.log
deleted file mode 100644
index e50b6d7..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7.1.1__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        348.73 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         13.36 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc-rocwmma__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc-rocwmma__fa1.log
deleted file mode 100644
index af74035..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc-rocwmma__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1493.41 ± 1.54 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.90 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc-rocwmma__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc-rocwmma__fa1__longctx16384.log
deleted file mode 100644
index 7919346..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc-rocwmma__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        139.87 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         13.35 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc-rocwmma__hblt0__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc-rocwmma__hblt0__fa1.log
deleted file mode 100644
index a461503..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc-rocwmma__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1494.79 ± 2.85 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.91 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
deleted file mode 100644
index 34ebd8c..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc-rocwmma__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        139.47 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         13.36 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc__fa1.log
deleted file mode 100644
index 9935294..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1529.76 ± 1.36 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.85 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc__fa1__longctx16384.log
deleted file mode 100644
index 5c0b82d..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        350.95 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         13.35 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc__hblt0__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc__hblt0__fa1.log
deleted file mode 100644
index a8d33d1..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc__hblt0__fa1.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1530.14 ± 1.62 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.88 ± 0.01 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc__hblt0__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc__hblt0__fa1__longctx16384.log
deleted file mode 100644
index fda552c..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__rocm7_rc__hblt0__fa1__longctx16384.log
+++ /dev/null
@@ -1,10 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |        347.70 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         13.35 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 8493016..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1494.56 ± 4.36 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           tg128 |         56.03 ± 0.06 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx16384.log
deleted file mode 100644
index 6f58f25..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        196.22 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         15.95 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__vulkan_radv__fa1.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__vulkan_radv__fa1.log
deleted file mode 100644
index 13624db..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1135.49 ± 4.16 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           tg128 |         55.73 ± 0.02 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx16384.log b/benchmark/results/20-12-2025/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx16384.log
deleted file mode 100644
index 549cd93..0000000
--- a/benchmark/results/20-12-2025/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx16384.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d16384 |        294.01 ± 0.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d16384 |         15.16 ± 0.00 |
-
-build: 2aa45ef9e (7423)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index f8dbba3..0000000
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         78.30 ± 0.11 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.98 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 1e98d22..0000000
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         78.37 ± 0.13 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.76 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index fe9db4c..0000000
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         78.63 ± 0.08 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.80 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index a0c8e97..0000000
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         17.58 ± 0.03 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.08 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index 29499ed..0000000
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         80.49 ± 0.14 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.99 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index dedaf41..0000000
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         21.15 ± 0.06 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.07 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 2223087..0000000
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         78.97 ± 0.09 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.99 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index a4d9675..0000000
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         21.17 ± 0.28 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.07 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 4540fd9..0000000
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,24 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x3c25) [0x7f8003c3bc25]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f8003c3bfeb]
-/lib64/libggml-base.so.0(+0x16669) [0x7f8003c4e669]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f80033b2bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f800339cd3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f80033b2ea8]
-/lib64/libggml-vulkan.so.0(+0x14f76) [0x7f8003cf6f76]
-/lib64/libggml-vulkan.so.0(+0x13597f) [0x7f8003e1797f]
-/lib64/libggml-vulkan.so.0(+0x136411) [0x7f8003e18411]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7f8003c574d3]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f800756ce70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f800756f445]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f8007575aaf]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f800757742e]
-/usr/sbin/llama-bench() [0x41cc3b]
-/usr/sbin/llama-bench() [0x41977f]
-/lib64/libc.so.6(+0x35b5) [0x7f80030835b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f8003083668]
-/usr/sbin/llama-bench() [0x41b595]
-terminate called after throwing an instance of 'vk::DeviceLostError'
-  what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__fa1 failed (exit 0)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 8d7f3ff..0000000
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,23 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x3c25) [0x7f34c7070c25]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f34c7070feb]
-/lib64/libggml-base.so.0(+0x16669) [0x7f34c7083669]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f34c67e7bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f34c67d1d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f34c67e7ea8]
-/lib64/libggml-vulkan.so.0(+0x16b68) [0x7f34c712db68]
-/lib64/libggml-vulkan.so.0(+0xfcfd0) [0x7f34c7213fd0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f34c708c092]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f34ca9a1e70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f34ca9a4445]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f34ca9aaaaf]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f34ca9ac42e]
-/usr/sbin/llama-bench() [0x41cc3b]
-/usr/sbin/llama-bench() [0x41977f]
-/lib64/libc.so.6(+0x35b5) [0x7f34c64b85b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f34c64b8668]
-/usr/sbin/llama-bench() [0x41b595]
-terminate called after throwing an instance of 'vk::DeviceLostError'
-  what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__fa1 __longctx32768 failed (exit 0)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 1c52acf..0000000
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 |           pp512 |         47.20 ± 0.03 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 |           tg128 |          2.99 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index eb50e64..0000000
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |          7.39 ± 0.02 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          2.60 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index 8bc6dfa..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        435.72 ± 2.11 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.38 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index ec45ef3..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         87.93 ± 0.11 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.33 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 6771408..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        403.83 ± 1.74 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.37 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 28c5c04..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         87.49 ± 0.23 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.33 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index f8af5b7..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        393.19 ± 1.92 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         19.36 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 431595d..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         94.74 ± 0.05 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.06 ± 0.18 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 5c65164..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        393.15 ± 2.00 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.11 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index adc0015..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         94.55 ± 0.30 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.18 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index 3da36d8..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        476.23 ± 2.21 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.36 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 8a819b6..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         86.79 ± 0.15 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.32 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index d0c28f4..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        401.13 ± 5.52 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.35 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 3da0eab..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         85.92 ± 0.13 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.30 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 1400856..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           pp512 |        114.88 ± 0.15 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           tg128 |         10.48 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 0c10034..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         10.15 ± 0.00 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          5.05 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 2c91e29..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           pp512 |        342.11 ± 5.47 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           tg128 |          9.48 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index d0eb79f..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        102.93 ± 0.11 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          7.82 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__fa1.log
deleted file mode 100644
index f49590b..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        905.79 ± 2.84 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         32.96 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index b2ff1c9..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         91.97 ± 0.19 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.58 ± 0.07 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index fcbe701..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |       893.68 ± 28.19 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         33.20 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 0835b60..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         91.83 ± 0.16 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.62 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index 5f26656..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        903.39 ± 1.04 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         31.04 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 8dc2481..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         99.84 ± 0.23 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.33 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index ada6b79..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        906.25 ± 1.64 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         32.57 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 41a8790..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        100.07 ± 0.20 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.35 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log
deleted file mode 100644
index 37c4d87..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |       902.85 ± 59.94 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         33.16 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index d4a1f68..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         89.75 ± 0.31 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.59 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index fd05cd0..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        906.42 ± 2.57 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         33.18 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d6e72bd..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         89.99 ± 0.26 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.60 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index fcd1fbe..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           pp512 |        499.98 ± 1.29 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           tg128 |         39.01 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index e9e723c..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,23 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x3c25) [0x7fa3d16b4c25]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fa3d16b4feb]
-/lib64/libggml-base.so.0(+0x16669) [0x7fa3d16c7669]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7fa3d0e2bbfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7fa3d0e15d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7fa3d0e2bea8]
-/lib64/libggml-vulkan.so.0(+0x16b68) [0x7fa3d1771b68]
-/lib64/libggml-vulkan.so.0(+0xfcfd0) [0x7fa3d1857fd0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7fa3d16d0092]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fa3d4fe5e70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7fa3d4fe8445]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7fa3d4feeaaf]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7fa3d4ff042e]
-/usr/sbin/llama-bench() [0x41cc3b]
-/usr/sbin/llama-bench() [0x419a10]
-/lib64/libc.so.6(+0x35b5) [0x7fa3d0afc5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fa3d0afc668]
-/usr/sbin/llama-bench() [0x41b595]
-terminate called after throwing an instance of 'vk::DeviceLostError'
-  what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] GLM-4.7-Flash-UD-Q8_K_XL__fa1 __longctx32768 failed (exit 0)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index c84fe4a..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           pp512 |        853.46 ± 6.70 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           tg128 |         40.38 ± 1.22 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 3b8a27e..0000000
--- a/benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        133.27 ± 0.03 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         21.17 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1.log
deleted file mode 100644
index 16a5422..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        673.71 ± 1.42 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.35 ± 0.17 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index fe8ad78..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        748.37 ± 4.67 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.42 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index ff41820..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        165.11 ± 1.59 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.08 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log
deleted file mode 100644
index aaf2ed8..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        706.16 ± 3.76 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.39 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index f03361f..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        166.61 ± 1.47 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.05 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 7a1e96f..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        710.61 ± 4.31 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.39 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 4ae9022..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        161.22 ± 2.64 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.06 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log
deleted file mode 100644
index 4bfad54..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        881.45 ± 3.04 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.44 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index e15c3d6..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        213.50 ± 0.70 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.08 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 710a6a1..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        735.98 ± 5.08 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.44 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 8ee4f6b..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        205.29 ± 0.84 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.08 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 47e06e4..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,23 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x3c25) [0x7f9947881c25]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f9947881feb]
-/lib64/libggml-base.so.0(+0x16669) [0x7f9947894669]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f9946ff8bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f9946fe2d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f9946ff8ea8]
-/lib64/libggml-vulkan.so.0(+0x16b68) [0x7f994793eb68]
-/lib64/libggml-vulkan.so.0(+0xfcfd0) [0x7f9947a24fd0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f994789d092]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f994b1b2e70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f994b1b5445]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f994b1bbaaf]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f994b1bd42e]
-/usr/sbin/llama-bench() [0x41cc3b]
-/usr/sbin/llama-bench() [0x419a10]
-/lib64/libc.so.6(+0x35b5) [0x7f9946cc95b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f9946cc9668]
-/usr/sbin/llama-bench() [0x41b595]
-terminate called after throwing an instance of 'vk::DeviceLostError'
-  what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] Ministral-3-14B-Instruct-2512-BF16__fa1 failed (exit 0)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 59b5b4b..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,23 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x3c25) [0x7f4bc4c80c25]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f4bc4c80feb]
-/lib64/libggml-base.so.0(+0x16669) [0x7f4bc4c93669]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f4bc43f7bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f4bc43e1d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f4bc43f7ea8]
-/lib64/libggml-vulkan.so.0(+0x16b68) [0x7f4bc4d3db68]
-/lib64/libggml-vulkan.so.0(+0xfcfd0) [0x7f4bc4e23fd0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f4bc4c9c092]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f4bc85b1e70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f4bc85b4445]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f4bc85baaaf]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f4bc85bc42e]
-/usr/sbin/llama-bench() [0x41cc3b]
-/usr/sbin/llama-bench() [0x41977f]
-/lib64/libc.so.6(+0x35b5) [0x7f4bc40c85b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f4bc40c8668]
-/usr/sbin/llama-bench() [0x41b595]
-terminate called after throwing an instance of 'vk::DeviceLostError'
-  what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] Ministral-3-14B-Instruct-2512-BF16__fa1 __longctx32768 failed (exit 0)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log
deleted file mode 100644
index 8311fc9..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           pp512 |        166.51 ± 0.62 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           tg128 |          7.94 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 645c378..0000000
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         72.65 ± 0.24 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          6.70 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 2d8a89d..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        263.95 ± 1.57 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.02 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index a6f51a2..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        305.52 ± 0.18 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.59 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index c181684..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        260.98 ± 1.56 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.60 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 6d79881..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        324.69 ± 0.27 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.38 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 6c0fee0..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        262.30 ± 1.42 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.86 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index ca1525c..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        324.11 ± 0.41 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.45 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
deleted file mode 100644
index 2b76326..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        228.50 ± 1.06 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.07 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 39e5779..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        284.43 ± 0.38 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.76 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index c5c9ddb..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        236.84 ± 1.32 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.04 ± 0.05 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index db1e7af..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        290.04 ± 0.11 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.76 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index ecc3dbc..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |        140.84 ± 0.27 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         13.97 ± 0.05 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index fe06b04..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        106.21 ± 0.16 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         13.21 ± 0.32 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index cd213bb..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |        190.66 ± 7.11 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         14.41 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 4618d0c..0000000
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        191.12 ± 0.46 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         13.85 ± 0.26 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1.log
deleted file mode 100644
index bebfbdf..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        194.62 ± 1.88 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.64 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 69580a4..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         49.93 ± 0.03 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.41 ± 0.37 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 3945af1..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        194.32 ± 1.84 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.80 ± 0.11 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 51d1402..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         49.97 ± 0.14 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.63 ± 0.11 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index 87dbc2a..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        197.89 ± 1.84 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         12.96 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 18f0388..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         70.32 ± 0.08 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.22 ± 0.11 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 2a8c48b..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        198.04 ± 1.90 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         13.11 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 5d89766..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         70.27 ± 0.05 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.02 ± 0.31 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
deleted file mode 100644
index e3df124..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        197.46 ± 1.16 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.31 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 7c95bbf..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         48.93 ± 0.02 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         10.34 ± 0.41 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index b2bd010..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           pp512 |        194.00 ± 1.99 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.33 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 61b7bff..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         49.33 ± 0.04 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         10.46 ± 0.34 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 865311b..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |           pp512 |        106.76 ± 0.77 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |           tg128 |         16.49 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index fdee904..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         26.54 ± 0.02 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          9.57 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index 9130533..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |           pp512 |        158.81 ± 2.40 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |           tg128 |         17.16 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index b16d990..0000000
--- a/benchmark/results/29-03-2026/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         28.25 ± 0.02 |
-| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         11.49 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index cfd6de4..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        474.42 ± 2.29 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         25.36 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index f391dea..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        189.58 ± 0.12 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         18.53 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 3b26f36..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        479.78 ± 2.72 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         25.37 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 820d3b1..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        180.26 ± 0.37 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         18.54 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index b60bde6..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        556.96 ± 2.81 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         23.05 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index d9f8066..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        238.10 ± 0.11 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         18.26 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 6dbd18b..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        556.80 ± 4.26 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         25.16 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d9525bd..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        237.85 ± 0.37 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         18.46 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index 8f2facd..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        503.77 ± 3.42 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         25.54 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index e21cfdb..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        182.59 ± 0.07 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         18.25 ± 1.21 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index e018831..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |        482.99 ± 0.86 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         25.49 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index dfd3f50..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        175.78 ± 0.14 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         18.94 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 05e1c62..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |        171.39 ± 0.81 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         10.48 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 05e07dc..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         73.62 ± 0.08 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          8.49 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index 7bc1638..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |        369.06 ± 2.58 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |          9.38 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 6636734..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        140.81 ± 0.52 |
-| qwen3moe 30B.A3B BF16          |  56.89 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          8.24 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1.log
deleted file mode 100644
index 8ac5001..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       994.33 ± 11.86 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         55.87 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 1d4eadc..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        209.15 ± 0.11 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         30.88 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 81f42e5..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       994.22 ± 10.06 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         55.84 ± 0.04 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index fc65e7c..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        208.93 ± 0.21 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         30.90 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index 8d03802..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1009.37 ± 9.64 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         53.13 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index e60bbfd..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        285.49 ± 0.18 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         30.53 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 9ab5751..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1012.69 ± 9.02 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         54.94 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 1d13bc6..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        285.25 ± 0.16 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         30.50 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1.log
deleted file mode 100644
index ae4042e..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1023.24 ± 11.08 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         56.92 ± 0.17 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index a37013f..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        206.37 ± 0.03 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         31.69 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 7aedfaa..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1012.11 ± 7.91 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         56.92 ± 0.21 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 4a658bc..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        205.56 ± 0.23 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         31.62 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 9239487..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |        828.53 ± 4.66 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         63.31 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index a200dea..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        110.50 ± 0.14 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         25.87 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index 29ec4d4..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1145.66 ± 9.68 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         68.15 ± 0.11 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 3dfccf2..0000000
--- a/benchmark/results/29-03-2026/Qwen3-30B-A3B-Instruct-2507-UD-Q6_K_XL__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        193.03 ± 0.86 |
-| qwen3moe 30B.A3B Q6_K          |  24.53 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         34.02 ± 0.04 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1.log
deleted file mode 100644
index e6d9faa..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1149.31 ± 13.28 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         67.72 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 31bdd1d..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        212.38 ± 0.33 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         33.49 ± 0.23 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 0031201..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1149.83 ± 7.05 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         68.05 ± 0.68 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 1b4492c..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        212.45 ± 0.21 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         33.52 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log
deleted file mode 100644
index 9642b43..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1170.01 ± 7.53 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         65.12 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index ad3e381..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        291.28 ± 0.14 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         32.65 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index a288f4e..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1167.76 ± 7.70 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         64.97 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index b190589..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        290.12 ± 2.48 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         32.65 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log
deleted file mode 100644
index 12481a0..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1178.27 ± 10.86 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         69.46 ± 0.62 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 95d24a2..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        209.14 ± 0.10 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         34.83 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index ed6c60e..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |      1172.23 ± 12.92 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         69.65 ± 0.62 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 8b694cc..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        209.05 ± 0.22 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         34.76 ± 0.25 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log
deleted file mode 100644
index b1b4b06..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |        882.98 ± 3.84 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         80.84 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 0e42549..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        111.65 ± 0.08 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         28.03 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log
deleted file mode 100644
index 3b18735..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1290.50 ± 7.83 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         83.79 ± 0.18 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 253d842..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        194.26 ± 0.86 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         37.04 ± 0.04 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm-7_2__fa1.log
deleted file mode 100644
index eaf9c62..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        613.58 ± 2.84 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         29.81 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index efca4d7..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        447.94 ± 2.75 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         26.06 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index ba7778f..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |      583.48 ± 105.06 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         29.85 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 4e99c3f..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       439.93 ± 32.65 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         26.09 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index 029c2b7..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        627.46 ± 3.12 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         28.02 ± 0.04 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index be93303..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        527.78 ± 1.05 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         26.22 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index b272219..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |       623.64 ± 17.23 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         29.16 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 9521439..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        512.14 ± 3.00 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         26.33 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
deleted file mode 100644
index 1ce170b..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        622.47 ± 8.34 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         29.93 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 1e6ca64..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       443.00 ± 30.64 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         26.97 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 1407abc..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |       600.08 ± 13.59 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         29.99 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 57f366f..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        457.86 ± 1.53 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         27.08 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index afbab06..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |        396.36 ± 1.71 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         30.90 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index a881bd7..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        208.44 ± 0.64 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         26.08 ± 0.05 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index 728a4b6..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |        500.88 ± 3.30 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         31.74 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 8b67617..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Coder-Next-UD-Q8_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        420.42 ± 0.15 |
-| qwen3next 80B.A3B Q8_0         |  86.94 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         28.05 ± 0.05 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index 2791109..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        629.03 ± 5.19 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         31.37 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index ffe0750..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        439.65 ± 0.69 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         27.25 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index ca5a35f..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        648.58 ± 5.44 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         30.86 ± 0.77 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index dbd302b..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        457.60 ± 1.26 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         27.22 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 4ce1624..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        634.72 ± 6.93 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         29.90 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 07b6bfd..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        511.83 ± 7.51 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         27.45 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index cd84abc..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |       632.78 ± 59.05 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         30.53 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 70d9321..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        507.17 ± 2.44 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         27.44 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index 713a371..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |       621.13 ± 82.64 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         31.52 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index aa85e0e..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        467.56 ± 0.55 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         28.26 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index ee04293..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        661.07 ± 5.80 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         31.54 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index ca2c10d..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        432.86 ± 0.82 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         28.20 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index c349ffd..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |        436.50 ± 7.59 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         34.72 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index e177acc..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        218.02 ± 0.85 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         28.50 ± 0.07 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index f7c91c0..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |       587.86 ± 37.36 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         36.28 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 25f1422..0000000
--- a/benchmark/results/29-03-2026/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        453.76 ± 0.75 |
-| qwen3next 80B.A3B Q8_0         |  86.68 GiB |    79.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         31.67 ± 0.06 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index c769c07..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        313.18 ± 2.32 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         19.62 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 6cacb57..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        243.74 ± 1.45 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         16.51 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index 83fa89a..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        311.68 ± 1.84 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         18.77 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 995277c..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        273.72 ± 1.22 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         17.91 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 9330711..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        306.06 ± 2.55 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         19.25 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 4066e47..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        270.84 ± 0.74 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         17.98 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
deleted file mode 100644
index 0bb4c16..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        314.27 ± 4.13 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         19.66 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 482822d..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        235.12 ± 5.24 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         18.36 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 3c70313..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        310.73 ± 1.65 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         19.61 ± 0.11 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 22ef68a..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        241.56 ± 5.74 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         18.35 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 2aa7628..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |           pp512 |        183.05 ± 1.84 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |           tg128 |         21.31 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 9159adc..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        102.85 ± 0.15 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         18.76 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index ebcc2cb..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |           pp512 |        239.56 ± 7.45 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |           tg128 |         21.68 ± 0.38 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index ba5992f..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        203.34 ± 0.47 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         20.09 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index 88d3daf..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |        508.43 ± 3.18 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         23.65 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 8693fc9..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        382.51 ± 1.54 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         21.63 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 955738e..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |        544.11 ± 3.06 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.40 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 19940ae..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        407.19 ± 1.96 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         21.59 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 23b8db1..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |        550.28 ± 3.67 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         23.23 ± 0.18 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 221ccc6..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        405.13 ± 1.85 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         21.72 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index 3151c98..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       541.57 ± 11.33 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         23.69 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 25444a3..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        441.64 ± 9.63 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         22.18 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 532dfcd..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |        509.57 ± 8.20 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         23.78 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 24b9961..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        379.36 ± 1.33 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         22.20 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index ec9c2e2..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |        122.56 ± 0.40 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         11.56 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index af1901a..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         97.32 ± 0.17 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.95 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index d914309..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |        309.96 ± 4.20 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         10.79 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index a17d6c3..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        258.85 ± 0.77 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.39 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm-7_2__fa1.log
deleted file mode 100644
index 91bf9ac..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |      1071.38 ± 11.20 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         47.78 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 1c89299..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       1093.96 ± 6.37 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         47.95 ± 0.30 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 18626cd..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        661.40 ± 0.76 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         40.13 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index 77bc2b3..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       1079.44 ± 6.76 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         46.46 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 28cdd02..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        762.29 ± 2.68 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         40.46 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index e8d6552..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       1082.35 ± 6.79 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         46.48 ± 0.04 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index e352549..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        777.89 ± 1.08 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         40.24 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1.log
deleted file mode 100644
index f34ef08..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       1092.86 ± 9.42 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         48.16 ± 0.30 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index ca64073..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        655.39 ± 2.00 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         41.90 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 90d466f..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       1113.86 ± 6.42 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         48.10 ± 0.31 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 558b9cf..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        655.89 ± 1.47 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         41.98 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index a947901..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |        661.63 ± 3.14 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         58.16 ± 0.08 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index dcdc814..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        288.86 ± 0.53 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         44.24 ± 0.09 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index bc44011..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |      1013.40 ± 39.22 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         59.13 ± 0.07 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 2945779..0000000
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        673.55 ± 0.64 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         48.93 ± 0.13 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1.log
deleted file mode 100644
index 5893093..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        854.20 ± 6.85 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.20 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index a017aaa..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       392.45 ± 16.97 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         11.63 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index f21710a..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        877.69 ± 1.71 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.19 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 526bc80..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       387.73 ± 18.58 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         11.63 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
deleted file mode 100644
index 797411c..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        869.53 ± 1.47 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.15 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index d664914..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        260.25 ± 4.30 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         11.60 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index e90965b..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        870.35 ± 1.54 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.15 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 829f89e..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        263.85 ± 5.85 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         11.61 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log
deleted file mode 100644
index 116b328..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        892.78 ± 1.04 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.19 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 77f2bd0..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       441.92 ± 38.71 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         11.63 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 2423cc9..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           pp512 |        895.09 ± 0.81 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.19 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 1f6940d..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       449.40 ± 28.98 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         11.63 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
deleted file mode 100644
index a32ca20..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |           pp512 |        122.72 ± 0.06 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |           tg128 |         14.31 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index ba09991..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        107.34 ± 0.03 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         11.32 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1.log
deleted file mode 100644
index c138142..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |           pp512 |       513.78 ± 16.10 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |           tg128 |         14.07 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 785012f..0000000
--- a/benchmark/results/29-03-2026/gemma-3-12b-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        258.63 ± 1.57 |
-| gemma3 12B Q8_0                |  13.40 GiB |    11.77 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         11.53 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index f1ccf93..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        324.28 ± 1.23 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.97 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index b0cb077..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        166.64 ± 2.06 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.70 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 8531f51..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        455.07 ± 0.55 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.97 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 714f6c0..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        188.26 ± 1.59 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.70 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index 47bc0c4..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        426.08 ± 0.81 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.83 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index ca28d62..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        209.69 ± 3.55 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.68 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index f180c49..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        440.33 ± 0.38 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.96 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 5dc5e20..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        202.85 ± 0.64 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.69 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index ea47115..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        466.09 ± 1.03 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.98 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index ceac7ca..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        182.83 ± 1.03 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.68 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 2c5cf1e..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           pp512 |        452.21 ± 1.03 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.98 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index e630a45..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        181.81 ± 3.26 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.70 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index f8267e7..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,23 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x3c25) [0x7f25ac3cbc25]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f25ac3cbfeb]
-/lib64/libggml-base.so.0(+0x16669) [0x7f25ac3de669]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f25abb42bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f25abb2cd3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f25abb42ea8]
-/lib64/libggml-vulkan.so.0(+0x16b68) [0x7f25ac488b68]
-/lib64/libggml-vulkan.so.0(+0xfcfd0) [0x7f25ac56efd0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f25ac3e7092]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f25afcfce70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f25afcff445]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f25afd05aaf]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f25afd0742e]
-/usr/sbin/llama-bench() [0x41cc3b]
-/usr/sbin/llama-bench() [0x419a10]
-/lib64/libc.so.6(+0x35b5) [0x7f25ab8135b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f25ab813668]
-/usr/sbin/llama-bench() [0x41b595]
-terminate called after throwing an instance of 'vk::DeviceLostError'
-  what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] gemma-3-27b-it-BF16-00001-of-00002__fa1 failed (exit 0)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 17cbfe3..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,23 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x3c25) [0x7f5b5347fc25]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f5b5347ffeb]
-/lib64/libggml-base.so.0(+0x16669) [0x7f5b53492669]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f5b52bf6bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f5b52be0d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f5b52bf6ea8]
-/lib64/libggml-vulkan.so.0(+0x16b68) [0x7f5b5353cb68]
-/lib64/libggml-vulkan.so.0(+0xfcfd0) [0x7f5b53622fd0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f5b5349b092]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f5b56db0e70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f5b56db3445]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f5b56db9aaf]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f5b56dbb42e]
-/usr/sbin/llama-bench() [0x41cc3b]
-/usr/sbin/llama-bench() [0x41977f]
-/lib64/libc.so.6(+0x35b5) [0x7f5b528c75b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f5b528c7668]
-/usr/sbin/llama-bench() [0x41b595]
-terminate called after throwing an instance of 'vk::DeviceLostError'
-  what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] gemma-3-27b-it-BF16-00001-of-00002__fa1 __longctx32768 failed (exit 0)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index be2d8b8..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 |           pp512 |         91.65 ± 0.40 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 |           tg128 |          3.98 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 8fe70f2..0000000
--- a/benchmark/results/29-03-2026/gemma-3-27b-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         81.52 ± 0.33 |
-| gemma3 27B BF16                |  50.31 GiB |    27.01 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          3.73 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1.log
deleted file mode 100644
index 92de1c6..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2700.94 ± 1.79 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         77.69 ± 0.66 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index ab19336..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      1564.27 ± 18.55 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         56.80 ± 0.48 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 76309ba..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2691.00 ± 4.60 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         77.86 ± 0.64 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 85d1206..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      1563.05 ± 19.77 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         56.95 ± 0.08 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1.log
deleted file mode 100644
index 7fd54a5..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2719.77 ± 6.47 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.32 ± 0.04 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index a88f066..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      1687.79 ± 33.87 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         54.09 ± 0.16 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index ba6263c..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2713.06 ± 9.07 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.85 ± 0.04 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index ea96dc2..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      1681.29 ± 17.71 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         54.11 ± 0.29 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1.log
deleted file mode 100644
index f37ea3e..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2666.49 ± 3.29 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         78.31 ± 0.89 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 769b2bf..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      1530.18 ± 18.45 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         57.72 ± 0.36 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 0e4d278..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           pp512 |       2674.53 ± 5.86 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |  1 |    0 |           tg128 |         78.93 ± 0.64 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d6379ae..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      1534.58 ± 20.23 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         57.65 ± 0.44 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 9325e89..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |           pp512 |        647.40 ± 0.53 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |           tg128 |         79.07 ± 0.08 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index e54640a..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        455.63 ± 1.03 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         54.86 ± 0.17 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1.log
deleted file mode 100644
index ed9c6f1..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |           pp512 |      2479.97 ± 41.40 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |           tg128 |         87.24 ± 0.13 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index c008393..0000000
--- a/benchmark/results/29-03-2026/gemma-3-4b-it-Q3_K_S__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |       1759.67 ± 6.72 |
-| gemma3 4B Q3_K - Small         |   1.80 GiB |     3.88 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         64.91 ± 0.04 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1.log
deleted file mode 100644
index 83045a0..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        631.59 ± 4.27 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.91 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index dd99083..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        303.63 ± 0.57 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         36.07 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 8ff8feb..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |       610.41 ± 53.09 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.34 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index d0b590a..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        302.82 ± 1.30 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         36.12 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log
deleted file mode 100644
index 492fecb..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        648.44 ± 6.33 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         49.85 ± 0.04 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index c6c534c..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        442.64 ± 0.84 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         39.69 ± 0.34 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index 8d1b5bf..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |       639.43 ± 31.93 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.99 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index c1bacb6..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        438.75 ± 1.06 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         39.75 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log
deleted file mode 100644
index d70f3d4..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |       596.69 ± 97.42 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.38 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 9670ec2..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        308.13 ± 1.66 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         35.43 ± 8.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 0857b4a..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |       641.07 ± 11.17 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.35 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index a189399..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        314.50 ± 0.28 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         40.04 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log
deleted file mode 100644
index e8675e6..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           pp512 |        576.81 ± 2.43 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           tg128 |         51.18 ± 0.04 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 315ed48..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        213.74 ± 0.68 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         34.52 ± 0.07 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log
deleted file mode 100644
index c0ad41d..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           pp512 |       633.21 ± 13.06 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           tg128 |         56.15 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index fae75b7..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        287.49 ± 1.21 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         42.67 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1.log
deleted file mode 100644
index 5e219c6..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1638.53 ± 13.50 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.67 ± 0.07 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index ca3a5f0..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        488.89 ± 0.54 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         51.91 ± 0.03 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index 6094846..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1631.29 ± 15.38 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.62 ± 0.09 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index f2a466a..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        492.09 ± 1.86 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         51.93 ± 0.04 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log
deleted file mode 100644
index a7cf323..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1691.01 ± 16.68 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.07 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 45bb062..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        730.11 ± 1.16 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         57.02 ± 0.07 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index a3bdd9d..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1690.30 ± 13.53 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.00 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index edcc380..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        730.49 ± 1.20 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         56.89 ± 0.05 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log
deleted file mode 100644
index 932a36a..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1668.50 ± 13.61 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.68 ± 0.10 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 65c4caa..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        507.77 ± 2.81 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         57.46 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index 50262df..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1642.70 ± 14.12 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.75 ± 0.10 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index f425c83..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        507.84 ± 1.97 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         57.32 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log
deleted file mode 100644
index fa14b74..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1303.99 ± 6.84 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           tg128 |         73.68 ± 0.08 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 99d5ac7..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        364.73 ± 0.33 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         49.99 ± 0.21 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log
deleted file mode 100644
index 917f775..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           pp512 |      1577.96 ± 12.64 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           tg128 |         78.94 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 97ddca5..0000000
--- a/benchmark/results/29-03-2026/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        541.12 ± 0.20 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         60.77 ± 0.08 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm-7_2__fa1.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm-7_2__fa1.log
deleted file mode 100644
index 99ce5e7..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1543.27 ± 2.63 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.61 ± 0.14 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm-7_2__fa1__longctx32768.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm-7_2__fa1__longctx32768.log
deleted file mode 100644
index 0dc8ee6..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm-7_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        165.68 ± 0.88 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.65 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1.log
deleted file mode 100644
index dd483cb..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1539.48 ± 5.61 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.57 ± 0.15 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 5a66ad0..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm-7_2__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        166.31 ± 1.47 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.65 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1.log
deleted file mode 100644
index 5103c84..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1528.37 ± 6.40 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.14 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index e0b84ff..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        196.35 ± 2.62 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.97 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1.log
deleted file mode 100644
index acaa1cf..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1521.12 ± 5.74 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.22 ± 0.02 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 15d1b59..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm6_4_4__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        196.03 ± 1.28 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.97 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log
deleted file mode 100644
index 6e1ba6a..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1507.68 ± 1.23 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.41 ± 0.14 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index c083ed2..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        206.97 ± 1.96 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.62 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1.log
deleted file mode 100644
index c51b8f4..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1508.58 ± 2.01 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.44 ± 0.15 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1__longctx32768.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1__longctx32768.log
deleted file mode 100644
index 25dcbb6..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__rocm7-nightlies__hblt0__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        206.54 ± 0.91 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.62 ± 0.00 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log
deleted file mode 100644
index a1c3881..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           pp512 |        324.82 ± 0.45 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           tg128 |         55.43 ± 0.14 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 0c19fe3..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        144.29 ± 1.09 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.20 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__vulkan_radv__fa1.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__vulkan_radv__fa1.log
deleted file mode 100644
index f0e5a2f..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1313.97 ± 1.29 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           tg128 |         55.59 ± 0.05 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log b/benchmark/results/29-03-2026/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 50c5ec7..0000000
--- a/benchmark/results/29-03-2026/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        208.18 ± 2.24 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          8.52 ± 0.01 |
-
-build: 2405d59cb (8577)
diff --git a/benchmark/results/29-03-2026/system_info.json b/benchmark/results/29-03-2026/system_info.json
deleted file mode 100644
index 2efeb7c..0000000
--- a/benchmark/results/29-03-2026/system_info.json
+++ /dev/null
@@ -1 +0,0 @@
-{"distro": "Fedora Linux 43 (Workstation Edition)", "kernel": "6.19.9-200.fc43.x86_64", "linux_firmware": "linux-firmware-20260309-1.fc43.noarch", "timestamp": "29 Mar 2026"}
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 486713e..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         72.06 ± 0.03 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.99 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 795fa30..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         16.76 ± 0.05 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.08 ± 0.01 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1.log
deleted file mode 100644
index 93d71b6..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         77.83 ± 0.10 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.95 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index 716d737..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         17.10 ± 0.04 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.06 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log
deleted file mode 100644
index f2036b6..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         78.40 ± 0.10 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.76 ± 0.00 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index d77eb38..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         17.57 ± 0.05 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.09 ± 0.01 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
deleted file mode 100644
index 2a7f425..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         78.06 ± 0.06 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.95 ± 0.00 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index b6b5178..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         18.34 ± 0.14 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.06 ± 0.01 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
deleted file mode 100644
index 5a74d0f..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |          8.45 ± 0.02 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          1.43 ± 0.02 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 7e0aed4..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,24 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7f656fe92465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f656fe9282b]
-/lib64/libggml-base.so.0(+0x16ed9) [0x7f656fea4ed9]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f656f608bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f656f5f2d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f656f608ea8]
-/lib64/libggml-vulkan.so.0(+0x157c0) [0x7f656ff607c0]
-/lib64/libggml-vulkan.so.0(+0x13fa8f) [0x7f657008aa8f]
-/lib64/libggml-vulkan.so.0(+0x140521) [0x7f657008b521]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7f656feae553]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f6573ac9530]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f6573acbb05]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f6573ad226f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f6573ad3bee]
-/usr/sbin/llama-bench() [0x41a1bb]
-/usr/sbin/llama-bench() [0x416c49]
-/lib64/libc.so.6(+0x35b5) [0x7f656f2d95b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f656f2d9668]
-/usr/sbin/llama-bench() [0x4189d5]
-terminate called after throwing an instance of 'vk::DeviceLostError'
-  what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__fa1 failed (exit 0)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 1ff0669..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,23 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7f178fa50465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f178fa5082b]
-/lib64/libggml-base.so.0(+0x16ed9) [0x7f178fa62ed9]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f178f1c6bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f178f1b0d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f178f1c6ea8]
-/lib64/libggml-vulkan.so.0(+0x173b2) [0x7f178fb203b2]
-/lib64/libggml-vulkan.so.0(+0x1075d0) [0x7f178fc105d0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f178fa6c112]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f1793687530]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f1793689b05]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f179369026f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f1793691bee]
-/usr/sbin/llama-bench() [0x41a1bb]
-/usr/sbin/llama-bench() [0x416c49]
-/lib64/libc.so.6(+0x35b5) [0x7f178ee975b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f178ee97668]
-/usr/sbin/llama-bench() [0x4189d5]
-terminate called after throwing an instance of 'vk::DeviceLostError'
-  what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__fa1 __longctx32768 failed (exit 0)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
deleted file mode 100644
index 2d1f68b..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
+++ /dev/null
@@ -1,23 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7f69acb1c465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f69acb1c83b]
-/lib64/libggml-base.so.0(+0x16f19) [0x7f69acb2ef19]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f69ac8bfbfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f69ac8a9d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f69ac8bfea8]
-/lib64/libggml-vulkan.so.0(+0x1728d) [0x7f69acbf128d]
-/lib64/libggml-vulkan.so.0(+0x10a410) [0x7f69acce4410]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f69acb38192]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f69b07a6c70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f69b07a9255]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f69b07af98f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f69b07b132e]
-/usr/sbin/llama-bench() [0x40663b]
-/usr/sbin/llama-bench() [0x4038b9]
-/lib64/libc.so.6(+0x35b5) [0x7f69ac5905b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f69ac590668]
-/usr/sbin/llama-bench() [0x404e65]
-terminate called after throwing an instance of 'vk::DeviceLostError'
-  what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__fa1 __longctx65536 failed (exit 0)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log
deleted file mode 100644
index b4ac940..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 |           pp512 |         46.91 ± 0.02 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 |           tg128 |          2.99 ± 0.00 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index cc9c080..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |          7.32 ± 0.01 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          2.59 ± 0.00 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx65536.log b/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx65536.log
deleted file mode 100644
index 9523902..0000000
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx65536.log
+++ /dev/null
@@ -1,25 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-radv/amdgpu: The CS has been cancelled because the context is lost. This context is innocent.
-/lib64/libggml-base.so.0(+0x4465) [0x7fd3fbd2c465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fd3fbd2c83b]
-/lib64/libggml-base.so.0(+0x16f19) [0x7fd3fbd3ef19]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7fd3fbacfbfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7fd3fbab9d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7fd3fbacfea8]
-/lib64/libggml-vulkan.so.0(+0x1569b) [0x7fd3fbdff69b]
-/lib64/libggml-vulkan.so.0(+0x14505a) [0x7fd3fbf2f05a]
-/lib64/libggml-vulkan.so.0(+0x145c31) [0x7fd3fbf2fc31]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7fd3fbd485d3]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fd3ff9b6c70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7fd3ff9b9255]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7fd3ff9bf98f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7fd3ff9c132e]
-/usr/sbin/llama-bench() [0x40663b]
-/usr/sbin/llama-bench() [0x403a5b]
-/lib64/libc.so.6(+0x35b5) [0x7fd3fb7a05b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fd3fb7a0668]
-/usr/sbin/llama-bench() [0x404e65]
-terminate called after throwing an instance of 'vk::DeviceLostError'
-  what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_radv] Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__fa1 __longctx65536 failed (exit 0)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index e284659..0000000
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        403.45 ± 2.11 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.04 ± 0.01 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index d97ca66..0000000
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         87.65 ± 0.31 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.25 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index 6d03d4a..0000000
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         46.54 ± 0.17 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         11.91 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1.log
deleted file mode 100644
index d1f71ff..0000000
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,2 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index ed83496..0000000
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         87.91 ± 0.20 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         14.84 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 503a469..0000000
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         46.47 ± 0.09 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         11.63 ± 0.19 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1.log
index b419d8c..02e945a 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        403.20 ± 2.04 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         20.43 ± 0.00 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        422.28 ± 1.78 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         23.17 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log
index 508de56..630aead 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         87.63 ± 0.21 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.31 ± 0.02 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         91.33 ± 0.25 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         16.56 ± 0.01 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log
index cf9cc2e..eb8984d 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         46.45 ± 0.01 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         11.98 ± 0.00 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         48.38 ± 0.18 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         12.90 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1.log
deleted file mode 100644
index d1f71ff..0000000
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1.log
+++ /dev/null
@@ -1,2 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log
index 9250782..691ac41 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        392.81 ± 1.96 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         18.97 ± 0.00 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        407.51 ± 1.36 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.00 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
index 4662794..d3ed5fa 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         94.55 ± 0.17 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         14.89 ± 0.01 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         97.18 ± 0.20 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         16.38 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
index 563fcbe..2db9935 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         50.95 ± 0.07 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         11.15 ± 0.09 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         51.90 ± 0.19 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         12.82 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log
index 5735731..09783e2 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        400.57 ± 4.10 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         19.39 ± 0.00 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        424.65 ± 3.93 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         23.30 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
index ef19e5d..0ea8bb1 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         86.01 ± 0.04 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         14.86 ± 0.03 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         90.71 ± 0.38 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         16.55 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
index 8f535c1..f2abb61 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         45.39 ± 0.36 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         11.22 ± 0.13 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         47.61 ± 0.13 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         12.92 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
index e2f4794..70128c6 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           pp512 |        114.94 ± 0.15 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           tg128 |         10.48 ± 0.01 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           pp512 |        115.47 ± 0.14 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           tg128 |         10.78 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
index 2dc96f2..708dcf2 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         10.14 ± 0.00 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          5.04 ± 0.00 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         10.42 ± 0.00 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          5.40 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
index 3dd0992..d1ac5ad 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,24 +1,24 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7fa5c507b465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fa5c507b83b]
-/lib64/libggml-base.so.0(+0x16f19) [0x7fa5c508df19]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7fa5c4e1ebfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7fa5c4e08d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7fa5c4e1eea8]
-/lib64/libggml-vulkan.so.0(+0x1569b) [0x7fa5c514e69b]
-/lib64/libggml-vulkan.so.0(+0x14505a) [0x7fa5c527e05a]
-/lib64/libggml-vulkan.so.0(+0x145c31) [0x7fa5c527ec31]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7fa5c50975d3]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fa5c8d05c70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7fa5c8d08255]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7fa5c8d0e98f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7fa5c8d1032e]
-/usr/sbin/llama-bench() [0x40663b]
-/usr/sbin/llama-bench() [0x403a5b]
-/lib64/libc.so.6(+0x35b5) [0x7fa5c4aef5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fa5c4aef668]
-/usr/sbin/llama-bench() [0x404e65]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7f4479108465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f447910883b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7f447911af59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7f4478eabbfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f4478e95d3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7f4478eabea8]
+/lib64/libggml-vulkan.so.0(+0x1381f) [0x7f44791d981f]
+/lib64/libggml-vulkan.so.0(+0x139c0a) [0x7f44792ffc0a]
+/lib64/libggml-vulkan.so.0(+0x13a7e1) [0x7f44793007e1]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7f4479124613]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f447cade9d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f447cae15b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7f447cae7f95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7f447cae9aae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403c06]
+/lib64/libc.so.6(+0x35b5) [0x7f4478b7c5b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7f4478b7c668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
 ✖ ! [vulkan_amdvlk] GLM-4.7-Flash-BF16-00001-of-00002__fa1 __longctx65536 failed (exit 0)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log
index a0c45c2..7002593 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           pp512 |        326.36 ± 4.62 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           tg128 |          9.50 ± 0.02 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           pp512 |        361.24 ± 1.76 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           tg128 |          9.46 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
index 8e29b1a..e653558 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        102.00 ± 0.06 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          7.82 ± 0.02 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        109.96 ± 0.16 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          7.96 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log
index 094219e..06a0f51 100644
--- a/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         61.52 ± 0.01 |
-| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          6.57 ± 0.00 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         65.21 ± 0.03 |
+| deepseek2 30B.A3B BF16         |  55.79 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          6.90 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 0252058..0000000
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |       1058.10 ± 2.19 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         33.18 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 6a9e2f9..0000000
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         91.21 ± 1.08 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.62 ± 0.01 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index 2445d7c..0000000
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         47.63 ± 0.04 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         14.98 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1.log
deleted file mode 100644
index 79c3029..0000000
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |       889.16 ± 36.98 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         32.61 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index b24dbc6..0000000
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         91.01 ± 1.51 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.38 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 0d48067..0000000
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         47.62 ± 0.15 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         14.87 ± 0.00 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1.log
index e64c344..39e42ba 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        905.99 ± 2.05 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         32.65 ± 0.00 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |       913.36 ± 47.37 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         35.50 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
index 7f43dd2..451f5ba 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         91.59 ± 0.16 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.36 ± 0.00 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         95.82 ± 0.14 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         21.96 ± 0.02 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
index b6a4eed..9b2cbce 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         47.51 ± 0.05 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         14.86 ± 0.00 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         49.39 ± 0.15 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         15.86 ± 0.20 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1.log
index 45a93b4..27f36de 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        903.01 ± 2.32 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         30.77 ± 0.00 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        929.42 ± 1.08 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         33.78 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
index 84cc9c5..db73952 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         99.70 ± 0.23 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         19.93 ± 0.00 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        102.84 ± 0.07 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         21.41 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
index 8344413..395fb88 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         51.79 ± 0.10 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         14.36 ± 0.00 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         53.18 ± 0.20 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         15.67 ± 0.10 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log
index a2cc15a..facd1b7 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |       888.70 ± 38.05 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         32.81 ± 0.00 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           pp512 |        931.06 ± 6.95 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |  1 |    0 |           tg128 |         35.52 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
index ac18d67..87c0228 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         89.53 ± 0.09 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.47 ± 0.00 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         94.94 ± 0.61 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         22.00 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
index 0d230ee..7d1f664 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         46.43 ± 0.14 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         14.43 ± 0.00 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         48.90 ± 0.07 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         15.99 ± 0.01 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
index 57fea0d..28b8c9b 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           pp512 |        499.74 ± 1.49 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           tg128 |         39.02 ± 0.02 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           pp512 |        543.44 ± 1.10 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           tg128 |         39.66 ± 0.02 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
index 7615176..5a7ec4f 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,24 +1,23 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7f03039c2465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f03039c282b]
-/lib64/libggml-base.so.0(+0x16ed9) [0x7f03039d4ed9]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f0303138bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f0303122d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f0303138ea8]
-/lib64/libggml-vulkan.so.0(+0x157c0) [0x7f0303a907c0]
-/lib64/libggml-vulkan.so.0(+0x13f96a) [0x7f0303bba96a]
-/lib64/libggml-vulkan.so.0(+0x140521) [0x7f0303bbb521]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7f03039de553]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f03075f9530]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f03075fbb05]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f030760226f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f0307603bee]
-/usr/sbin/llama-bench() [0x41a1bb]
-/usr/sbin/llama-bench() [0x416ec6]
-/lib64/libc.so.6(+0x35b5) [0x7f0302e095b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f0302e09668]
-/usr/sbin/llama-bench() [0x4189d5]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7fecbc5ac465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fecbc5ac83b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7fecbc5bef59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7fecbc34fbfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7fecbc339d3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7fecbc34fea8]
+/lib64/libggml-vulkan.so.0(+0x15411) [0x7fecbc67f411]
+/lib64/libggml-vulkan.so.0(+0xfefd0) [0x7fecbc768fd0]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7fecbc5c81d2]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fecbff829d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7fecbff855b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7fecbff8bf95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7fecbff8daae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403c06]
+/lib64/libc.so.6(+0x35b5) [0x7fecbc0205b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7fecbc020668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
 ✖ ! [vulkan_amdvlk] GLM-4.7-Flash-UD-Q8_K_XL__fa1 __longctx32768 failed (exit 0)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
index 8abe7d5..fec9079 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,23 +1,23 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7fb35126c465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fb35126c83b]
-/lib64/libggml-base.so.0(+0x16f19) [0x7fb35127ef19]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7fb35100fbfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7fb350ff9d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7fb35100fea8]
-/lib64/libggml-vulkan.so.0(+0x1728d) [0x7fb35134128d]
-/lib64/libggml-vulkan.so.0(+0x10a410) [0x7fb351434410]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7fb351288192]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fb354ef6c70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7fb354ef9255]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7fb354eff98f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7fb354f0132e]
-/usr/sbin/llama-bench() [0x40663b]
-/usr/sbin/llama-bench() [0x403a5b]
-/lib64/libc.so.6(+0x35b5) [0x7fb350ce05b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fb350ce0668]
-/usr/sbin/llama-bench() [0x404e65]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7fcec5e44465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fcec5e4483b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7fcec5e56f59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7fcec5be7bfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7fcec5bd1d3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7fcec5be7ea8]
+/lib64/libggml-vulkan.so.0(+0x15411) [0x7fcec5f17411]
+/lib64/libggml-vulkan.so.0(+0xfefd0) [0x7fcec6000fd0]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7fcec5e601d2]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fcec981a9d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7fcec981d5b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7fcec9823f95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7fcec9825aae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403b2b]
+/lib64/libc.so.6(+0x35b5) [0x7fcec58b85b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7fcec58b8668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
 ✖ ! [vulkan_amdvlk] GLM-4.7-Flash-UD-Q8_K_XL__fa1 __longctx65536 failed (exit 0)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log
index fd4bd04..e1ba2c3 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           pp512 |        875.80 ± 2.04 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           tg128 |         40.19 ± 1.39 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1092.07 ± 1.21 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |           tg128 |         41.17 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
index df14c95..52b137d 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        134.10 ± 0.07 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         21.18 ± 0.03 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        136.78 ± 0.02 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         22.55 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
index 7257eea..4f1d641 100644
--- a/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         72.80 ± 0.00 |
-| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         14.39 ± 0.01 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         74.06 ± 0.01 |
+| deepseek2 30B.A3B Q8_0         |  32.70 GiB |    29.94 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         15.70 ± 0.01 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 6151c79..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           pp512 |        356.93 ± 1.86 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.97 ± 0.13 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 73cec88..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         65.86 ± 0.58 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.10 ± 0.04 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index 20d42cf..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,6 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-main: error: failed to load model '/home/kyuz0/models/mini-max-m2.7/UD-Q3_K_S/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003.gguf'
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-✖ ! [rocm-7_2_2-pr21344] MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__fa1 __longctx65536 failed (exit 0)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1.log
deleted file mode 100644
index e2db156..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           pp512 |        236.39 ± 1.24 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.64 ± 0.01 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index 52c1dcd..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         64.53 ± 0.53 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.04 ± 0.17 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 9680e41..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,6 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-main: error: failed to load model '/home/kyuz0/models/mini-max-m2.7/UD-Q3_K_S/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003.gguf'
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-✖ ! [rocm-7_2_2] MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__fa1 __longctx65536 failed (exit 0)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_3__fa1.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_3__fa1.log
index b78725e..e896482 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_3__fa1.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_3__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           pp512 |        236.96 ± 1.25 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.81 ± 0.01 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           pp512 |        243.25 ± 1.32 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           tg128 |         23.11 ± 0.01 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log
index 7ffced2..688571b 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         62.13 ± 0.56 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.15 ± 0.00 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        142.92 ± 0.05 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.19 ± 0.05 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1.log
index 322f9e6..cafc5e6 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           pp512 |        236.56 ± 1.44 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.55 ± 0.00 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           pp512 |        238.97 ± 1.17 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.98 ± 0.02 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
index ffb9df3..191d49b 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         77.34 ± 1.21 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.54 ± 0.09 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        119.83 ± 0.18 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.73 ± 0.18 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1__longctx65536.log
index 05901f6..97dbdbc 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1__longctx65536.log
@@ -1,3 +1,3 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-✖ ! [rocm6_4_4] MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__fa1 __longctx65536 failed (exit 0)
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+✖ [rocm6_4_4] MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__fa1 __longctx65536 failed (exit 137)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm7-nightlies__fa1.log
index 19ff545..7a85604 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm7-nightlies__fa1.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           pp512 |        225.65 ± 0.79 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.44 ± 0.01 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           pp512 |        235.98 ± 0.63 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.13 ± 0.03 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
index 1c157f4..4441b35 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         64.10 ± 0.50 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.35 ± 0.28 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        156.27 ± 0.25 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.45 ± 0.03 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1.log
index e820993..ba92219 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |           pp512 |        202.08 ± 0.31 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |           tg128 |         24.94 ± 0.01 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |           pp512 |        205.56 ± 0.83 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |           tg128 |         25.91 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
index d513da2..1440940 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         31.48 ± 0.09 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.27 ± 0.01 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         30.23 ± 0.05 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.74 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log
index 62960ea..3fade02 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,3 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-✖ ! [vulkan_amdvlk] MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__fa1 __longctx65536 failed (exit 0)
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         13.90 ± 0.02 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          6.67 ± 0.05 |
+
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1.log
index b06e345..3c8a8ed 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |           pp512 |        231.25 ± 0.79 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |           tg128 |         30.70 ± 0.02 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |           pp512 |        242.43 ± 0.92 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |           tg128 |         31.08 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx32768.log
index cecf7a1..e941ced 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         32.00 ± 0.03 |
-| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         14.47 ± 0.04 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         33.86 ± 0.21 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         14.65 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx65536.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx65536.log
index 302dcdd..ebfb237 100644
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx65536.log
@@ -1,3 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-✖ ! [vulkan_radv] MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__fa1 __longctx65536 failed (exit 0)
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         13.79 ± 0.08 |
+| minimax-m2 230B.A10B Q3_K - Small |  87.20 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          8.84 ± 0.09 |
+
+build: 0253fb21f (9187)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm6_4_4__fa1.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm6_4_4__fa1.log
deleted file mode 100644
index 6caf1c2..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           pp512 |        245.87 ± 1.58 |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.72 ± 0.01 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index bed6c9f..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         81.20 ± 1.25 |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.69 ± 0.12 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm7-nightlies__fa1.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm7-nightlies__fa1.log
deleted file mode 100644
index 4f7dba8..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           pp512 |        238.13 ± 1.15 |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.54 ± 0.27 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index 24d6e3f..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         61.41 ± 0.44 |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.64 ± 0.15 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_amdvlk__fa1.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_amdvlk__fa1.log
deleted file mode 100644
index 192b97a..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |           pp512 |        203.65 ± 1.00 |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |           tg128 |         24.07 ± 0.02 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index 621b928..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         31.64 ± 0.19 |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.06 ± 0.00 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_radv__fa1.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_radv__fa1.log
deleted file mode 100644
index 366fe9b..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |           pp512 |        230.72 ± 8.67 |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |           tg128 |         29.48 ± 0.01 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_radv__fa1__longctx32768.log b/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 9c5cc93..0000000
--- a/benchmark/results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         32.06 ± 0.16 |
-| minimax-m2 230B.A10B Q3_K - Medium |  94.93 GiB |   228.69 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         14.12 ± 0.02 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index d1d295f..0000000
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        745.50 ± 9.08 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.42 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 5c893b9..0000000
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        164.19 ± 1.96 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.08 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1.log
deleted file mode 100644
index d702654..0000000
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        732.09 ± 2.49 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.42 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index 2bb5fa9..0000000
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        182.01 ± 2.08 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.07 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 9bdaeeb..0000000
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         80.08 ± 2.35 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          6.11 ± 0.00 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__fa1.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1.log
similarity index 64%
rename from benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__fa1.log
rename to benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1.log
index d804f02..93b98ee 100644
--- a/benchmark/results/09-01-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7.1.1__fa1.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1.log
@@ -1,8 +1,8 @@
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        792.78 ± 1.08 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        801.06 ± 9.05 |
 | mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.50 ± 0.00 |
 
-build: 9c142e3a2 (7670)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1__longctx32768.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1__longctx32768.log
similarity index 79%
rename from benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1__longctx32768.log
rename to benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1__longctx32768.log
index edb719b..a35b9dc 100644
--- a/benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2__fa1__longctx32768.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        160.30 ± 0.21 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.08 ± 0.00 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        349.46 ± 1.14 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.12 ± 0.00 |
 
-build: 2405d59cb (8577)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1__longctx65536.log
similarity index 79%
rename from benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1__longctx65536.log
rename to benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1__longctx65536.log
index c268707..90b0de8 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         75.02 ± 1.56 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          6.11 ± 0.00 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        209.94 ± 1.05 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          6.14 ± 0.00 |
 
-build: 7957de9dc (8645)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log
index 7303ffc..b92fa3f 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        716.84 ± 2.31 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.39 ± 0.00 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        762.49 ± 6.93 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.49 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log
index 644eac3..2a17883 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        166.12 ± 0.76 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.05 ± 0.00 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        318.28 ± 0.86 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.11 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx65536.log
index b1347cd..930d54c 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         71.24 ± 0.32 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          6.09 ± 0.00 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        188.86 ± 0.24 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          6.14 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log
index 3609ee1..f2831cf 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |        726.92 ± 2.95 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.43 ± 0.00 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           pp512 |       808.74 ± 14.71 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |  1 |    0 |           tg128 |          8.51 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log
index 37b0832..722e29e 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        162.40 ± 0.45 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.07 ± 0.00 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        395.07 ± 0.76 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.12 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx65536.log
index 3162148..8940c51 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         73.23 ± 0.94 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          6.11 ± 0.00 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        239.54 ± 0.33 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          6.14 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log
index 43d4db9..1b8385d 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log
@@ -1,23 +1,23 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7f143de79465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f143de7982b]
-/lib64/libggml-base.so.0(+0x16ed9) [0x7f143de8bed9]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f143d5efbfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f143d5d9d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f143d5efea8]
-/lib64/libggml-vulkan.so.0(+0x173b2) [0x7f143df493b2]
-/lib64/libggml-vulkan.so.0(+0x1075d0) [0x7f143e0395d0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f143de95112]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f1441ab0530]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f1441ab2b05]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f1441ab926f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f1441ababee]
-/usr/sbin/llama-bench() [0x41a1bb]
-/usr/sbin/llama-bench() [0x416ec6]
-/lib64/libc.so.6(+0x35b5) [0x7f143d2c05b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f143d2c0668]
-/usr/sbin/llama-bench() [0x4189d5]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7f5169267465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f516926783b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7f5169279f59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7f516900abfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f5168ff4d3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7f516900aea8]
+/lib64/libggml-vulkan.so.0(+0x15411) [0x7f516933a411]
+/lib64/libggml-vulkan.so.0(+0xfefd0) [0x7f5169423fd0]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f51692831d2]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f516cc3d9d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f516cc405b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7f516cc46f95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7f516cc48aae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403c06]
+/lib64/libc.so.6(+0x35b5) [0x7f5168cdb5b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7f5168cdb668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] Ministral-3-14B-Instruct-2512-BF16__fa1 failed (exit 0)
+✖ [vulkan_amdvlk] Ministral-3-14B-Instruct-2512-BF16__fa1 failed (exit 134)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log
index 3fad042..978d981 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,23 +1,23 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7f4a99ca1465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f4a99ca182b]
-/lib64/libggml-base.so.0(+0x16ed9) [0x7f4a99cb3ed9]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f4a99417bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f4a99401d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f4a99417ea8]
-/lib64/libggml-vulkan.so.0(+0x173b2) [0x7f4a99d713b2]
-/lib64/libggml-vulkan.so.0(+0x1075d0) [0x7f4a99e615d0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f4a99cbd112]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f4a9d8d8530]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f4a9d8dab05]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f4a9d8e126f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f4a9d8e2bee]
-/usr/sbin/llama-bench() [0x41a1bb]
-/usr/sbin/llama-bench() [0x416c49]
-/lib64/libc.so.6(+0x35b5) [0x7f4a990e85b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f4a990e8668]
-/usr/sbin/llama-bench() [0x4189d5]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7f388e92f465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f388e92f83b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7f388e941f59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7f388e6d2bfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f388e6bcd3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7f388e6d2ea8]
+/lib64/libggml-vulkan.so.0(+0x15411) [0x7f388ea02411]
+/lib64/libggml-vulkan.so.0(+0xfefd0) [0x7f388eaebfd0]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f388e94b1d2]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f38923059d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f38923085b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7f389230ef95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7f3892310aae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403989]
+/lib64/libc.so.6(+0x35b5) [0x7f388e3a35b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7f388e3a3668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] Ministral-3-14B-Instruct-2512-BF16__fa1 __longctx32768 failed (exit 0)
+✖ [vulkan_amdvlk] Ministral-3-14B-Instruct-2512-BF16__fa1 __longctx32768 failed (exit 134)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx65536.log
index fe396cf..419a93f 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,23 +1,23 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7f41f07d5465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f41f07d583b]
-/lib64/libggml-base.so.0(+0x16f19) [0x7f41f07e7f19]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f41f0578bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f41f0562d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f41f0578ea8]
-/lib64/libggml-vulkan.so.0(+0x1728d) [0x7f41f08aa28d]
-/lib64/libggml-vulkan.so.0(+0x10a410) [0x7f41f099d410]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f41f07f1192]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f41f445fc70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f41f4462255]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f41f446898f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f41f446a32e]
-/usr/sbin/llama-bench() [0x40663b]
-/usr/sbin/llama-bench() [0x4038b9]
-/lib64/libc.so.6(+0x35b5) [0x7f41f02495b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f41f0249668]
-/usr/sbin/llama-bench() [0x404e65]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7f25ade02465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f25ade0283b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7f25ade14f59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7f25adba5bfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f25adb8fd3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7f25adba5ea8]
+/lib64/libggml-vulkan.so.0(+0x15411) [0x7f25aded5411]
+/lib64/libggml-vulkan.so.0(+0xfefd0) [0x7f25adfbefd0]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f25ade1e1d2]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f25b17d89d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f25b17db5b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7f25b17e1f95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7f25b17e3aae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403989]
+/lib64/libc.so.6(+0x35b5) [0x7f25ad8765b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7f25ad876668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
-✖ ! [vulkan_amdvlk] Ministral-3-14B-Instruct-2512-BF16__fa1 __longctx65536 failed (exit 0)
+✖ [vulkan_amdvlk] Ministral-3-14B-Instruct-2512-BF16__fa1 __longctx65536 failed (exit 134)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log
index 22f733b..346cbd3 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           pp512 |        164.20 ± 0.55 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           tg128 |          7.94 ± 0.00 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           pp512 |        255.05 ± 0.70 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |           tg128 |          7.83 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log
index d95eba5..b52e9f5 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         72.41 ± 0.21 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          6.71 ± 0.00 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         79.54 ± 0.26 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          6.62 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx65536.log b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx65536.log
index 8439df5..47d2c14 100644
--- a/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         42.48 ± 0.11 |
-| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          5.81 ± 0.00 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         35.49 ± 0.19 |
+| mistral3 14B BF16              |  25.16 GiB |    13.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          5.74 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index bf06839..0000000
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        333.46 ± 1.20 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.99 ± 0.04 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index c4e622b..0000000
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        307.47 ± 0.55 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.60 ± 0.01 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1.log
deleted file mode 100644
index 44934bf..0000000
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        267.67 ± 1.70 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.18 ± 0.07 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index b1d3089..0000000
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        304.86 ± 0.24 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.79 ± 0.04 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index b578d21..0000000
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        249.66 ± 0.90 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         15.42 ± 0.07 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2__fa1.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1.log
similarity index 62%
rename from benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2__fa1.log
rename to benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1.log
index 3e633b5..f14062e 100644
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2__fa1.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        259.71 ± 1.51 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.74 ± 0.00 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        275.89 ± 1.53 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.85 ± 0.04 |
 
-build: 2405d59cb (8577)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log
similarity index 77%
rename from benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
rename to benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log
index 9cd5891..39c6022 100644
--- a/benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        302.29 ± 0.44 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.58 ± 0.01 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        382.38 ± 5.98 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         14.40 ± 0.07 |
 
-build: 2405d59cb (8577)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log
similarity index 77%
rename from benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log
rename to benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log
index fa48132..68129b7 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        250.79 ± 0.32 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         15.03 ± 0.04 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        348.10 ± 0.16 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         13.96 ± 0.16 |
 
-build: 7957de9dc (8645)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1.log
index 86597fe..973f984 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        261.51 ± 1.50 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.63 ± 0.00 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        272.39 ± 5.98 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.50 ± 0.06 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
index 7d5f443..b3d6c5b 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        323.83 ± 0.28 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.34 ± 0.01 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        379.64 ± 0.46 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         14.15 ± 0.16 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log
index 9198c57..f1698ec 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        290.67 ± 0.26 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         15.11 ± 0.05 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        333.18 ± 0.36 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         13.88 ± 0.17 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
index 229bdfb..3041cd7 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        232.54 ± 0.27 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         15.81 ± 0.00 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           pp512 |        239.02 ± 1.74 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |  1 |    0 |           tg128 |         14.43 ± 0.02 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
index 0446b85..a1f4485 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        284.15 ± 0.23 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.34 ± 0.61 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        353.69 ± 0.07 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         14.32 ± 0.08 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log
index 97ba25a..e8ffdc4 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        242.09 ± 0.33 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         15.47 ± 0.10 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        326.04 ± 0.10 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         14.05 ± 0.19 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
index 7fc882e..6f0b3ce 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |        138.78 ± 0.58 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         14.45 ± 0.09 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |        160.64 ± 0.60 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         14.44 ± 0.10 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
index 83f7a70..66e01f5 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        104.96 ± 0.02 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         13.66 ± 0.21 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        120.39 ± 0.10 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         13.67 ± 0.16 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log
index 42c0726..47862c7 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         85.09 ± 0.16 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         13.21 ± 0.04 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         96.89 ± 0.18 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         13.30 ± 0.25 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1.log
index 412da30..fd3678b 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |        191.97 ± 8.53 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         14.70 ± 0.35 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |           pp512 |        270.56 ± 2.02 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |           tg128 |         14.86 ± 0.06 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
index e3a63f3..1713890 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        185.90 ± 0.18 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         14.21 ± 0.38 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        242.77 ± 0.08 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         14.30 ± 0.15 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log
index f07472e..2c645a7 100644
--- a/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        174.22 ± 0.19 |
-| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         13.55 ± 0.03 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        216.34 ± 0.78 |
+| nemotron_h_moe 120B.A12B Q4_K - Medium |  78.02 GiB |   120.67 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         13.69 ± 0.35 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 2296b6a..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1498.50 ± 6.83 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         68.29 ± 0.67 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 608ce77..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        213.71 ± 0.11 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         33.47 ± 0.03 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index 51d8bed..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        110.56 ± 0.10 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         22.59 ± 0.01 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1.log
deleted file mode 100644
index 7f4faac..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1149.95 ± 8.12 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         67.57 ± 0.10 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index ac5a4b3..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        194.74 ± 0.07 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         33.36 ± 0.13 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 8280acb..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        102.26 ± 0.13 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         22.53 ± 0.07 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log
deleted file mode 100644
index 5011d54..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1167.77 ± 7.34 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         65.06 ± 0.02 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log
deleted file mode 100644
index 724b2a2..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        291.69 ± 0.26 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         32.81 ± 0.02 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx65536.log
deleted file mode 100644
index fa76777..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        159.40 ± 0.10 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         22.26 ± 0.17 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log
deleted file mode 100644
index acd091b..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           pp512 |       1185.86 ± 2.21 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |  1 |    0 |           tg128 |         68.95 ± 0.01 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log
deleted file mode 100644
index c0601b2..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        203.21 ± 0.11 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         34.77 ± 0.02 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx65536.log
deleted file mode 100644
index ec90521..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        102.96 ± 0.05 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         23.26 ± 0.12 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log
deleted file mode 100644
index f60bdd6..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |        882.68 ± 3.83 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         80.48 ± 0.06 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log
deleted file mode 100644
index ec6b25c..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        111.69 ± 0.03 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         28.03 ± 0.03 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx65536.log
deleted file mode 100644
index 7744f9d..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         60.42 ± 0.04 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         17.44 ± 0.01 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log
deleted file mode 100644
index 7d5e7d4..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1145.07 ± 8.85 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |           tg128 |         82.16 ± 3.05 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log
deleted file mode 100644
index 36ea50e..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        194.38 ± 0.33 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         37.09 ± 0.05 |
-
-build: 3f8752b55 (8743)
diff --git a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx65536.log b/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx65536.log
deleted file mode 100644
index 210de70..0000000
--- a/benchmark/results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         63.65 ± 0.38 |
-| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         24.54 ± 0.02 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 5b9a82c..0000000
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        418.75 ± 3.97 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         19.62 ± 0.03 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index c456266..0000000
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        207.29 ± 0.79 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         15.34 ± 2.04 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1.log
deleted file mode 100644
index b9f6f0d..0000000
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        314.59 ± 2.13 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         19.96 ± 0.15 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index 2143a5b..0000000
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        235.56 ± 9.94 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         16.82 ± 0.08 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 4b7d92a..0000000
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        152.12 ± 0.13 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         14.21 ± 0.22 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2__fa1.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1.log
similarity index 63%
rename from benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2__fa1.log
rename to benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1.log
index 6190e09..5fb9e16 100644
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2__fa1.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        305.11 ± 1.38 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         19.18 ± 0.00 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        336.69 ± 1.88 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         20.12 ± 0.02 |
 
-build: 2405d59cb (8577)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log
similarity index 78%
rename from benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
rename to benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log
index 3f12afa..16389cd 100644
--- a/benchmark/results/29-03-2026/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        246.59 ± 0.45 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         16.49 ± 0.01 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        276.15 ± 4.08 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         17.73 ± 0.08 |
 
-build: 2405d59cb (8577)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log
similarity index 78%
rename from benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log
rename to benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log
index db2483b..5cc2573 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        170.79 ± 0.22 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         14.19 ± 0.06 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        192.28 ± 2.56 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         15.74 ± 0.27 |
 
-build: 7957de9dc (8645)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1.log
index 23f86cc..aaa49bd 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        316.73 ± 1.27 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         18.86 ± 0.02 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        336.02 ± 1.65 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         19.73 ± 0.02 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
index 4842808..64b1ace 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        275.81 ± 1.62 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         17.90 ± 0.01 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        300.11 ± 0.10 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         18.47 ± 0.15 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log
index 81ec5de..dc3af70 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        187.01 ± 3.13 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         16.91 ± 0.09 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        202.62 ± 1.80 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         17.30 ± 0.22 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
index dbdb4c2..9dee21d 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        318.17 ± 2.18 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         19.57 ± 0.14 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           pp512 |        338.52 ± 3.41 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |  1 |    0 |           tg128 |         20.09 ± 0.02 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
index 5607dd4..3cda8ac 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        223.61 ± 0.83 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         18.32 ± 0.01 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        272.63 ± 0.34 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         18.69 ± 0.18 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log
index 8c2d444..7762da5 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        138.80 ± 0.17 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         17.26 ± 0.15 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        184.62 ± 0.08 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         17.52 ± 0.15 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
index f1a164d..49290c8 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |           pp512 |        182.98 ± 1.82 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |           tg128 |         21.33 ± 0.00 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |           pp512 |        202.91 ± 1.91 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |           tg128 |         21.42 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
index c587a4f..68d82e3 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        102.84 ± 0.14 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         18.77 ± 0.04 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        115.36 ± 0.12 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         19.04 ± 0.02 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log
index 2e87f68..a20596f 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         70.58 ± 0.04 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         16.94 ± 0.01 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         81.21 ± 0.05 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         17.29 ± 0.02 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1.log
index f531091..9fc261b 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |           pp512 |        241.15 ± 7.43 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |           tg128 |         21.75 ± 0.24 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |           pp512 |        308.78 ± 4.29 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |           tg128 |         22.30 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
index 502caf7..3a01aae 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        202.99 ± 0.41 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         20.09 ± 0.02 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        224.82 ± 0.40 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         20.38 ± 0.02 |
 
-build: 3f8752b55 (8743)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log
index 51a6b87..2f6d40c 100644
--- a/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        156.41 ± 0.22 |
-| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         18.78 ± 0.01 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        165.14 ± 0.48 |
+| qwen35moe 122B.A10B Q5_K - Medium |  85.60 GiB |   122.11 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         18.99 ± 0.01 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1.log
new file mode 100644
index 0000000..cf438be
--- /dev/null
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1.log
@@ -0,0 +1,8 @@
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |  1 |    0 |           pp512 |        367.47 ± 2.44 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |  1 |    0 |           tg128 |          6.51 ± 0.00 |
+
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
similarity index 62%
rename from benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
rename to benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
index 8d95301..53dcfef 100644
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__hblt0__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         17.08 ± 0.02 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.08 ± 0.00 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        197.70 ± 2.94 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.14 ± 0.01 |
 
-build: 2405d59cb (8577)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
similarity index 62%
rename from benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log
rename to benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
index 80cc879..7c79694 100644
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |          8.12 ± 0.03 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          1.48 ± 0.00 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        106.30 ± 1.69 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          5.83 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1.log
new file mode 100644
index 0000000..174cd2e
--- /dev/null
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1.log
@@ -0,0 +1,8 @@
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |  1 |    0 |           pp512 |        364.12 ± 1.62 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |  1 |    0 |           tg128 |          6.48 ± 0.00 |
+
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
similarity index 62%
rename from benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
rename to benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
index 8999b32..6cbb4ec 100644
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         17.56 ± 0.07 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.09 ± 0.00 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        172.59 ± 0.84 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.13 ± 0.01 |
 
-build: 2405d59cb (8577)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
similarity index 62%
rename from benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
rename to benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
index ddb75a1..6c739d4 100644
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |          8.16 ± 0.08 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          1.58 ± 0.00 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         93.78 ± 1.21 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          5.81 ± 0.01 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1.log
similarity index 60%
rename from benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1.log
rename to benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1.log
index 35551cf..103bef0 100644
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1.log
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           pp512 |         78.37 ± 0.10 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.96 ± 0.00 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |  1 |    0 |           pp512 |        371.42 ± 4.05 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |  1 |    0 |           tg128 |          6.51 ± 0.00 |
 
-build: 2405d59cb (8577)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
similarity index 75%
rename from benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
rename to benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
index 4f22b35..ad9e34b 100644
--- a/benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         17.07 ± 0.06 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.07 ± 0.00 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        190.65 ± 0.72 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.14 ± 0.01 |
 
-build: 2405d59cb (8577)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
similarity index 75%
rename from benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log
rename to benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
index a4c2062..f38be6a 100644
--- a/benchmark/results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |          7.94 ± 0.04 |
-| llama ?B Q4_K - Medium         |  70.31 GiB |   125.03 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          1.49 ± 0.00 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        105.08 ± 0.72 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          5.82 ± 0.01 |
 
-build: 7957de9dc (8645)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
new file mode 100644
index 0000000..b8d023a
--- /dev/null
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
@@ -0,0 +1,8 @@
+ggml_vulkan: Found 1 Vulkan devices:
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | Vulkan     |  99 |  1 |    0 |           pp512 |         80.22 ± 0.05 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | Vulkan     |  99 |  1 |    0 |           tg128 |          5.88 ± 0.00 |
+
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
new file mode 100644
index 0000000..3e31791
--- /dev/null
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
@@ -0,0 +1,8 @@
+ggml_vulkan: Found 1 Vulkan devices:
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         60.84 ± 0.01 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          5.42 ± 0.00 |
+
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
new file mode 100644
index 0000000..1234ef9
--- /dev/null
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
@@ -0,0 +1,8 @@
+ggml_vulkan: Found 1 Vulkan devices:
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         48.97 ± 0.01 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          5.03 ± 0.00 |
+
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1.log
new file mode 100644
index 0000000..98057f8
--- /dev/null
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1.log
@@ -0,0 +1,8 @@
+ggml_vulkan: Found 1 Vulkan devices:
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | Vulkan     |  99 |  1 |    0 |           pp512 |        274.97 ± 0.46 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | Vulkan     |  99 |  1 |    0 |           tg128 |          6.30 ± 0.00 |
+
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
new file mode 100644
index 0000000..72dde4a
--- /dev/null
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
@@ -0,0 +1,8 @@
+ggml_vulkan: Found 1 Vulkan devices:
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        123.12 ± 0.56 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          5.92 ± 0.00 |
+
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
new file mode 100644
index 0000000..d2fffe4
--- /dev/null
+++ b/benchmark/results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
@@ -0,0 +1,8 @@
+ggml_vulkan: Found 1 Vulkan devices:
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         63.14 ± 0.36 |
+| qwen35 27B Q8_0                |  33.31 GiB |    27.32 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          5.60 ± 0.00 |
+
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index ffdd0be..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |        513.08 ± 3.80 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         23.46 ± 0.06 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 1800190..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        379.44 ± 1.48 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         21.65 ± 0.01 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1.log
deleted file mode 100644
index f836f61..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |        509.14 ± 3.56 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         23.98 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index d1591bf..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        391.47 ± 0.29 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         21.88 ± 0.01 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index fe7a028..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        264.06 ± 0.65 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         20.14 ± 0.01 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1.log
similarity index 64%
rename from benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
rename to benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1.log
index 388df5f..a792f4c 100644
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm-7_2__hblt0__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |        512.17 ± 3.01 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         23.69 ± 0.00 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       525.94 ± 13.39 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         26.01 ± 0.00 |
 
-build: 2405d59cb (8577)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log
similarity index 79%
rename from benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
rename to benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log
index b6be3f4..bf2df28 100644
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-BF16-00001-of-00002__rocm-7_2__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        423.84 ± 0.73 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         21.64 ± 0.01 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        417.86 ± 0.42 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         23.91 ± 0.01 |
 
-build: 2405d59cb (8577)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log
similarity index 79%
rename from benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log
rename to benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log
index cea0bad..f49abb1 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        296.43 ± 1.09 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         19.60 ± 0.00 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        322.65 ± 7.40 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         22.20 ± 0.01 |
 
-build: 7957de9dc (8645)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
index 7f7683b..2de5d30 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       538.42 ± 12.90 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.91 ± 0.00 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |        573.71 ± 4.26 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         25.07 ± 0.00 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
index 1393959..2711ec9 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        412.11 ± 0.60 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         21.59 ± 0.09 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        426.65 ± 0.74 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         23.77 ± 0.01 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
index de6b985..54d9d88 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        326.10 ± 1.55 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         19.05 ± 0.01 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        343.79 ± 1.31 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         22.22 ± 0.01 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log
index f3a2f0e..57102f9 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log
@@ -1,6 +1,8 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 64055 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 64055 MiB
-main: error: failed to load model '/home/kyuz0/models/qwen-3.6-35b-a3b/BF16/Qwen3.6-35B-A3B-BF16-00001-of-00002.gguf'
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
+  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-✖ ! [rocm7-nightlies] Qwen3.6-35B-A3B-BF16-00001-of-00002__fa1 failed (exit 0)
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |        528.97 ± 5.27 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         25.91 ± 0.01 |
+
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
index c843485..e82db29 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
@@ -1,6 +1,8 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 64055 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 64055 MiB
-main: error: failed to load model '/home/kyuz0/models/qwen-3.6-35b-a3b/BF16/Qwen3.6-35B-A3B-BF16-00001-of-00002.gguf'
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
+  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-✖ ! [rocm7-nightlies] Qwen3.6-35B-A3B-BF16-00001-of-00002__fa1 __longctx32768 failed (exit 0)
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        409.92 ± 1.94 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         23.96 ± 0.44 |
+
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
index a164cdd..f21d85e 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        256.49 ± 1.26 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         20.93 ± 0.06 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        314.48 ± 1.45 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         22.60 ± 0.02 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
index 6a5a268..e538e85 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |        122.43 ± 0.23 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         11.55 ± 0.01 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |        122.89 ± 0.24 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         11.60 ± 0.03 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
index 028a15c..8193e40 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         98.18 ± 0.09 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.94 ± 0.01 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        100.51 ± 0.04 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.93 ± 0.01 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
index dd4c0ad..dd5ed27 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         82.44 ± 0.12 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         10.40 ± 0.01 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         85.86 ± 0.08 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         10.46 ± 0.01 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
index 9cb2ce7..dca1ef1 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |        316.59 ± 1.87 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         10.80 ± 0.01 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |        328.40 ± 1.64 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         10.68 ± 0.01 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
index b336a9a..72bffa4 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        267.07 ± 0.91 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.37 ± 0.01 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        275.80 ± 0.89 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.23 ± 0.00 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log
index 4f99e4f..a611cbf 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        226.78 ± 1.49 |
-| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         10.05 ± 0.00 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        241.42 ± 0.58 |
+| qwen35moe 35B.A3B BF16         |  64.60 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          9.94 ± 0.02 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index d480c06..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       1359.62 ± 5.74 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         48.34 ± 0.43 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index fc7af7d..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        670.26 ± 1.77 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         40.42 ± 0.04 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1.log
deleted file mode 100644
index 195ecab..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       1089.21 ± 5.82 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         49.27 ± 0.08 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index 5ced017..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        703.58 ± 0.51 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         41.08 ± 0.33 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 3df355f..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        460.71 ± 1.18 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         35.40 ± 0.24 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1.log
new file mode 100644
index 0000000..0b531cb
--- /dev/null
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1.log
@@ -0,0 +1,8 @@
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |  1 |    0 |           pp512 |       1120.00 ± 8.54 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.53 ± 0.16 |
+
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm-7_2__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log
similarity index 62%
rename from benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm-7_2__fa1__longctx32768.log
rename to benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log
index 15f4159..50b9233 100644
--- a/benchmark/results/29-03-2026/Qwen3.5-35B-A3B-UD-Q4_K_XL__rocm-7_2__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        669.09 ± 1.13 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.70 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         40.19 ± 0.03 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        746.60 ± 0.16 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         43.84 ± 0.36 |
 
-build: 2405d59cb (8577)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log
similarity index 62%
rename from benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
rename to benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log
index 983e08c..de749bd 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        412.60 ± 0.33 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         34.90 ± 0.02 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        493.36 ± 0.38 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         38.39 ± 0.27 |
 
-build: 7957de9dc (8645)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1.log
index 0b2707e..14ac0c4 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       1067.33 ± 6.85 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         48.23 ± 0.11 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |  1 |    0 |           pp512 |       1116.91 ± 7.21 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.32 ± 0.12 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log
index e6f9da3..5fe75e6 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        769.99 ± 2.79 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         41.91 ± 0.49 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        807.33 ± 0.34 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         43.47 ± 0.44 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log
index 676bb76..19ebddf 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        473.84 ± 2.27 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         36.70 ± 0.97 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        547.71 ± 1.13 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         38.47 ± 0.41 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1.log
index 083496a..84b8cc0 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1.log
@@ -1,8 +1,8 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 64055 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 64055 MiB
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
+  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |      1091.62 ± 10.45 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         49.10 ± 0.15 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |  1 |    0 |           pp512 |      1085.14 ± 11.75 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.47 ± 0.18 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log
index 382213d..427231d 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log
@@ -1,8 +1,8 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 64055 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 64055 MiB
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
+  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        664.89 ± 1.48 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         42.73 ± 0.65 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        716.94 ± 1.11 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         44.45 ± 0.70 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log
index 222c0d8..c617aae 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        432.39 ± 0.18 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         38.08 ± 0.59 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        468.89 ± 0.65 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         39.30 ± 0.58 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1.log
index ad0830a..a6f017e 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |        663.94 ± 2.80 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         57.13 ± 0.04 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | Vulkan     |  99 |  1 |    0 |           pp512 |        696.54 ± 2.95 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | Vulkan     |  99 |  1 |    0 |           tg128 |         59.57 ± 0.07 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log
index beda67a..45ebe64 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        292.49 ± 0.18 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         43.97 ± 0.06 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        319.31 ± 0.43 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         45.30 ± 0.20 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log
index 122b511..89c4ad0 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        183.42 ± 0.98 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         36.62 ± 0.05 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        207.54 ± 0.69 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         36.74 ± 0.04 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1.log
index fea8098..c73d477 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1045.16 ± 5.79 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         58.82 ± 0.08 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1113.64 ± 7.57 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | Vulkan     |  99 |  1 |    0 |           tg128 |         60.43 ± 0.06 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log
index e344a93..6add865 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        682.80 ± 1.45 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         48.77 ± 0.07 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        715.52 ± 1.18 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         49.21 ± 0.38 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log
index 44d692e..0501d15 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        507.32 ± 1.33 |
-| qwen35moe 35B.A3B Q4_K - Medium |  20.81 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         42.75 ± 0.05 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        524.70 ± 1.20 |
+| qwen35moe 35B.A3B Q4_K - Medium |  21.27 GiB |    35.51 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         43.23 ± 0.23 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 42341b2..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       1222.77 ± 2.88 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         43.72 ± 0.08 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index 14a3929..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        544.04 ± 1.79 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         37.70 ± 0.21 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 8f2da45..0000000
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        414.74 ± 1.74 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         32.74 ± 0.02 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1.log
similarity index 64%
rename from benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1.log
rename to benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1.log
index 06efdb6..1248374 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |      1043.12 ± 46.10 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         44.53 ± 0.01 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       1093.92 ± 7.09 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         46.18 ± 0.01 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
similarity index 79%
rename from benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
rename to benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
index d92138b..058eb53 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        653.49 ± 0.18 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         37.21 ± 0.02 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        728.92 ± 0.14 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         40.00 ± 0.01 |
 
-build: 7957de9dc (8645)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
similarity index 79%
rename from benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
rename to benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
index 6b183a9..30aab69 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        414.31 ± 0.61 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         32.30 ± 0.01 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        486.17 ± 0.85 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         35.45 ± 0.05 |
 
-build: 7957de9dc (8645)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log
index 1c3f7f0..a5eaa51 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       1049.22 ± 7.25 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         43.10 ± 0.01 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |      1095.53 ± 11.56 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         45.13 ± 0.01 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
index a75e2df..781378f 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        758.76 ± 0.99 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         38.70 ± 0.01 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        787.54 ± 0.45 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         39.88 ± 0.00 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
index 39a638f..5a7dfbf 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        515.64 ± 0.13 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         34.53 ± 0.03 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        537.23 ± 1.72 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         35.72 ± 0.04 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log
index 637ab39..f08b815 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log
@@ -1,8 +1,8 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 64055 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 64055 MiB
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
+  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |       1036.89 ± 9.24 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         44.20 ± 0.00 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           pp512 |      1088.56 ± 10.42 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |  1 |    0 |           tg128 |         46.21 ± 0.01 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
index 104e4f9..8f7c6b6 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
@@ -1,8 +1,8 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 64055 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 64055 MiB
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
+  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        510.38 ± 2.64 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         38.18 ± 1.71 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        704.46 ± 1.78 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         40.54 ± 0.21 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
index 464aae3..f4e0208 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        375.11 ± 2.14 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         35.02 ± 0.03 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        462.61 ± 1.15 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         36.31 ± 0.01 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
index 1ed9fe0..2107ef3 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |        576.25 ± 2.16 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         45.25 ± 0.01 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |        640.78 ± 2.08 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         46.53 ± 0.07 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
index e078197..0055266 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        273.77 ± 0.30 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         36.71 ± 0.08 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        307.07 ± 1.01 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         37.93 ± 0.07 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
index 9224ec2..52334ce 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        175.02 ± 1.28 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         31.29 ± 0.05 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        202.18 ± 0.79 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         32.35 ± 0.04 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log
index 78c10fa..d6c6942 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |        893.79 ± 4.74 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         46.05 ± 0.07 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1045.00 ± 7.30 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |           tg128 |         46.33 ± 0.02 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
index a6902ba..146be11 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        610.84 ± 1.99 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         39.67 ± 0.15 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        682.80 ± 0.96 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         40.08 ± 0.17 |
 
-build: f53577432 (8942)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
index cc7bef5..6724eb3 100644
--- a/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        468.83 ± 1.17 |
-| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         35.46 ± 0.04 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        504.87 ± 1.05 |
+| qwen35moe 35B.A3B Q8_0         |  35.80 GiB |    34.66 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         35.80 ± 0.05 |
 
-build: ab6120cde (8997)
+build: 1a68ec937 (9193)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index ad4fb02..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |        836.74 ± 6.15 |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.95 ± 0.01 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 2cd4471..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        619.85 ± 6.10 |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.38 ± 0.01 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index c5eeaa8..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        408.04 ± 3.58 |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         19.52 ± 0.01 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1.log
deleted file mode 100644
index ba410bb..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |       823.65 ± 83.54 |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.67 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index 27ef16c..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        628.14 ± 5.88 |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.15 ± 0.01 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 5f077bd..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        403.29 ± 1.83 |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         19.35 ± 0.19 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log
index d61700a..33baec1 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |      796.06 ± 126.76 |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.68 ± 0.00 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |        909.04 ± 7.54 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         24.64 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log
index 063ea41..132b7e9 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        628.20 ± 5.71 |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.12 ± 0.00 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        667.44 ± 3.68 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         21.59 ± 0.01 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log
index 4909514..c581b57 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        426.61 ± 3.98 |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         19.55 ± 0.01 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        446.88 ± 2.91 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         20.79 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
index 27617ae..0b67b9c 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |        839.13 ± 7.61 |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         20.75 ± 0.00 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |        893.46 ± 6.97 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         23.37 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
index 38d2b4a..7a3983a 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       646.67 ± 11.22 |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         19.86 ± 0.01 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        691.94 ± 9.26 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         21.39 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
index 898ec96..ac4f51b 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        441.14 ± 2.12 |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         17.76 ± 0.01 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        470.28 ± 3.56 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         20.66 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log
index bdb11b0..1a14f0c 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |       844.42 ± 19.66 |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         22.69 ± 0.00 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |       870.72 ± 80.56 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         24.65 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
index cf1a239..49f28c3 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        546.66 ± 5.96 |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         20.11 ± 0.07 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        638.41 ± 6.72 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         21.58 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
index 95707ae..351e768 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        345.70 ± 0.13 |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         19.16 ± 0.47 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        421.88 ± 0.33 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         20.88 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
index 1a94edd..ce565f3 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           pp512 |        117.96 ± 0.28 |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           tg128 |         16.18 ± 0.02 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           pp512 |        123.59 ± 0.22 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           tg128 |         18.85 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
index 62d20bc..80e1308 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         58.94 ± 0.07 |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         15.26 ± 0.01 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         60.07 ± 0.05 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         16.44 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
index 383b4a9..134402a 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         39.57 ± 0.04 |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         14.10 ± 0.02 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         39.84 ± 0.03 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         15.10 ± 0.01 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1.log
index 882cb85..6c4a2a0 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           pp512 |        376.52 ± 7.37 |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           tg128 |         14.74 ± 0.01 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           pp512 |        587.64 ± 6.59 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           tg128 |         18.60 ± 0.02 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
index 98e4674..4b47ff7 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        351.87 ± 1.49 |
-| gemma4 ?B BF16                 |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         13.71 ± 0.06 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        428.06 ± 1.23 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         16.99 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log
index 034abc4..02195b8 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        281.32 ± 0.70 |
-| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         13.15 ± 0.03 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        331.82 ± 0.61 |
+| gemma4 26B.A4B BF16            |  47.02 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         16.12 ± 0.02 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 3363194..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |       1608.01 ± 5.26 |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         47.13 ± 0.29 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index ba51449..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       689.30 ± 12.21 |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         37.14 ± 0.01 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index 2ce4c72..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        446.29 ± 0.87 |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         35.16 ± 0.03 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1.log
deleted file mode 100644
index e24d90b..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |      1299.90 ± 11.12 |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         46.59 ± 0.02 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index e2f02bd..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        693.48 ± 4.74 |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         36.76 ± 0.16 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 1bb9c55..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        453.92 ± 5.18 |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         34.85 ± 0.16 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log
index 88d1ad9..b2a616e 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |      1301.66 ± 10.69 |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         46.42 ± 0.00 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |       1329.04 ± 9.98 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         48.46 ± 0.11 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log
index 00152b7..e55bad8 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       693.56 ± 10.92 |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         36.83 ± 0.15 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        725.46 ± 9.61 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         37.84 ± 0.15 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log
index dd17c6d..64b339e 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        450.75 ± 0.77 |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         34.84 ± 0.16 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        471.30 ± 1.61 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         35.61 ± 0.14 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log
index 20ea166..0c61a87 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |      1288.38 ± 10.85 |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         45.58 ± 0.01 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |       1334.91 ± 9.90 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         47.74 ± 0.07 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log
index aad4700..65cf0f1 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       721.84 ± 10.97 |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         36.17 ± 0.02 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       761.50 ± 12.52 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         37.29 ± 0.26 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log
index ccd7682..dbf422d 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        469.64 ± 3.17 |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         34.31 ± 0.25 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        499.49 ± 1.70 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         35.18 ± 0.24 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log
index 1db2c43..ce3bc9d 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |      1219.74 ± 15.04 |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         46.57 ± 0.01 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |       1194.14 ± 6.99 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         48.47 ± 0.12 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log
index 8f086cb..efe677e 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        600.42 ± 7.55 |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         36.88 ± 0.01 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        674.07 ± 6.75 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         37.68 ± 0.38 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log
index 7a8df40..506b436 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        374.22 ± 0.13 |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         34.77 ± 0.32 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        437.20 ± 0.38 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         35.52 ± 0.34 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log
index 0699b0b..b8addde 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           pp512 |        720.78 ± 2.94 |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           tg128 |         50.21 ± 0.06 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           pp512 |        765.18 ± 2.13 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           tg128 |         52.04 ± 0.02 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log
index 5ec6aae..515bd29 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        100.79 ± 0.03 |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         35.41 ± 0.01 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        101.11 ± 0.10 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         36.34 ± 0.06 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log
index df37dff..951c510 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         54.90 ± 0.12 |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         29.52 ± 0.03 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         54.46 ± 0.03 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         30.34 ± 0.05 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1.log
index ef94402..7c46ef6 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1213.78 ± 7.36 |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           tg128 |         53.04 ± 0.10 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1324.66 ± 8.47 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           tg128 |         54.67 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log
index 6ac9013..cf99b7e 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        638.76 ± 1.76 |
-| gemma4 ?B Q4_K - Medium        |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         40.93 ± 0.09 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        682.46 ± 2.10 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         42.40 ± 0.03 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log
index 69589cc..3f5ede9 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        445.00 ± 0.08 |
-| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         36.64 ± 0.06 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        469.95 ± 0.40 |
+| gemma4 26B.A4B Q4_K - Medium   |  15.90 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         37.46 ± 0.07 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 48ba107..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |       1524.44 ± 7.11 |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.68 ± 0.02 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 48432ce..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        689.36 ± 8.08 |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         33.67 ± 0.01 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index c26efe2..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        436.86 ± 3.98 |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         31.55 ± 0.72 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1.log
deleted file mode 100644
index 64550ee..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |      1272.69 ± 74.50 |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.31 ± 0.01 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index 9ed35f3..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        696.25 ± 7.26 |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         32.95 ± 0.77 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index dc7f2ce..0000000
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        452.10 ± 0.82 |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         29.87 ± 3.28 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log
index b8fb75e..ef920f8 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |      1301.93 ± 18.07 |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.35 ± 0.03 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |       1354.00 ± 8.88 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         42.35 ± 0.01 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
index d15b313..640a540 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        695.00 ± 7.26 |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         33.41 ± 0.01 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       725.86 ± 11.87 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         34.01 ± 0.02 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
index c68c3d3..f44d2de 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        447.63 ± 3.37 |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         31.76 ± 0.04 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        471.81 ± 3.32 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         32.14 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
index 5fbd898..b225ff8 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |       1303.32 ± 8.77 |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         38.81 ± 0.01 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |       1356.47 ± 9.00 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.28 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
index 84431bb..8d13e96 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        727.92 ± 8.62 |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         32.52 ± 0.35 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       769.66 ± 14.85 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         33.66 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
index c1f0b33..1c943d4 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        451.48 ± 1.29 |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         30.76 ± 0.03 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        504.88 ± 0.76 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         32.09 ± 0.02 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log
index 834db9b..d2d1d2d 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |      1294.40 ± 31.54 |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         41.09 ± 0.01 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           pp512 |       1342.08 ± 8.02 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |  1 |    0 |           tg128 |         42.32 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
index 1203ea4..da10573 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        596.78 ± 7.98 |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         33.35 ± 0.00 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        692.34 ± 8.31 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         34.00 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
index 9ee129b..b12cf40 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        363.83 ± 2.85 |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         31.68 ± 0.02 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        447.95 ± 0.85 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         32.16 ± 0.02 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
index c5a3b18..13e1450 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           pp512 |        589.20 ± 1.79 |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           tg128 |         43.18 ± 0.07 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           pp512 |        669.11 ± 1.51 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           tg128 |         43.93 ± 0.02 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
index 9b69152..2284da8 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         97.70 ± 0.10 |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         31.82 ± 0.05 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         99.13 ± 0.12 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         32.25 ± 0.02 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
index 7fccbf2..64467df 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         54.04 ± 0.08 |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         26.65 ± 0.03 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         53.86 ± 0.04 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         27.52 ± 0.03 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1.log
index 4ab7368..b20277c 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           pp512 |       944.96 ± 19.24 |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           tg128 |         43.58 ± 1.06 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1291.45 ± 8.95 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |           tg128 |         45.48 ± 0.03 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
index fd0cc70..0ee5b2d 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        579.88 ± 0.09 |
-| gemma4 ?B Q8_0                 |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         35.52 ± 0.07 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        672.93 ± 0.29 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         36.81 ± 0.04 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
index 7f18752..9bf0690 100644
--- a/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        415.94 ± 0.55 |
-| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         31.99 ± 0.06 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        465.71 ± 0.32 |
+| gemma4 26B.A4B Q8_0            |  25.94 GiB |    25.23 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         32.90 ± 0.01 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 7870940..0000000
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        380.58 ± 1.49 |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.49 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 9ff1bae..0000000
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        150.63 ± 1.53 |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.17 ± 0.03 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index 1ab9d0f..0000000
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         96.84 ± 0.57 |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.07 ± 0.02 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1.log
deleted file mode 100644
index 64f36dd..0000000
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        377.93 ± 2.48 |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.44 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index d58c607..0000000
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        153.11 ± 1.83 |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.18 ± 0.01 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 61898ee..0000000
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         98.23 ± 0.26 |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.08 ± 0.01 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log
index 9ea45f7..ea11a28 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        376.47 ± 1.77 |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.44 ± 0.00 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        428.07 ± 3.61 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.55 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log
index d6a45c5..0f9ed06 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        152.54 ± 2.43 |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.18 ± 0.00 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        167.20 ± 1.91 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.23 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log
index 982b824..db4d5ee 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         98.53 ± 0.11 |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.08 ± 0.02 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        108.99 ± 0.35 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.13 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
index e55d3f2..0548fac 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        360.84 ± 1.60 |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.30 ± 0.00 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        410.15 ± 2.16 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.44 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
index 9885e65..53ababf 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        154.86 ± 2.16 |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.14 ± 0.01 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        172.49 ± 2.42 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.21 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
index 3fdc9fe..c1c545d 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        101.08 ± 0.13 |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          2.94 ± 0.02 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        114.16 ± 0.63 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.12 ± 0.01 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log
index 1b9e740..8aab84e 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        373.84 ± 0.93 |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.41 ± 0.00 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        423.05 ± 6.41 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          3.53 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
index 1c1e2dd..d141254 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        139.53 ± 1.98 |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.16 ± 0.05 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        158.77 ± 2.07 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          3.25 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
index 125b805..66fea99 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         83.27 ± 0.13 |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.04 ± 0.00 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        102.83 ± 0.08 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.14 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
index 2cf8a6a..f324469 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log
@@ -1,23 +1,23 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7f2a944d5465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f2a944d582b]
-/lib64/libggml-base.so.0(+0x16ed9) [0x7f2a944e7ed9]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f2a93c4bbfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f2a93c35d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f2a93c4bea8]
-/lib64/libggml-vulkan.so.0(+0x173b2) [0x7f2a945a53b2]
-/lib64/libggml-vulkan.so.0(+0x1075d0) [0x7f2a946955d0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f2a944f1112]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f2a9810c530]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f2a9810eb05]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f2a9811526f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f2a98116bee]
-/usr/sbin/llama-bench() [0x41a1bb]
-/usr/sbin/llama-bench() [0x416ec6]
-/lib64/libc.so.6(+0x35b5) [0x7f2a9391c5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f2a9391c668]
-/usr/sbin/llama-bench() [0x4189d5]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7ff75610f465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7ff75610f83b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7ff756121f59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7ff755eb2bfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7ff755e9cd3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7ff755eb2ea8]
+/lib64/libggml-vulkan.so.0(+0x15411) [0x7ff7561e2411]
+/lib64/libggml-vulkan.so.0(+0xfefd0) [0x7ff7562cbfd0]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7ff75612b1d2]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7ff759ae59d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7ff759ae85b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7ff759aeef95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7ff759af0aae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403c06]
+/lib64/libc.so.6(+0x35b5) [0x7ff755b835b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7ff755b83668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
 ✖ ! [vulkan_amdvlk] gemma-4-31B-it-BF16-00001-of-00002__fa1 failed (exit 0)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
index 3a53be1..e41f6f1 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,23 +1,23 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7f7c4a4ab465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f7c4a4ab82b]
-/lib64/libggml-base.so.0(+0x16ed9) [0x7f7c4a4bded9]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f7c49c21bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f7c49c0bd3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f7c49c21ea8]
-/lib64/libggml-vulkan.so.0(+0x173b2) [0x7f7c4a57b3b2]
-/lib64/libggml-vulkan.so.0(+0x1075d0) [0x7f7c4a66b5d0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f7c4a4c7112]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f7c4e0e2530]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f7c4e0e4b05]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f7c4e0eb26f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f7c4e0ecbee]
-/usr/sbin/llama-bench() [0x41a1bb]
-/usr/sbin/llama-bench() [0x416c49]
-/lib64/libc.so.6(+0x35b5) [0x7f7c498f25b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f7c498f2668]
-/usr/sbin/llama-bench() [0x4189d5]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7fb7d3fe1465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fb7d3fe183b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7fb7d3ff3f59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7fb7d3d84bfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7fb7d3d6ed3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7fb7d3d84ea8]
+/lib64/libggml-vulkan.so.0(+0x15411) [0x7fb7d40b4411]
+/lib64/libggml-vulkan.so.0(+0xfefd0) [0x7fb7d419dfd0]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7fb7d3ffd1d2]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fb7d79b79d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7fb7d79ba5b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7fb7d79c0f95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7fb7d79c2aae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403989]
+/lib64/libc.so.6(+0x35b5) [0x7fb7d3a555b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7fb7d3a55668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
 ✖ ! [vulkan_amdvlk] gemma-4-31B-it-BF16-00001-of-00002__fa1 __longctx32768 failed (exit 0)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
index 6a066ed..87f932a 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,23 +1,23 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7fa1d7fcc465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fa1d7fcc83b]
-/lib64/libggml-base.so.0(+0x16f19) [0x7fa1d7fdef19]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7fa1d7d6fbfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7fa1d7d59d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7fa1d7d6fea8]
-/lib64/libggml-vulkan.so.0(+0x1728d) [0x7fa1d80a128d]
-/lib64/libggml-vulkan.so.0(+0x10a410) [0x7fa1d8194410]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7fa1d7fe8192]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fa1dbc56c70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7fa1dbc59255]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7fa1dbc5f98f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7fa1dbc6132e]
-/usr/sbin/llama-bench() [0x40663b]
-/usr/sbin/llama-bench() [0x4038b9]
-/lib64/libc.so.6(+0x35b5) [0x7fa1d7a405b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fa1d7a40668]
-/usr/sbin/llama-bench() [0x404e65]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7f0020d59465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f0020d5983b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7f0020d6bf59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7f0020afcbfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f0020ae6d3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7f0020afcea8]
+/lib64/libggml-vulkan.so.0(+0x15411) [0x7f0020e2c411]
+/lib64/libggml-vulkan.so.0(+0xfefd0) [0x7f0020f15fd0]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f0020d751d2]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f002472f9d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f00247325b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7f0024738f95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7f002473aaae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403989]
+/lib64/libc.so.6(+0x35b5) [0x7f00207cd5b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7f00207cd668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
 ✖ ! [vulkan_amdvlk] gemma-4-31B-it-BF16-00001-of-00002__fa1 __longctx65536 failed (exit 0)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1.log
index ac42bfc..348e87f 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           pp512 |         79.53 ± 0.26 |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           tg128 |          3.50 ± 0.00 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           pp512 |        135.94 ± 0.46 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           tg128 |          3.50 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
index 7b3e175..43a29bd 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         60.29 ± 0.37 |
-| gemma4 ?B BF16                 |  57.18 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          3.24 ± 0.00 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         92.40 ± 1.45 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          3.25 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log
index 8f3488f..20d4186 100644
--- a/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         45.96 ± 1.09 |
-| gemma4 31B BF16                |  57.18 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          3.06 ± 0.00 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         67.88 ± 1.23 |
+| gemma4 31B BF16                |  57.18 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          3.08 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 41a06c4..0000000
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        293.27 ± 0.37 |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |         10.50 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 3338698..0000000
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        136.84 ± 1.27 |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.11 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index 86912e3..0000000
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         91.86 ± 0.33 |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          7.51 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1.log
deleted file mode 100644
index 2c3483f..0000000
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        310.54 ± 0.73 |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |         10.52 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index 79d530a..0000000
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        140.71 ± 1.29 |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.11 ± 0.01 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 94ec84e..0000000
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         93.57 ± 0.20 |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          7.52 ± 0.01 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log
index 60b508a..0b06c8a 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        309.16 ± 0.31 |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |         10.49 ± 0.00 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        322.69 ± 0.81 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |         10.55 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log
index aa2a5d6..e50fce3 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        139.29 ± 1.61 |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.10 ± 0.01 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        150.43 ± 1.59 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.13 ± 0.01 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log
index 785a7fd..8ea35cf 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         92.80 ± 0.18 |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          7.51 ± 0.01 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        101.59 ± 0.41 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          7.54 ± 0.01 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log
index 66aac4c..aff532d 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        309.22 ± 0.13 |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |         10.02 ± 0.00 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        317.49 ± 0.94 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |         10.20 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log
index d7be7d7..244b954 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        143.91 ± 1.64 |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.81 ± 0.00 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        154.73 ± 1.21 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.95 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log
index b9aae5c..3aba455 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         96.98 ± 0.37 |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          7.26 ± 0.01 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        106.22 ± 0.22 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          7.38 ± 0.01 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log
index e3b7d2d..675c530 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        315.72 ± 0.30 |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |         10.51 ± 0.00 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        313.83 ± 2.55 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |         10.55 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log
index dd44eae..536aa19 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        130.80 ± 1.45 |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.20 ± 0.00 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        142.53 ± 1.40 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.21 ± 0.02 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log
index 43b05b9..b4b6fbb 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         84.96 ± 0.07 |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          7.56 ± 0.02 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         95.32 ± 0.10 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          7.58 ± 0.02 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log
index 2e5a7d2..9544d55 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           pp512 |         69.12 ± 0.02 |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           tg128 |         10.71 ± 0.00 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           pp512 |         70.48 ± 0.02 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           tg128 |         11.32 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log
index 4f6b526..b7bec69 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,23 +1,23 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7f0f8a452465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f0f8a45282b]
-/lib64/libggml-base.so.0(+0x16ed9) [0x7f0f8a464ed9]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f0f89bc8bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f0f89bb2d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f0f89bc8ea8]
-/lib64/libggml-vulkan.so.0(+0x173b2) [0x7f0f8a5223b2]
-/lib64/libggml-vulkan.so.0(+0x1075d0) [0x7f0f8a6125d0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f0f8a46e112]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f0f8e089530]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f0f8e08bb05]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f0f8e09226f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f0f8e093bee]
-/usr/sbin/llama-bench() [0x41a1bb]
-/usr/sbin/llama-bench() [0x416ec6]
-/lib64/libc.so.6(+0x35b5) [0x7f0f898995b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f0f89899668]
-/usr/sbin/llama-bench() [0x4189d5]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7f74311b6465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f74311b683b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7f74311c8f59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7f7430f59bfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f7430f43d3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7f7430f59ea8]
+/lib64/libggml-vulkan.so.0(+0x15411) [0x7f7431289411]
+/lib64/libggml-vulkan.so.0(+0xfefd0) [0x7f7431372fd0]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f74311d21d2]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f7434b8c9d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f7434b8f5b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7f7434b95f95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7f7434b97aae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403b2b]
+/lib64/libc.so.6(+0x35b5) [0x7f7430c2a5b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7f7430c2a668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
 ✖ ! [vulkan_amdvlk] gemma-4-31B-it-UD-Q4_K_XL__fa1 __longctx32768 failed (exit 0)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log
index 7913e47..b235fe6 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,23 +1,23 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7fbb0ecdb465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fbb0ecdb83b]
-/lib64/libggml-base.so.0(+0x16f19) [0x7fbb0ecedf19]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7fbb0ea7ebfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7fbb0ea68d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7fbb0ea7eea8]
-/lib64/libggml-vulkan.so.0(+0x1728d) [0x7fbb0edb028d]
-/lib64/libggml-vulkan.so.0(+0x10a410) [0x7fbb0eea3410]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7fbb0ecf7192]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fbb12965c70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7fbb12968255]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7fbb1296e98f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7fbb1297032e]
-/usr/sbin/llama-bench() [0x40663b]
-/usr/sbin/llama-bench() [0x403a5b]
-/lib64/libc.so.6(+0x35b5) [0x7fbb0e74f5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7fbb0e74f668]
-/usr/sbin/llama-bench() [0x404e65]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7fe26f226465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fe26f22683b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7fe26f238f59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7fe26efc9bfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7fe26efb3d3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7fe26efc9ea8]
+/lib64/libggml-vulkan.so.0(+0x15411) [0x7fe26f2f9411]
+/lib64/libggml-vulkan.so.0(+0xfefd0) [0x7fe26f3e2fd0]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7fe26f2421d2]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fe272bfc9d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7fe272bff5b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7fe272c05f95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7fe272c07aae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403b2b]
+/lib64/libc.so.6(+0x35b5) [0x7fe26ec9a5b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7fe26ec9a668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
 ✖ ! [vulkan_amdvlk] gemma-4-31B-it-UD-Q4_K_XL__fa1 __longctx65536 failed (exit 0)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1.log
index a66892a..b85e109 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           pp512 |        244.46 ± 0.27 |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           tg128 |         11.02 ± 0.01 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           pp512 |        301.53 ± 0.16 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           tg128 |         11.45 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log
index 67bf195..a5684ef 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        124.32 ± 1.04 |
-| gemma4 ?B Q4_K - Medium        |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          8.78 ± 0.00 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        146.08 ± 1.04 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          9.09 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log
index f764fbf..ec78316 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         76.03 ± 1.52 |
-| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          7.78 ± 0.00 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         86.61 ± 2.84 |
+| gemma4 31B Q4_K - Medium       |  17.46 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          7.88 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 704cc50..0000000
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        282.42 ± 0.40 |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          6.16 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 1835f21..0000000
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        132.60 ± 1.37 |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.26 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index d77f293..0000000
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         90.15 ± 0.22 |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          5.00 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1.log
deleted file mode 100644
index b397b89..0000000
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        308.17 ± 0.62 |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          6.16 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index fbdfc88..0000000
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        138.85 ± 1.58 |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.26 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 7d96053..0000000
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         93.92 ± 0.29 |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          5.00 ± 0.00 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log
index 8b2bd22..00abf73 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        309.61 ± 0.81 |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          6.15 ± 0.00 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        326.41 ± 1.50 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          6.17 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
index 3e6693a..9981324 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        139.33 ± 1.73 |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.25 ± 0.00 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        150.02 ± 1.17 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.26 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
index ea393fd..9305e1c 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         93.30 ± 0.54 |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          5.00 ± 0.00 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        100.93 ± 0.14 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          5.01 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
index 4c792b4..35a3e6a 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        303.40 ± 0.23 |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          6.07 ± 0.00 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        325.63 ± 1.72 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          6.14 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
index 45ec95f..dd9ae50 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        141.37 ± 1.98 |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.18 ± 0.08 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        155.90 ± 2.50 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.27 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
index 5ae9fb9..d2a4c4f 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         93.78 ± 0.36 |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          4.95 ± 0.00 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        106.02 ± 0.86 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          5.01 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log
index 0ef79f1..ce2694a 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        319.75 ± 0.26 |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          6.12 ± 0.00 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           pp512 |        333.66 ± 1.84 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |  1 |    0 |           tg128 |          6.16 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
index 513ed3b..0b5532a 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        128.00 ± 1.30 |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.30 ± 0.00 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        143.80 ± 1.56 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.30 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
index f00e31f..117f557 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         81.58 ± 0.14 |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          4.99 ± 0.05 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         96.12 ± 0.10 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          5.03 ± 0.01 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
index c3c9edf..7a5c7b9 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           pp512 |         72.26 ± 0.02 |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           tg128 |          6.28 ± 0.00 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           pp512 |         73.59 ± 0.03 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           tg128 |          6.21 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
index cc5da6c..7e3a66c 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,23 +1,24 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7f8de9476465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f8de947682b]
-/lib64/libggml-base.so.0(+0x16ed9) [0x7f8de9488ed9]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f8de8becbfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f8de8bd6d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f8de8becea8]
-/lib64/libggml-vulkan.so.0(+0x173b2) [0x7f8de95463b2]
-/lib64/libggml-vulkan.so.0(+0x1075d0) [0x7f8de96365d0]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f8de9492112]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f8ded0ad530]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f8ded0afb05]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f8ded0b626f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f8ded0b7bee]
-/usr/sbin/llama-bench() [0x41a1bb]
-/usr/sbin/llama-bench() [0x416ec6]
-/lib64/libc.so.6(+0x35b5) [0x7f8de88bd5b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f8de88bd668]
-/usr/sbin/llama-bench() [0x4189d5]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7f995798c465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f995798c83b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7f995799ef59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7f995772fbfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f9957719d3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7f995772fea8]
+/lib64/libggml-vulkan.so.0(+0x1381f) [0x7f9957a5d81f]
+/lib64/libggml-vulkan.so.0(+0x139c0a) [0x7f9957b83c0a]
+/lib64/libggml-vulkan.so.0(+0x13a7e1) [0x7f9957b847e1]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x7f3) [0x7f99579a8613]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f995b3629d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f995b3655b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7f995b36bf95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7f995b36daae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403c06]
+/lib64/libc.so.6(+0x35b5) [0x7f99574005b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7f9957400668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
 ✖ ! [vulkan_amdvlk] gemma-4-31B-it-UD-Q8_K_XL__fa1 __longctx32768 failed (exit 0)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
index f4d2122..14e6d70 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,23 +1,23 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
-/lib64/libggml-base.so.0(+0x4465) [0x7f2f37a93465]
-/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f2f37a9383b]
-/lib64/libggml-base.so.0(+0x16f19) [0x7f2f37aa5f19]
-/lib64/libstdc++.so.6(+0x1ebfc) [0x7f2f37836bfc]
-/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f2f37820d3a]
-/lib64/libstdc++.so.6(+0x1eea8) [0x7f2f37836ea8]
-/lib64/libggml-vulkan.so.0(+0x1728d) [0x7f2f37b6828d]
-/lib64/libggml-vulkan.so.0(+0x10a410) [0x7f2f37c5b410]
-/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7f2f37aaf192]
-/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7f2f3b71dc70]
-/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7f2f3b720255]
-/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x35f) [0x7f2f3b72698f]
-/lib64/libllama.so.0(llama_decode+0xe) [0x7f2f3b72832e]
-/usr/sbin/llama-bench() [0x40663b]
-/usr/sbin/llama-bench() [0x403a5b]
-/lib64/libc.so.6(+0x35b5) [0x7f2f375075b5]
-/lib64/libc.so.6(__libc_start_main+0x88) [0x7f2f37507668]
-/usr/sbin/llama-bench() [0x404e65]
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+/lib64/libggml-base.so.0(+0x4465) [0x7fcccc1e5465]
+/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7fcccc1e583b]
+/lib64/libggml-base.so.0(+0x16f59) [0x7fcccc1f7f59]
+/lib64/libstdc++.so.6(+0x1ebfc) [0x7fcccbf88bfc]
+/lib64/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7fcccbf72d3a]
+/lib64/libstdc++.so.6(+0x1eea8) [0x7fcccbf88ea8]
+/lib64/libggml-vulkan.so.0(+0x15411) [0x7fcccc2b8411]
+/lib64/libggml-vulkan.so.0(+0xfefd0) [0x7fcccc3a1fd0]
+/lib64/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x3b2) [0x7fcccc2011d2]
+/lib64/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa0) [0x7fcccfbbb9d0]
+/lib64/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe5) [0x7fcccfbbe5b5]
+/lib64/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x355) [0x7fcccfbc4f95]
+/lib64/libllama.so.0(llama_decode+0xe) [0x7fcccfbc6aae]
+/usr/sbin/llama-bench() [0x4066fb]
+/usr/sbin/llama-bench() [0x403b2b]
+/lib64/libc.so.6(+0x35b5) [0x7fcccbc595b5]
+/lib64/libc.so.6(__libc_start_main+0x88) [0x7fcccbc59668]
+/usr/sbin/llama-bench() [0x404f35]
 terminate called after throwing an instance of 'vk::DeviceLostError'
   what():  vk::Queue::submit: ErrorDeviceLost
 ✖ ! [vulkan_amdvlk] gemma-4-31B-it-UD-Q8_K_XL__fa1 __longctx65536 failed (exit 0)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1.log
index 9f5b317..5faedde 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           pp512 |        209.20 ± 6.85 |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           tg128 |          6.28 ± 0.00 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           pp512 |        272.10 ± 0.28 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |           tg128 |          6.30 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
index 6b46536..fdba976 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        112.72 ± 1.26 |
-| gemma4 ?B Q8_0                 |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          5.49 ± 0.00 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        141.06 ± 0.76 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          5.52 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
index 7ae88af..83149b4 100644
--- a/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         71.68 ± 0.77 |
-| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          5.02 ± 0.00 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         88.45 ± 3.85 |
+| gemma4 31B Q8_0                |  32.60 GiB |    30.70 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          5.05 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 00caf55..0000000
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |       945.82 ± 60.17 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.25 ± 0.01 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index fcc4071..0000000
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        311.38 ± 1.36 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         31.94 ± 7.16 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index 76a7342..0000000
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        175.99 ± 0.26 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         27.44 ± 0.31 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1.log
deleted file mode 100644
index 6ee9804..0000000
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        634.64 ± 3.80 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.70 ± 0.04 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index 5b9fb05..0000000
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        304.96 ± 1.71 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         35.84 ± 0.12 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 97ab7bd..0000000
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        175.32 ± 0.50 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         24.52 ± 5.66 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1.log
index b9a196c..7b20187 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        635.33 ± 4.03 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.99 ± 0.05 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |      625.58 ± 109.36 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.81 ± 0.04 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log
index ba6f437..0f79407 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        302.31 ± 0.46 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         35.94 ± 0.12 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        589.25 ± 0.34 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         36.32 ± 0.13 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log
index 88e55ac..bdf6e2f 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        175.25 ± 0.49 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         27.78 ± 0.06 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        371.47 ± 2.43 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         26.11 ± 3.50 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log
index 58b8b3f..5420599 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        648.77 ± 4.37 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         49.76 ± 0.02 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        675.76 ± 4.24 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.02 ± 0.07 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
index fc17a5b..2c0c422 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        439.58 ± 0.57 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         39.77 ± 0.03 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        603.27 ± 1.54 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         40.02 ± 0.24 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx65536.log
index f636f42..f9fce77 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        273.87 ± 0.30 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         32.17 ± 0.38 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        379.11 ± 1.07 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         32.99 ± 0.14 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log
index e805b34..627487e 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        641.05 ± 2.79 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.73 ± 0.01 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           pp512 |        684.87 ± 3.91 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.48 ± 0.07 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
index 3360e12..1a544d8 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        305.81 ± 0.24 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         36.03 ± 0.02 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        653.09 ± 1.74 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         36.34 ± 0.22 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log
index dba40ef..11a79c6 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        177.27 ± 0.45 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         27.69 ± 0.13 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        412.83 ± 4.50 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         28.15 ± 0.13 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log
index ccde36b..1358ab6 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           pp512 |        575.01 ± 3.11 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           tg128 |         51.20 ± 0.06 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           pp512 |        664.67 ± 3.10 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           tg128 |         51.42 ± 0.05 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
index bf15ed0..b2a44e8 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        214.22 ± 0.50 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         34.54 ± 0.04 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        234.02 ± 0.46 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         35.87 ± 0.04 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log
index 1aedfef..0822a03 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        129.83 ± 0.04 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         26.17 ± 0.07 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        143.42 ± 0.11 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         27.87 ± 0.02 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log
index 7dfab3a..fd58775 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           pp512 |        635.12 ± 4.23 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           tg128 |         56.31 ± 0.08 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           pp512 |        719.91 ± 4.36 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |           tg128 |         56.61 ± 0.08 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log
index b654b3f..a6e0d21 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        290.19 ± 1.03 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         42.28 ± 0.06 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        307.50 ± 0.37 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         43.01 ± 0.04 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx65536.log b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx65536.log
index 3dd3243..0863e64 100644
--- a/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        139.61 ± 0.56 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         35.00 ± 0.05 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        145.92 ± 0.15 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         34.97 ± 0.10 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index 0593f0d..0000000
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |       2029.48 ± 6.68 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.53 ± 0.11 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 5cf4d57..0000000
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        494.19 ± 1.26 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         51.81 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index 07c8d4a..0000000
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        277.16 ± 1.25 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         40.39 ± 0.04 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1.log
deleted file mode 100644
index da0f1ce..0000000
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1631.69 ± 17.03 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.53 ± 0.10 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index 0c002a7..0000000
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        490.56 ± 2.76 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         51.69 ± 0.16 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index b3b5c05..0000000
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        275.76 ± 1.52 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         40.22 ± 0.34 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1.log
index 80a1680..aeee0ed 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1636.30 ± 15.57 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.58 ± 0.05 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1786.17 ± 19.42 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.05 ± 0.06 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx32768.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx32768.log
index ac847ab..7581448 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx32768.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        487.06 ± 3.04 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         51.73 ± 0.18 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1003.11 ± 2.13 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         52.17 ± 0.17 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx65536.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx65536.log
index 5d62b87..ba58bce 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx65536.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        276.26 ± 0.34 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         40.32 ± 0.08 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        601.63 ± 1.09 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         40.68 ± 0.09 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log
index 5d0098e..62d7a4e 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1694.78 ± 17.20 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.19 ± 0.02 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1787.57 ± 15.77 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.68 ± 0.10 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log
index 293a0cb..e35d23c 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        730.43 ± 1.15 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         57.22 ± 0.03 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1027.43 ± 1.91 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         57.29 ± 0.37 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx65536.log
index 879519f..2f5474a 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        440.23 ± 0.27 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         47.51 ± 0.29 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        613.02 ± 1.23 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         47.67 ± 0.25 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log
index fa84182..80f0ba5 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1635.73 ± 10.16 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         72.47 ± 0.03 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           pp512 |      1812.57 ± 16.28 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |  1 |    0 |           tg128 |         73.03 ± 0.24 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log
index 1c1adcb..651c6f0 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        493.84 ± 1.85 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         52.05 ± 0.04 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1124.14 ± 2.36 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         52.23 ± 0.24 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx65536.log
index af03959..e2f05e7 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        277.70 ± 0.33 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         40.28 ± 0.41 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |        679.48 ± 9.08 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |         40.79 ± 0.14 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log b/benchmark/results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log
index a2ddf3c..fb91016 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1306.44 ± 9.57 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           tg128 |         73.63 ± 0.06 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           pp512 |      1376.47 ± 10.73 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           tg128 |         74.15 ± 0.04 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log
index b322708..778904e 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        365.33 ± 0.21 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         49.83 ± 0.07 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        388.64 ± 0.03 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         51.77 ± 0.12 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx65536.log
index 3c9107e..ec86ef4 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        210.72 ± 0.07 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         38.53 ± 0.01 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        228.50 ± 0.15 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         40.52 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log b/benchmark/results/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log
index 795bbaa..3900f95 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           pp512 |      1574.45 ± 16.63 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           tg128 |         79.03 ± 0.17 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           pp512 |      1692.49 ± 16.61 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |           tg128 |         79.78 ± 0.23 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log b/benchmark/results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log
index c1e124d..5043525 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        545.65 ± 0.37 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         60.70 ± 0.02 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        555.73 ± 1.50 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         61.12 ± 0.08 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx65536.log b/benchmark/results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx65536.log
index ee2ba5a..47e0e34 100644
--- a/benchmark/results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        229.28 ± 0.91 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         50.14 ± 0.11 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        237.74 ± 1.25 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |         50.13 ± 0.03 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1.log b/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1.log
deleted file mode 100644
index e97dc44..0000000
--- a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1512.85 ± 4.84 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.54 ± 0.15 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1__longctx32768.log b/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1__longctx32768.log
deleted file mode 100644
index 8f01f6a..0000000
--- a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        163.39 ± 1.12 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.65 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1__longctx65536.log b/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1__longctx65536.log
deleted file mode 100644
index a44a4d5..0000000
--- a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         72.65 ± 0.35 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.00 ± 0.00 |
-
-build: 7957de9dc (8645)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2__fa1.log b/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2__fa1.log
deleted file mode 100644
index ef4d459..0000000
--- a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2__fa1.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1542.98 ± 7.94 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.52 ± 0.12 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2__fa1__longctx32768.log b/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2__fa1__longctx32768.log
deleted file mode 100644
index 3bbc460..0000000
--- a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2__fa1__longctx32768.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        162.64 ± 0.31 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.64 ± 0.00 |
-
-build: f53577432 (8942)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2__fa1__longctx65536.log b/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2__fa1__longctx65536.log
deleted file mode 100644
index 859f3de..0000000
--- a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_2__fa1__longctx65536.log
+++ /dev/null
@@ -1,8 +0,0 @@
-ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         72.92 ± 0.24 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.00 ± 0.00 |
-
-build: ab6120cde (8997)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_3__fa1.log b/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_3__fa1.log
index 0e24968..f1c5103 100644
--- a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_3__fa1.log
+++ b/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_3__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1543.35 ± 8.00 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.50 ± 0.12 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |      1545.36 ± 22.91 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.59 ± 0.03 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_3__fa1__longctx32768.log b/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_3__fa1__longctx32768.log
index 5f07961..5478ac6 100644
--- a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_3__fa1__longctx32768.log
+++ b/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_3__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        159.17 ± 0.71 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.64 ± 0.00 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        159.95 ± 1.29 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.60 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_3__fa1__longctx65536.log b/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_3__fa1__longctx65536.log
index f85d75e..5c192e4 100644
--- a/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_3__fa1__longctx65536.log
+++ b/benchmark/results/llama-2-7b.Q4_0__rocm-7_2_3__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         72.03 ± 0.23 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         71.44 ± 0.29 |
 | llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.00 ± 0.00 |
 
-build: 8e1f9d083 (9112)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm6_4_4__fa1.log b/benchmark/results/llama-2-7b.Q4_0__rocm6_4_4__fa1.log
index ef866a8..33d9df4 100644
--- a/benchmark/results/llama-2-7b.Q4_0__rocm6_4_4__fa1.log
+++ b/benchmark/results/llama-2-7b.Q4_0__rocm6_4_4__fa1.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1524.35 ± 2.98 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.78 ± 0.02 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |      1532.82 ± 13.52 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         50.83 ± 0.04 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log b/benchmark/results/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log
index 4667568..0374677 100644
--- a/benchmark/results/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log
+++ b/benchmark/results/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        196.44 ± 1.93 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.96 ± 0.00 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        187.91 ± 1.55 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.94 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx65536.log b/benchmark/results/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx65536.log
index 3d8bc49..eb826b8 100644
--- a/benchmark/results/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx65536.log
+++ b/benchmark/results/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
+  Device 0: AMD Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         83.57 ± 0.98 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.74 ± 0.02 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         82.46 ± 0.59 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.76 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log b/benchmark/results/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log
index 27dfff5..ad2b078 100644
--- a/benchmark/results/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log
+++ b/benchmark/results/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |       1513.21 ± 2.79 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.41 ± 0.13 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           pp512 |      1516.11 ± 11.19 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |  1 |    0 |           tg128 |         51.27 ± 0.05 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log b/benchmark/results/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log
index 182d7c6..9ffdff9 100644
--- a/benchmark/results/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log
+++ b/benchmark/results/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        189.28 ± 0.40 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          5.62 ± 0.00 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |        204.17 ± 1.96 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.00 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx65536.log b/benchmark/results/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx65536.log
index c32cd3d..3ca0151 100644
--- a/benchmark/results/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx65536.log
+++ b/benchmark/results/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx65536.log
@@ -2,7 +2,7 @@ ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
 | model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         85.33 ± 0.30 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          2.90 ± 0.03 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d65536 |         94.78 ± 1.11 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d65536 |          3.79 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log b/benchmark/results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log
index 8d077eb..8e20b9d 100644
--- a/benchmark/results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log
+++ b/benchmark/results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           pp512 |        326.38 ± 0.44 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           tg128 |         55.44 ± 0.11 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           pp512 |        347.25 ± 0.61 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           tg128 |         55.69 ± 0.15 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log b/benchmark/results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log
index 2853f6a..e88a013 100644
--- a/benchmark/results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log
+++ b/benchmark/results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        145.14 ± 0.20 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.21 ± 0.00 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        155.90 ± 0.09 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          9.42 ± 0.00 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx65536.log b/benchmark/results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx65536.log
index 9685755..6465b83 100644
--- a/benchmark/results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx65536.log
+++ b/benchmark/results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         89.90 ± 0.68 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          4.61 ± 0.00 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        102.05 ± 0.15 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          5.12 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__vulkan_radv__fa1.log b/benchmark/results/llama-2-7b.Q4_0__vulkan_radv__fa1.log
index 68e0eb3..02ba413 100644
--- a/benchmark/results/llama-2-7b.Q4_0__vulkan_radv__fa1.log
+++ b/benchmark/results/llama-2-7b.Q4_0__vulkan_radv__fa1.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1311.79 ± 0.38 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           tg128 |         55.59 ± 0.03 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           pp512 |       1337.70 ± 0.75 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |           tg128 |         55.73 ± 0.04 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log b/benchmark/results/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log
index d0a49c1..72add42 100644
--- a/benchmark/results/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log
+++ b/benchmark/results/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        210.35 ± 0.60 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          8.53 ± 0.01 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |        216.79 ± 1.07 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          8.47 ± 0.01 |
 
-build: 3f8752b55 (8743)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx65536.log b/benchmark/results/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx65536.log
index 6e478db..ef64030 100644
--- a/benchmark/results/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx65536.log
+++ b/benchmark/results/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx65536.log
@@ -1,8 +1,8 @@
 ggml_vulkan: Found 1 Vulkan devices:
-ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
+ggml_vulkan: 0 = AMD Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
 | model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
 | ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |        104.40 ± 1.78 |
-| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          4.60 ± 0.00 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d65536 |         79.73 ± 0.65 |
+| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d65536 |          4.58 ± 0.00 |
 
-build: ab6120cde (8997)
+build: 0253fb21f (9187)
diff --git a/benchmark/results/system_info.json b/benchmark/results/system_info.json
index d9ce179..fdbbdd0 100644
--- a/benchmark/results/system_info.json
+++ b/benchmark/results/system_info.json
@@ -1 +1 @@
-{"distro": "Fedora Linux 43 (Workstation Edition)", "kernel": "6.19.9-200.fc43.x86_64", "linux_firmware": "linux-firmware-20260309-1.fc43.noarch", "timestamp": "10 Apr 2026"}
+{"distro": "Fedora Linux 43 (Workstation Edition)", "kernel": "6.19.12-200.fc43.x86_64", "linux_firmware": "linux-firmware-20260309-1.fc43.noarch", "timestamp": "16 May 2026"}
diff --git a/docs/results.json b/docs/results.json
index 2f3bbe4..9988617 100644
--- a/docs/results.json
+++ b/docs/results.json
@@ -1,38 +1,23 @@
 {
   "meta": {
-    "generated_at": "2026-05-12T11:31:47Z",
+    "generated_at": "2026-05-18T06:42:13Z",
     "system_info": {
       "distro": "Fedora Linux 43 (Workstation Edition)",
-      "kernel": "6.19.9-200.fc43.x86_64",
+      "kernel": "6.19.12-200.fc43.x86_64",
       "linux_firmware": "linux-firmware-20260309-1.fc43.noarch",
-      "timestamp": "10 Apr 2026"
+      "timestamp": "16 May 2026"
     },
     "llamacpp_builds": [
       {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       },
       {
-        "hash": "7957de9dc",
-        "number": "8645"
-      },
-      {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      },
-      {
-        "hash": "ab6120cde",
-        "number": "8997"
-      },
-      {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     ],
     "environments": [
-      "rocm-7_2",
-      "rocm-7_2_2",
-      "rocm-7_2_2-pr21344",
       "rocm-7_2_3",
       "rocm6_4_4",
       "rocm7-nightlies",
@@ -42,902 +27,18 @@
     "notes": "pp512 = prompt processing; tg128 = text generation; t/s = tokens/second"
   },
   "runs": [
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 72.06,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 2.99,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 16.76,
-      "tps_std": 0.05,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 2.08,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 7.94,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 1.49,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 77.83,
-      "tps_std": 0.1,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 2.95,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 17.1,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 2.06,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 8.12,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 1.48,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 78.4,
-      "tps_std": 0.1,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 2.76,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 17.57,
-      "tps_std": 0.05,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 2.09,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 8.16,
-      "tps_std": 0.08,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 1.58,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 78.06,
-      "tps_std": 0.06,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 2.95,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 18.34,
-      "tps_std": 0.14,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 2.06,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 8.45,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 1.43,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": null,
-      "tps_mean": null,
-      "tps_std": null,
-      "error": true,
-      "error_type": "runtime",
-      "backend": null,
-      "ngl": null,
-      "mmap": null,
-      "params_b": null,
-      "file_size_gib": null,
-      "name_params_b": 123.0,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": null
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": null,
-      "tps_mean": null,
-      "tps_std": null,
-      "error": true,
-      "error_type": "runtime",
-      "backend": null,
-      "ngl": null,
-      "mmap": null,
-      "params_b": null,
-      "file_size_gib": null,
-      "name_params_b": 123.0,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": null
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": null,
-      "tps_mean": null,
-      "tps_std": null,
-      "error": true,
-      "error_type": "runtime",
-      "backend": null,
-      "ngl": null,
-      "mmap": null,
-      "params_b": null,
-      "file_size_gib": null,
-      "name_params_b": 123.0,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": null
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 46.91,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 2.99,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 7.32,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 2.59,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 125.03,
-      "file_size_gib": 70.31,
-      "name_params_b": 125.03,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002",
-      "model_clean": "Devstral-2-123B-Instruct-2512-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": null,
-      "tps_mean": null,
-      "tps_std": null,
-      "error": true,
-      "error_type": "runtime",
-      "backend": null,
-      "ngl": null,
-      "mmap": null,
-      "params_b": null,
-      "file_size_gib": null,
-      "name_params_b": 123.0,
-      "quant": "Q4_K_XL",
-      "log": "results/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": null
-    },
     {
       "model": "GLM-4.7-Flash-BF16-00001-of-00002",
       "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 403.45,
-      "tps_std": 2.11,
+      "tps_mean": 422.28,
+      "tps_std": 1.78,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -947,24 +48,80 @@
       "file_size_gib": 55.79,
       "name_params_b": 29.94,
       "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log",
+      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "GLM-4.7-Flash-BF16-00001-of-00002",
       "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 21.04,
+      "tps_mean": 23.17,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 29.94,
+      "file_size_gib": 55.79,
+      "name_params_b": 29.94,
+      "quant": "BF16",
+      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
+      "model_clean": "GLM-4.7-Flash-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 91.33,
+      "tps_std": 0.25,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 29.94,
+      "file_size_gib": 55.79,
+      "name_params_b": 29.94,
+      "quant": "BF16",
+      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
+      "model_clean": "GLM-4.7-Flash-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 16.56,
       "tps_std": 0.01,
       "error": false,
       "error_type": null,
@@ -975,81 +132,25 @@
       "file_size_gib": 55.79,
       "name_params_b": 29.94,
       "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log",
+      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "GLM-4.7-Flash-BF16-00001-of-00002",
       "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 87.65,
-      "tps_std": 0.31,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 15.25,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 46.54,
-      "tps_std": 0.17,
+      "tps_mean": 48.38,
+      "tps_std": 0.18,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -1059,24 +160,24 @@
       "file_size_gib": 55.79,
       "name_params_b": 29.94,
       "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log",
+      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "GLM-4.7-Flash-BF16-00001-of-00002",
       "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 11.91,
+      "tps_mean": 12.9,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -1087,49 +188,80 @@
       "file_size_gib": 55.79,
       "name_params_b": 29.94,
       "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log",
+      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "GLM-4.7-Flash-BF16-00001-of-00002",
       "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
       "fa": true,
       "context": "default",
       "context_tokens": null,
-      "test": null,
-      "tps_mean": null,
-      "tps_std": null,
+      "test": "pp512",
+      "tps_mean": 407.51,
+      "tps_std": 1.36,
       "error": false,
       "error_type": null,
-      "backend": null,
-      "ngl": null,
-      "mmap": null,
-      "params_b": null,
-      "file_size_gib": null,
-      "name_params_b": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 29.94,
+      "file_size_gib": 55.79,
+      "name_params_b": 29.94,
       "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1.log",
+      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log",
       "rpc": false,
-      "build": null
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
     },
     {
       "model": "GLM-4.7-Flash-BF16-00001-of-00002",
       "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 22.0,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 29.94,
+      "file_size_gib": 55.79,
+      "name_params_b": 29.94,
+      "quant": "BF16",
+      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
+      "model_clean": "GLM-4.7-Flash-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 87.91,
+      "tps_mean": 97.18,
       "tps_std": 0.2,
       "error": false,
       "error_type": null,
@@ -1140,24 +272,24 @@
       "file_size_gib": 55.79,
       "name_params_b": 29.94,
       "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log",
+      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "GLM-4.7-Flash-BF16-00001-of-00002",
       "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 14.84,
+      "tps_mean": 16.38,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -1168,52 +300,24 @@
       "file_size_gib": 55.79,
       "name_params_b": 29.94,
       "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log",
+      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "GLM-4.7-Flash-BF16-00001-of-00002",
       "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 46.47,
-      "tps_std": 0.09,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 11.63,
+      "tps_mean": 51.9,
       "tps_std": 0.19,
       "error": false,
       "error_type": null,
@@ -1224,344 +328,11 @@
       "file_size_gib": 55.79,
       "name_params_b": 29.94,
       "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 403.2,
-      "tps_std": 2.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 20.43,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 87.63,
-      "tps_std": 0.21,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 15.31,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 46.45,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 11.98,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm-7_2",
-      "env_base": "rocm",
-      "env_variant": "7_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": null,
-      "tps_mean": null,
-      "tps_std": null,
-      "error": false,
-      "error_type": null,
-      "backend": null,
-      "ngl": null,
-      "mmap": null,
-      "params_b": null,
-      "file_size_gib": null,
-      "name_params_b": null,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1.log",
-      "rpc": false,
-      "build": null
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 392.81,
-      "tps_std": 1.96,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 18.97,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 94.55,
-      "tps_std": 0.17,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 14.89,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 50.95,
-      "tps_std": 0.07,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1574,8 +345,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 11.15,
-      "tps_std": 0.09,
+      "tps_mean": 12.82,
+      "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -1588,8 +359,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1602,8 +373,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 400.57,
-      "tps_std": 4.1,
+      "tps_mean": 424.65,
+      "tps_std": 3.93,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -1616,8 +387,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1630,7 +401,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 19.39,
+      "tps_mean": 23.3,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -1644,8 +415,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1658,8 +429,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 86.01,
-      "tps_std": 0.04,
+      "tps_mean": 90.71,
+      "tps_std": 0.38,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -1672,8 +443,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1686,8 +457,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 14.86,
-      "tps_std": 0.03,
+      "tps_mean": 16.55,
+      "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -1700,8 +471,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1714,35 +485,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 45.39,
-      "tps_std": 0.36,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 55.79,
-      "name_params_b": 29.94,
-      "quant": "BF16",
-      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
-      "model_clean": "GLM-4.7-Flash-BF16",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 11.22,
+      "tps_mean": 47.61,
       "tps_std": 0.13,
       "error": false,
       "error_type": null,
@@ -1756,8 +499,36 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "GLM-4.7-Flash-BF16-00001-of-00002",
+      "model_clean": "GLM-4.7-Flash-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 12.92,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 29.94,
+      "file_size_gib": 55.79,
+      "name_params_b": 29.94,
+      "quant": "BF16",
+      "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1770,8 +541,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 114.94,
-      "tps_std": 0.15,
+      "tps_mean": 115.47,
+      "tps_std": 0.14,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -1784,8 +555,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1798,7 +569,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 10.48,
+      "tps_mean": 10.78,
       "tps_std": 0.01,
       "error": false,
       "error_type": null,
@@ -1812,8 +583,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1826,7 +597,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 10.14,
+      "tps_mean": 10.42,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -1840,8 +611,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1854,7 +625,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 5.04,
+      "tps_mean": 5.4,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -1868,8 +639,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1907,8 +678,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 326.36,
-      "tps_std": 4.62,
+      "tps_mean": 361.24,
+      "tps_std": 1.76,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -1921,8 +692,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1935,8 +706,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 9.5,
-      "tps_std": 0.02,
+      "tps_mean": 9.46,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -1949,8 +720,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1963,8 +734,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 102.0,
-      "tps_std": 0.06,
+      "tps_mean": 109.96,
+      "tps_std": 0.16,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -1977,8 +748,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -1991,8 +762,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 7.82,
-      "tps_std": 0.02,
+      "tps_mean": 7.96,
+      "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -2005,8 +776,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2019,8 +790,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 61.52,
-      "tps_std": 0.01,
+      "tps_mean": 65.21,
+      "tps_std": 0.03,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -2033,8 +804,8 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2047,7 +818,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 6.57,
+      "tps_mean": 6.9,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -2061,22 +832,22 @@
       "log": "results/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "GLM-4.7-Flash-UD-Q8_K_XL",
       "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1058.1,
-      "tps_std": 2.19,
+      "tps_mean": 913.36,
+      "tps_std": 47.37,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -2086,24 +857,24 @@
       "file_size_gib": 32.7,
       "name_params_b": 29.94,
       "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log",
+      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "GLM-4.7-Flash-UD-Q8_K_XL",
       "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 33.18,
+      "tps_mean": 35.5,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -2114,25 +885,25 @@
       "file_size_gib": 32.7,
       "name_params_b": 29.94,
       "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log",
+      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "GLM-4.7-Flash-UD-Q8_K_XL",
       "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 91.21,
-      "tps_std": 1.08,
+      "tps_mean": 95.82,
+      "tps_std": 0.14,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -2142,25 +913,25 @@
       "file_size_gib": 32.7,
       "name_params_b": 29.94,
       "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
+      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "GLM-4.7-Flash-UD-Q8_K_XL",
       "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 20.62,
-      "tps_std": 0.01,
+      "tps_mean": 21.96,
+      "tps_std": 0.02,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -2170,192 +941,24 @@
       "file_size_gib": 32.7,
       "name_params_b": 29.94,
       "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
+      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "GLM-4.7-Flash-UD-Q8_K_XL",
       "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 47.63,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 14.98,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 889.16,
-      "tps_std": 36.98,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 32.61,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 91.01,
-      "tps_std": 1.51,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 20.38,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 47.62,
+      "tps_mean": 49.39,
       "tps_std": 0.15,
       "error": false,
       "error_type": null,
@@ -2366,179 +969,11 @@
       "file_size_gib": 32.7,
       "name_params_b": 29.94,
       "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 14.87,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 905.99,
-      "tps_std": 2.05,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 32.65,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 91.59,
-      "tps_std": 0.16,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 20.36,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 47.51,
-      "tps_std": 0.05,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2551,8 +986,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 14.86,
-      "tps_std": 0.0,
+      "tps_mean": 15.86,
+      "tps_std": 0.2,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -2565,8 +1000,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2579,8 +1014,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 903.01,
-      "tps_std": 2.32,
+      "tps_mean": 929.42,
+      "tps_std": 1.08,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -2593,8 +1028,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2607,8 +1042,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 30.77,
-      "tps_std": 0.0,
+      "tps_mean": 33.78,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -2621,8 +1056,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2635,8 +1070,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 99.7,
-      "tps_std": 0.23,
+      "tps_mean": 102.84,
+      "tps_std": 0.07,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -2649,8 +1084,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2663,8 +1098,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 19.93,
-      "tps_std": 0.0,
+      "tps_mean": 21.41,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -2677,8 +1112,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2691,7 +1126,35 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 51.79,
+      "tps_mean": 53.18,
+      "tps_std": 0.2,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 29.94,
+      "file_size_gib": 32.7,
+      "name_params_b": 29.94,
+      "quant": "Q8_K_XL",
+      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
+      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 15.67,
       "tps_std": 0.1,
       "error": false,
       "error_type": null,
@@ -2705,36 +1168,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "model_clean": "GLM-4.7-Flash-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 14.36,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 29.94,
-      "file_size_gib": 32.7,
-      "name_params_b": 29.94,
-      "quant": "Q8_K_XL",
-      "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2747,8 +1182,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 888.7,
-      "tps_std": 38.05,
+      "tps_mean": 931.06,
+      "tps_std": 6.95,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -2761,8 +1196,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2775,7 +1210,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 32.81,
+      "tps_mean": 35.52,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -2789,8 +1224,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2803,8 +1238,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 89.53,
-      "tps_std": 0.09,
+      "tps_mean": 94.94,
+      "tps_std": 0.61,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -2817,8 +1252,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2831,7 +1266,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 20.47,
+      "tps_mean": 22.0,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -2845,8 +1280,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2859,8 +1294,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 46.43,
-      "tps_std": 0.14,
+      "tps_mean": 48.9,
+      "tps_std": 0.07,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -2873,8 +1308,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2887,8 +1322,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 14.43,
-      "tps_std": 0.0,
+      "tps_mean": 15.99,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -2901,8 +1336,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2915,8 +1350,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 499.74,
-      "tps_std": 1.49,
+      "tps_mean": 543.44,
+      "tps_std": 1.1,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -2929,8 +1364,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -2943,7 +1378,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 39.02,
+      "tps_mean": 39.66,
       "tps_std": 0.02,
       "error": false,
       "error_type": null,
@@ -2957,8 +1392,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3021,8 +1456,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 875.8,
-      "tps_std": 2.04,
+      "tps_mean": 1092.07,
+      "tps_std": 1.21,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -3035,8 +1470,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3049,8 +1484,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 40.19,
-      "tps_std": 1.39,
+      "tps_mean": 41.17,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -3063,8 +1498,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3077,8 +1512,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 134.1,
-      "tps_std": 0.07,
+      "tps_mean": 136.78,
+      "tps_std": 0.02,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -3091,8 +1526,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3105,8 +1540,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 21.18,
-      "tps_std": 0.03,
+      "tps_mean": 22.55,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -3119,8 +1554,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3133,8 +1568,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 72.8,
-      "tps_std": 0.0,
+      "tps_mean": 74.06,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -3147,8 +1582,8 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3161,7 +1596,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 14.39,
+      "tps_mean": 15.7,
       "tps_std": 0.01,
       "error": false,
       "error_type": null,
@@ -3175,284 +1610,10 @@
       "log": "results/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 356.93,
-      "tps_std": 1.86,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 22.97,
-      "tps_std": 0.13,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 65.86,
-      "tps_std": 0.58,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 6.1,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": null,
-      "tps_mean": null,
-      "tps_std": null,
-      "error": true,
-      "error_type": "load",
-      "backend": null,
-      "ngl": null,
-      "mmap": null,
-      "params_b": null,
-      "file_size_gib": null,
-      "name_params_b": null,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": null
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 236.39,
-      "tps_std": 1.24,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 22.64,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 64.53,
-      "tps_std": 0.53,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 6.04,
-      "tps_std": 0.17,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": null,
-      "tps_mean": null,
-      "tps_std": null,
-      "error": true,
-      "error_type": "load",
-      "backend": null,
-      "ngl": null,
-      "mmap": null,
-      "params_b": null,
-      "file_size_gib": null,
-      "name_params_b": null,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": null
-    },
     {
       "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
       "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
@@ -3463,8 +1624,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 236.96,
-      "tps_std": 1.25,
+      "tps_mean": 243.25,
+      "tps_std": 1.32,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -3477,8 +1638,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3491,7 +1652,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 22.81,
+      "tps_mean": 23.11,
       "tps_std": 0.01,
       "error": false,
       "error_type": null,
@@ -3505,8 +1666,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3519,8 +1680,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 62.13,
-      "tps_std": 0.56,
+      "tps_mean": 142.92,
+      "tps_std": 0.05,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -3533,8 +1694,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3547,8 +1708,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 6.15,
-      "tps_std": 0.0,
+      "tps_mean": 6.19,
+      "tps_std": 0.05,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -3561,8 +1722,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3600,8 +1761,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 236.56,
-      "tps_std": 1.44,
+      "tps_mean": 238.97,
+      "tps_std": 1.17,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -3614,8 +1775,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -3628,8 +1789,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 21.55,
-      "tps_std": 0.0,
+      "tps_mean": 21.98,
+      "tps_std": 0.02,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -3642,8 +1803,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -3656,8 +1817,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 77.34,
-      "tps_std": 1.21,
+      "tps_mean": 119.83,
+      "tps_std": 0.18,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -3670,8 +1831,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -3684,8 +1845,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 6.54,
-      "tps_std": 0.09,
+      "tps_mean": 6.73,
+      "tps_std": 0.18,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -3698,8 +1859,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -3737,8 +1898,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 225.65,
-      "tps_std": 0.79,
+      "tps_mean": 235.98,
+      "tps_std": 0.63,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -3751,8 +1912,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3765,8 +1926,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 21.44,
-      "tps_std": 0.01,
+      "tps_mean": 22.13,
+      "tps_std": 0.03,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -3779,8 +1940,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3793,8 +1954,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 64.1,
-      "tps_std": 0.5,
+      "tps_mean": 156.27,
+      "tps_std": 0.25,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -3807,8 +1968,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3821,8 +1982,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 6.35,
-      "tps_std": 0.28,
+      "tps_mean": 6.45,
+      "tps_std": 0.03,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -3835,8 +1996,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3874,8 +2035,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 202.08,
-      "tps_std": 0.31,
+      "tps_mean": 205.56,
+      "tps_std": 0.83,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -3888,8 +2049,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3902,7 +2063,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 24.94,
+      "tps_mean": 25.91,
       "tps_std": 0.01,
       "error": false,
       "error_type": null,
@@ -3916,8 +2077,8 @@
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -3930,7 +2091,259 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 31.48,
+      "tps_mean": 30.23,
+      "tps_std": 0.05,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 228.69,
+      "file_size_gib": 87.2,
+      "name_params_b": 228.69,
+      "quant": "Q3_K_S",
+      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
+      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 10.74,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 228.69,
+      "file_size_gib": 87.2,
+      "name_params_b": 228.69,
+      "quant": "Q3_K_S",
+      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
+      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 13.9,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 228.69,
+      "file_size_gib": 87.2,
+      "name_params_b": 228.69,
+      "quant": "Q3_K_S",
+      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
+      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 6.67,
+      "tps_std": 0.05,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 228.69,
+      "file_size_gib": 87.2,
+      "name_params_b": 228.69,
+      "quant": "Q3_K_S",
+      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
+      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 242.43,
+      "tps_std": 0.92,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 228.69,
+      "file_size_gib": 87.2,
+      "name_params_b": 228.69,
+      "quant": "Q3_K_S",
+      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
+      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 31.08,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 228.69,
+      "file_size_gib": 87.2,
+      "name_params_b": 228.69,
+      "quant": "Q3_K_S",
+      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
+      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 33.86,
+      "tps_std": 0.21,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 228.69,
+      "file_size_gib": 87.2,
+      "name_params_b": 228.69,
+      "quant": "Q3_K_S",
+      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
+      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 14.65,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 228.69,
+      "file_size_gib": 87.2,
+      "name_params_b": 228.69,
+      "quant": "Q3_K_S",
+      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
+      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 13.79,
+      "tps_std": 0.08,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 228.69,
+      "file_size_gib": 87.2,
+      "name_params_b": 228.69,
+      "quant": "Q3_K_S",
+      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
+      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 8.84,
       "tps_std": 0.09,
       "error": false,
       "error_type": null,
@@ -3941,663 +2354,25 @@
       "file_size_gib": 87.2,
       "name_params_b": 228.69,
       "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 10.27,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": null,
-      "tps_mean": null,
-      "tps_std": null,
-      "error": true,
-      "error_type": "runtime",
-      "backend": null,
-      "ngl": null,
-      "mmap": null,
-      "params_b": null,
-      "file_size_gib": null,
-      "name_params_b": null,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": null
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 231.25,
-      "tps_std": 0.79,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 30.7,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 32.0,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 14.47,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 87.2,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_S",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_S-00001-of-00003",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_S",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": null,
-      "tps_mean": null,
-      "tps_std": null,
-      "error": true,
-      "error_type": "runtime",
-      "backend": null,
-      "ngl": null,
-      "mmap": null,
-      "params_b": null,
-      "file_size_gib": null,
-      "name_params_b": null,
-      "quant": "Q3_K_S",
       "log": "results/MiniMax-M2.7-UD-Q3_K_S-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
-      "build": null
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 245.87,
-      "tps_std": 1.58,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm6_4_4__fa1.log",
-      "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 22.72,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 81.2,
-      "tps_std": 1.25,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 6.69,
-      "tps_std": 0.12,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 238.13,
-      "tps_std": 1.15,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 22.54,
-      "tps_std": 0.27,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 61.41,
-      "tps_std": 0.44,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 6.64,
-      "tps_std": 0.15,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 203.65,
-      "tps_std": 1.0,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 24.07,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 31.64,
-      "tps_std": 0.19,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 10.06,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 230.72,
-      "tps_std": 8.67,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 29.48,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 32.06,
-      "tps_std": 0.16,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004",
-      "model_clean": "MiniMax-M2.7-UD-Q3_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 14.12,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 228.69,
-      "file_size_gib": 94.93,
-      "name_params_b": 228.69,
-      "quant": "Q3_K_XL",
-      "log": "results/MiniMax-M2.7-UD-Q3_K_XL-00001-of-00004__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "Ministral-3-14B-Instruct-2512-BF16",
       "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 745.5,
-      "tps_std": 9.08,
+      "tps_mean": 801.06,
+      "tps_std": 9.05,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -4607,24 +2382,24 @@
       "file_size_gib": 25.16,
       "name_params_b": 13.51,
       "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1.log",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
       "model": "Ministral-3-14B-Instruct-2512-BF16",
       "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 8.42,
+      "tps_mean": 8.5,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -4635,25 +2410,25 @@
       "file_size_gib": 25.16,
       "name_params_b": 13.51,
       "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1.log",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
       "model": "Ministral-3-14B-Instruct-2512-BF16",
       "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 164.19,
-      "tps_std": 1.96,
+      "tps_mean": 349.46,
+      "tps_std": 1.14,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -4663,24 +2438,24 @@
       "file_size_gib": 25.16,
       "name_params_b": 13.51,
       "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1__longctx32768.log",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
       "model": "Ministral-3-14B-Instruct-2512-BF16",
       "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 7.08,
+      "tps_mean": 7.12,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -4691,25 +2466,25 @@
       "file_size_gib": 25.16,
       "name_params_b": 13.51,
       "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1__longctx32768.log",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
       "model": "Ministral-3-14B-Instruct-2512-BF16",
       "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 75.02,
-      "tps_std": 1.56,
+      "tps_mean": 209.94,
+      "tps_std": 1.05,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -4719,24 +2494,24 @@
       "file_size_gib": 25.16,
       "name_params_b": 13.51,
       "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1__longctx65536.log",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
       "model": "Ministral-3-14B-Instruct-2512-BF16",
       "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 6.11,
+      "tps_mean": 6.14,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -4747,179 +2522,11 @@
       "file_size_gib": 25.16,
       "name_params_b": 13.51,
       "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2-pr21344__fa1__longctx65536.log",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 732.09,
-      "tps_std": 2.49,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 8.42,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 182.01,
-      "tps_std": 2.08,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 7.07,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 80.08,
-      "tps_std": 2.35,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 6.11,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -4932,8 +2539,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 716.84,
-      "tps_std": 2.31,
+      "tps_mean": 762.49,
+      "tps_std": 6.93,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -4946,8 +2553,8 @@
       "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -4960,7 +2567,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 8.39,
+      "tps_mean": 8.49,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -4974,8 +2581,8 @@
       "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -4988,7 +2595,175 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 166.12,
+      "tps_mean": 318.28,
+      "tps_std": 0.86,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 13.51,
+      "file_size_gib": 25.16,
+      "name_params_b": 13.51,
+      "quant": "BF16",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Ministral-3-14B-Instruct-2512-BF16",
+      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 7.11,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 13.51,
+      "file_size_gib": 25.16,
+      "name_params_b": 13.51,
+      "quant": "BF16",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Ministral-3-14B-Instruct-2512-BF16",
+      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 188.86,
+      "tps_std": 0.24,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 13.51,
+      "file_size_gib": 25.16,
+      "name_params_b": 13.51,
+      "quant": "BF16",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Ministral-3-14B-Instruct-2512-BF16",
+      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 6.14,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 13.51,
+      "file_size_gib": 25.16,
+      "name_params_b": 13.51,
+      "quant": "BF16",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Ministral-3-14B-Instruct-2512-BF16",
+      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 808.74,
+      "tps_std": 14.71,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 13.51,
+      "file_size_gib": 25.16,
+      "name_params_b": 13.51,
+      "quant": "BF16",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Ministral-3-14B-Instruct-2512-BF16",
+      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 8.51,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 13.51,
+      "file_size_gib": 25.16,
+      "name_params_b": 13.51,
+      "quant": "BF16",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Ministral-3-14B-Instruct-2512-BF16",
+      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 395.07,
       "tps_std": 0.76,
       "error": false,
       "error_type": null,
@@ -4999,179 +2774,11 @@
       "file_size_gib": 25.16,
       "name_params_b": 13.51,
       "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 7.05,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 71.24,
-      "tps_std": 0.32,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 6.09,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 726.92,
-      "tps_std": 2.95,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 8.43,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 162.4,
-      "tps_std": 0.45,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
       "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -5184,7 +2791,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 7.07,
+      "tps_mean": 7.12,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -5198,8 +2805,8 @@
       "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -5212,8 +2819,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 73.23,
-      "tps_std": 0.94,
+      "tps_mean": 239.54,
+      "tps_std": 0.33,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -5226,8 +2833,8 @@
       "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -5240,7 +2847,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 6.11,
+      "tps_mean": 6.14,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -5254,8 +2861,8 @@
       "log": "results/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -5343,8 +2950,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 164.2,
-      "tps_std": 0.55,
+      "tps_mean": 255.05,
+      "tps_std": 0.7,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -5357,8 +2964,8 @@
       "log": "results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -5371,7 +2978,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 7.94,
+      "tps_mean": 7.83,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -5385,8 +2992,8 @@
       "log": "results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -5399,570 +3006,402 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 72.41,
-      "tps_std": 0.21,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 6.71,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 42.48,
-      "tps_std": 0.11,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Ministral-3-14B-Instruct-2512-BF16",
-      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 5.81,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 13.51,
-      "file_size_gib": 25.16,
-      "name_params_b": 13.51,
-      "quant": "BF16",
-      "log": "results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 333.46,
-      "tps_std": 1.2,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 15.99,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 307.47,
-      "tps_std": 0.55,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 15.6,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 250.79,
-      "tps_std": 0.32,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 15.03,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 267.67,
-      "tps_std": 1.7,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 16.18,
-      "tps_std": 0.07,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 304.86,
-      "tps_std": 0.24,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 15.79,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 249.66,
-      "tps_std": 0.9,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 15.42,
-      "tps_std": 0.07,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 261.51,
-      "tps_std": 1.5,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 15.63,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 323.83,
-      "tps_std": 0.28,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 15.34,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 290.67,
+      "tps_mean": 79.54,
       "tps_std": 0.26,
       "error": false,
       "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 13.51,
+      "file_size_gib": 25.16,
+      "name_params_b": 13.51,
+      "quant": "BF16",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Ministral-3-14B-Instruct-2512-BF16",
+      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 6.62,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 13.51,
+      "file_size_gib": 25.16,
+      "name_params_b": 13.51,
+      "quant": "BF16",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Ministral-3-14B-Instruct-2512-BF16",
+      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 35.49,
+      "tps_std": 0.19,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 13.51,
+      "file_size_gib": 25.16,
+      "name_params_b": 13.51,
+      "quant": "BF16",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Ministral-3-14B-Instruct-2512-BF16",
+      "model_clean": "Ministral-3-14B-Instruct-2512-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 5.74,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 13.51,
+      "file_size_gib": 25.16,
+      "name_params_b": 13.51,
+      "quant": "BF16",
+      "log": "results/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 275.89,
+      "tps_std": 1.53,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 14.85,
+      "tps_std": 0.04,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 382.38,
+      "tps_std": 5.98,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 14.4,
+      "tps_std": 0.07,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 348.1,
+      "tps_std": 0.16,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 13.96,
+      "tps_std": 0.16,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 272.39,
+      "tps_std": 5.98,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 14.5,
+      "tps_std": 0.06,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 379.64,
+      "tps_std": 0.46,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 14.15,
+      "tps_std": 0.16,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 333.18,
+      "tps_std": 0.36,
+      "error": false,
+      "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
@@ -5973,8 +3412,8 @@
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -5987,8 +3426,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 15.11,
-      "tps_std": 0.05,
+      "tps_mean": 13.88,
+      "tps_std": 0.17,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -6001,8 +3440,8 @@
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -6015,8 +3454,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 232.54,
-      "tps_std": 0.27,
+      "tps_mean": 239.02,
+      "tps_std": 1.74,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -6029,8 +3468,8 @@
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -6043,8 +3482,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 15.81,
-      "tps_std": 0.0,
+      "tps_mean": 14.43,
+      "tps_std": 0.02,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -6057,8 +3496,8 @@
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -6071,8 +3510,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 284.15,
-      "tps_std": 0.23,
+      "tps_mean": 353.69,
+      "tps_std": 0.07,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -6085,8 +3524,8 @@
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -6099,8 +3538,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 15.34,
-      "tps_std": 0.61,
+      "tps_mean": 14.32,
+      "tps_std": 0.08,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -6113,8 +3552,8 @@
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -6127,35 +3566,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 242.09,
-      "tps_std": 0.33,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 15.47,
+      "tps_mean": 326.04,
       "tps_std": 0.1,
       "error": false,
       "error_type": null,
@@ -6169,8 +3580,36 @@
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 14.05,
+      "tps_std": 0.19,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -6183,8 +3622,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 138.78,
-      "tps_std": 0.58,
+      "tps_mean": 160.64,
+      "tps_std": 0.6,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -6197,8 +3636,8 @@
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -6211,8 +3650,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 14.45,
-      "tps_std": 0.09,
+      "tps_mean": 14.44,
+      "tps_std": 0.1,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -6225,8 +3664,8 @@
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -6239,8 +3678,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 104.96,
-      "tps_std": 0.02,
+      "tps_mean": 120.39,
+      "tps_std": 0.1,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -6253,8 +3692,8 @@
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -6267,35 +3706,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 13.66,
-      "tps_std": 0.21,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 85.09,
+      "tps_mean": 13.67,
       "tps_std": 0.16,
       "error": false,
       "error_type": null,
@@ -6306,11 +3717,39 @@
       "file_size_gib": 78.02,
       "name_params_b": 120.67,
       "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 96.89,
+      "tps_std": 0.18,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -6323,8 +3762,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 13.21,
-      "tps_std": 0.04,
+      "tps_mean": 13.3,
+      "tps_std": 0.25,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -6337,8 +3776,8 @@
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -6351,8 +3790,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 191.97,
-      "tps_std": 8.53,
+      "tps_mean": 270.56,
+      "tps_std": 2.02,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -6365,8 +3804,8 @@
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -6379,7 +3818,119 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 14.7,
+      "tps_mean": 14.86,
+      "tps_std": 0.06,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 242.77,
+      "tps_std": 0.08,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 14.3,
+      "tps_std": 0.15,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 216.34,
+      "tps_std": 0.78,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 120.67,
+      "file_size_gib": 78.02,
+      "name_params_b": 120.67,
+      "quant": "Q4_K_XL",
+      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
+      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 13.69,
       "tps_std": 0.35,
       "error": false,
       "error_type": null,
@@ -6390,1396 +3941,108 @@
       "file_size_gib": 78.02,
       "name_params_b": 120.67,
       "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 185.9,
-      "tps_std": 0.18,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 14.21,
-      "tps_std": 0.38,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 174.22,
-      "tps_std": 0.19,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
       "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003",
-      "model_clean": "NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 13.55,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 120.67,
-      "file_size_gib": 78.02,
-      "name_params_b": 120.67,
-      "quant": "Q4_K_XL",
-      "log": "results/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm-7_2_2-pr21344",
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1498.5,
-      "tps_std": 6.83,
+      "tps_mean": 336.69,
+      "tps_std": 1.88,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1.log",
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm-7_2_2-pr21344",
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 68.29,
-      "tps_std": 0.67,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 213.71,
-      "tps_std": 0.11,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 33.47,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 110.56,
-      "tps_std": 0.1,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 22.59,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1149.95,
-      "tps_std": 8.12,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 67.57,
-      "tps_std": 0.1,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 194.74,
-      "tps_std": 0.07,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 33.36,
-      "tps_std": 0.13,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 102.26,
-      "tps_std": 0.13,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 22.53,
-      "tps_std": 0.07,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1167.77,
-      "tps_std": 7.34,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 65.06,
+      "tps_mean": 20.12,
       "tps_std": 0.02,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1.log",
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 291.69,
-      "tps_std": 0.26,
+      "tps_mean": 276.15,
+      "tps_std": 4.08,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log",
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 32.81,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 159.4,
-      "tps_std": 0.1,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 22.26,
-      "tps_std": 0.17,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1185.86,
-      "tps_std": 2.21,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 68.95,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 203.21,
-      "tps_std": 0.11,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 34.77,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 102.96,
-      "tps_std": 0.05,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 23.26,
-      "tps_std": 0.12,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 882.68,
-      "tps_std": 3.83,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 80.48,
-      "tps_std": 0.06,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 111.69,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 28.03,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 60.42,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 17.44,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1145.07,
-      "tps_std": 8.85,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 82.16,
-      "tps_std": 3.05,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 194.38,
-      "tps_std": 0.33,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 37.09,
-      "tps_std": 0.05,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 63.65,
-      "tps_std": 0.38,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "model_clean": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 24.54,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.53,
-      "file_size_gib": 17.35,
-      "name_params_b": 30.53,
-      "quant": "Q4_K_M",
-      "log": "results/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 418.75,
-      "tps_std": 3.97,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 19.62,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 207.29,
-      "tps_std": 0.79,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 15.34,
-      "tps_std": 2.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 170.79,
-      "tps_std": 0.22,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 14.19,
-      "tps_std": 0.06,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 314.59,
-      "tps_std": 2.13,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 19.96,
-      "tps_std": 0.15,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 235.56,
-      "tps_std": 9.94,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 16.82,
+      "tps_mean": 17.73,
       "tps_std": 0.08,
       "error": false,
       "error_type": null,
@@ -7790,25 +4053,25 @@
       "file_size_gib": 85.6,
       "name_params_b": 122.11,
       "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
       "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
       "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm-7_2_2",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 152.12,
-      "tps_std": 0.13,
+      "tps_mean": 192.28,
+      "tps_std": 2.56,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -7818,25 +4081,25 @@
       "file_size_gib": 85.6,
       "name_params_b": 122.11,
       "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
       "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
       "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm-7_2_2",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 14.21,
-      "tps_std": 0.22,
+      "tps_mean": 15.74,
+      "tps_std": 0.27,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -7846,11 +4109,11 @@
       "file_size_gib": 85.6,
       "name_params_b": 122.11,
       "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -7863,8 +4126,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 316.73,
-      "tps_std": 1.27,
+      "tps_mean": 336.02,
+      "tps_std": 1.65,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -7877,8 +4140,8 @@
       "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -7891,7 +4154,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 18.86,
+      "tps_mean": 19.73,
       "tps_std": 0.02,
       "error": false,
       "error_type": null,
@@ -7905,8 +4168,8 @@
       "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -7919,1169 +4182,1063 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 275.81,
+      "tps_mean": 300.11,
+      "tps_std": 0.1,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 18.47,
+      "tps_std": 0.15,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 202.62,
+      "tps_std": 1.8,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 17.3,
+      "tps_std": 0.22,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 338.52,
+      "tps_std": 3.41,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 20.09,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 272.63,
+      "tps_std": 0.34,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 18.69,
+      "tps_std": 0.18,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 184.62,
+      "tps_std": 0.08,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 17.52,
+      "tps_std": 0.15,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 202.91,
+      "tps_std": 1.91,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 21.42,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 115.36,
+      "tps_std": 0.12,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 19.04,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 81.21,
+      "tps_std": 0.05,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 17.29,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 308.78,
+      "tps_std": 4.29,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 22.3,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 224.82,
+      "tps_std": 0.4,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 20.38,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 165.14,
+      "tps_std": 0.48,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
+      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 18.99,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 122.11,
+      "file_size_gib": 85.6,
+      "name_params_b": 122.11,
+      "quant": "Q5_K_XL",
+      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 367.47,
+      "tps_std": 2.44,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 6.51,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 197.7,
+      "tps_std": 2.94,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 6.14,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 106.3,
+      "tps_std": 1.69,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 5.83,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 364.12,
       "tps_std": 1.62,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
       "env": "rocm6_4_4",
       "env_base": "rocm6_4_4",
       "env_variant": null,
       "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 17.9,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 187.01,
-      "tps_std": 3.13,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 16.91,
-      "tps_std": 0.09,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 318.17,
-      "tps_std": 2.18,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 19.57,
-      "tps_std": 0.14,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 223.61,
-      "tps_std": 0.83,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 18.32,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 138.8,
-      "tps_std": 0.17,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 17.26,
-      "tps_std": 0.15,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 182.98,
-      "tps_std": 1.82,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 21.33,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 102.84,
-      "tps_std": 0.14,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 18.77,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 70.58,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 16.94,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 241.15,
-      "tps_std": 7.43,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 21.75,
-      "tps_std": 0.24,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 202.99,
-      "tps_std": 0.41,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 20.09,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 156.41,
-      "tps_std": 0.22,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003",
-      "model_clean": "Qwen3.5-122B-A10B-UD-Q5_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 18.78,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 122.11,
-      "file_size_gib": 85.6,
-      "name_params_b": 122.11,
-      "quant": "Q5_K_XL",
-      "log": "results/Qwen3.5-122B-A10B-UD-Q5_K_XL-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 513.08,
-      "tps_std": 3.8,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 23.46,
-      "tps_std": 0.06,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 379.44,
-      "tps_std": 1.48,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 21.65,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 296.43,
-      "tps_std": 1.09,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 19.6,
+      "tps_mean": 6.48,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log",
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
       "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 509.14,
-      "tps_std": 3.56,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 172.59,
+      "tps_std": 0.84,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1.log",
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 6.13,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 93.78,
+      "tps_std": 1.21,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 5.81,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 371.42,
+      "tps_std": 4.05,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 23.98,
+      "tps_mean": 6.51,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1.log",
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 391.47,
-      "tps_std": 0.29,
+      "tps_mean": 190.65,
+      "tps_std": 0.72,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log",
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 21.88,
+      "tps_mean": 6.14,
       "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log",
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 264.06,
-      "tps_std": 0.65,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 20.14,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 538.42,
-      "tps_std": 12.9,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 21.91,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 412.11,
-      "tps_std": 0.6,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 21.59,
-      "tps_std": 0.09,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 326.1,
-      "tps_std": 1.55,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 19.05,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": null,
-      "tps_mean": null,
-      "tps_std": null,
-      "error": true,
-      "error_type": "load",
-      "backend": null,
-      "ngl": null,
-      "mmap": null,
-      "params_b": null,
-      "file_size_gib": null,
-      "name_params_b": 35.0,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": null
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": null,
-      "tps_mean": null,
-      "tps_std": null,
-      "error": true,
-      "error_type": "load",
-      "backend": null,
-      "ngl": null,
-      "mmap": null,
-      "params_b": null,
-      "file_size_gib": null,
-      "name_params_b": 35.0,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": null
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
       "env": "rocm7-nightlies",
       "env_base": "rocm7",
       "env_variant": "nightlies",
@@ -9089,27 +5246,27 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 256.49,
-      "tps_std": 1.26,
+      "tps_mean": 105.08,
+      "tps_std": 0.72,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
       "env": "rocm7-nightlies",
       "env_base": "rocm7",
       "env_variant": "nightlies",
@@ -9117,27 +5274,27 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 20.93,
-      "tps_std": 0.06,
+      "tps_mean": 5.82,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
       "env": "vulkan_amdvlk",
       "env_base": "vulkan_amdvlk",
       "env_variant": null,
@@ -9145,1170 +5302,1842 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 122.43,
-      "tps_std": 0.23,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 11.55,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 98.18,
-      "tps_std": 0.09,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 10.94,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 82.44,
-      "tps_std": 0.12,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 10.4,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 316.59,
-      "tps_std": 1.87,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 10.8,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 267.07,
-      "tps_std": 0.91,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 10.37,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 226.78,
-      "tps_std": 1.49,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
-      "model_clean": "Qwen3.6-35B-A3B-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 10.05,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 64.6,
-      "name_params_b": 34.66,
-      "quant": "BF16",
-      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1359.62,
-      "tps_std": 5.74,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 48.34,
-      "tps_std": 0.43,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 670.26,
-      "tps_std": 1.77,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 40.42,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 412.6,
-      "tps_std": 0.33,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 34.9,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1089.21,
-      "tps_std": 5.82,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 49.27,
-      "tps_std": 0.08,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 703.58,
-      "tps_std": 0.51,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 41.08,
-      "tps_std": 0.33,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 460.71,
-      "tps_std": 1.18,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 35.4,
-      "tps_std": 0.24,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1067.33,
-      "tps_std": 6.85,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 48.23,
-      "tps_std": 0.11,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 769.99,
-      "tps_std": 2.79,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 41.91,
-      "tps_std": 0.49,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 473.84,
-      "tps_std": 2.27,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 36.7,
-      "tps_std": 0.97,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1091.62,
-      "tps_std": 10.45,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 49.1,
-      "tps_std": 0.15,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 664.89,
-      "tps_std": 1.48,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 42.73,
-      "tps_std": 0.65,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 432.39,
-      "tps_std": 0.18,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 38.08,
-      "tps_std": 0.59,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 663.94,
-      "tps_std": 2.8,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 57.13,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 292.49,
-      "tps_std": 0.18,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 43.97,
-      "tps_std": 0.06,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 183.42,
-      "tps_std": 0.98,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 36.62,
+      "tps_mean": 80.22,
       "tps_std": 0.05,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 5.88,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 60.84,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 5.42,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 48.97,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 5.03,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 274.97,
+      "tps_std": 0.46,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 6.3,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 123.12,
+      "tps_std": 0.56,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 5.92,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 63.14,
+      "tps_std": 0.36,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-27B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-27B-UD-Q8_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 5.6,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 27.32,
+      "file_size_gib": 33.31,
+      "name_params_b": 27.32,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-27B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 525.94,
+      "tps_std": 13.39,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
       "params_b": 34.66,
-      "file_size_gib": 20.81,
+      "file_size_gib": 64.6,
       "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 26.01,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 417.86,
+      "tps_std": 0.42,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 23.91,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 322.65,
+      "tps_std": 7.4,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 22.2,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 573.71,
+      "tps_std": 4.26,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 25.07,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 426.65,
+      "tps_std": 0.74,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 23.77,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 343.79,
+      "tps_std": 1.31,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 22.22,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 528.97,
+      "tps_std": 5.27,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 25.91,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 409.92,
+      "tps_std": 1.94,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 23.96,
+      "tps_std": 0.44,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 314.48,
+      "tps_std": 1.45,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 22.6,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 122.89,
+      "tps_std": 0.24,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 11.6,
+      "tps_std": 0.03,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 100.51,
+      "tps_std": 0.04,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 10.93,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 85.86,
+      "tps_std": 0.08,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 10.46,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 328.4,
+      "tps_std": 1.64,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 10.68,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 275.8,
+      "tps_std": 0.89,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 10.23,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 241.42,
+      "tps_std": 0.58,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-BF16-00001-of-00002",
+      "model_clean": "Qwen3.6-35B-A3B-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 9.94,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 64.6,
+      "name_params_b": 34.66,
+      "quant": "BF16",
+      "log": "results/Qwen3.6-35B-A3B-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1120.0,
+      "tps_std": 8.54,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 51.53,
+      "tps_std": 0.16,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 746.6,
+      "tps_std": 0.16,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 43.84,
+      "tps_std": 0.36,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 493.36,
+      "tps_std": 0.38,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 38.39,
+      "tps_std": 0.27,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1116.91,
+      "tps_std": 7.21,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 50.32,
+      "tps_std": 0.12,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 807.33,
+      "tps_std": 0.34,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 43.47,
+      "tps_std": 0.44,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 547.71,
+      "tps_std": 1.13,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 38.47,
+      "tps_std": 0.41,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1085.14,
+      "tps_std": 11.75,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 51.47,
+      "tps_std": 0.18,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 716.94,
+      "tps_std": 1.11,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 44.45,
+      "tps_std": 0.7,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 468.89,
+      "tps_std": 0.65,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 39.3,
+      "tps_std": 0.58,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 696.54,
+      "tps_std": 2.95,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 59.57,
+      "tps_std": 0.07,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 319.31,
+      "tps_std": 0.43,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 45.3,
+      "tps_std": 0.2,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 207.54,
+      "tps_std": 0.69,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
       "quant": "Q4_K_XL",
       "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 36.74,
+      "tps_std": 0.04,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -10321,22 +7150,22 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1045.16,
-      "tps_std": 5.79,
+      "tps_mean": 1113.64,
+      "tps_std": 7.57,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 20.81,
-      "name_params_b": 34.66,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
       "quant": "Q4_K_XL",
       "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
@@ -10349,27 +7178,867 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 58.82,
-      "tps_std": 0.08,
+      "tps_mean": 60.43,
+      "tps_std": 0.06,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 715.52,
+      "tps_std": 1.18,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 49.21,
+      "tps_std": 0.38,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 524.7,
+      "tps_std": 1.2,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 43.23,
+      "tps_std": 0.23,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 35.51,
+      "file_size_gib": 21.27,
+      "name_params_b": 35.51,
+      "quant": "Q4_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1093.92,
+      "tps_std": 7.09,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 46.18,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 728.92,
+      "tps_std": 0.14,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 40.0,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 486.17,
+      "tps_std": 0.85,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 35.45,
+      "tps_std": 0.05,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1095.53,
+      "tps_std": 11.56,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 45.13,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 787.54,
+      "tps_std": 0.45,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 39.88,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 537.23,
+      "tps_std": 1.72,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 35.72,
+      "tps_std": 0.04,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1088.56,
+      "tps_std": 10.42,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 46.21,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 704.46,
+      "tps_std": 1.78,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 40.54,
+      "tps_std": 0.21,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 462.61,
+      "tps_std": 1.15,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 36.31,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 640.78,
+      "tps_std": 2.08,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
       "params_b": 34.66,
-      "file_size_gib": 20.81,
+      "file_size_gib": 35.8,
       "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1.log",
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 46.53,
+      "tps_std": 0.07,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 307.07,
+      "tps_std": 1.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 37.93,
+      "tps_std": 0.07,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 202.18,
+      "tps_std": 0.79,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 32.35,
+      "tps_std": 0.04,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1045.0,
+      "tps_std": 7.3,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 46.33,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 34.66,
+      "file_size_gib": 35.8,
+      "name_params_b": 34.66,
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "1a68ec937",
+        "number": "9193"
+      }
+    },
+    {
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
       "env": "vulkan_radv",
       "env_base": "vulkan_radv",
       "env_variant": null,
@@ -10378,26 +8047,26 @@
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
       "tps_mean": 682.8,
-      "tps_std": 1.45,
+      "tps_std": 0.96,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
       "params_b": 34.66,
-      "file_size_gib": 20.81,
+      "file_size_gib": 35.8,
       "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log",
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
       "env": "vulkan_radv",
       "env_base": "vulkan_radv",
       "env_variant": null,
@@ -10405,27 +8074,27 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 48.77,
-      "tps_std": 0.07,
+      "tps_mean": 40.08,
+      "tps_std": 0.17,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
       "params_b": 34.66,
-      "file_size_gib": 20.81,
+      "file_size_gib": 35.8,
       "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log",
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
       "env": "vulkan_radv",
       "env_base": "vulkan_radv",
       "env_variant": null,
@@ -10433,27 +8102,27 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 507.32,
-      "tps_std": 1.33,
+      "tps_mean": 504.87,
+      "tps_std": 1.05,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
       "params_b": 34.66,
-      "file_size_gib": 20.81,
+      "file_size_gib": 35.8,
       "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log",
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q4_K_XL",
+      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
       "env": "vulkan_radv",
       "env_base": "vulkan_radv",
       "env_variant": null,
@@ -10461,7 +8130,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 42.75,
+      "tps_mean": 35.8,
       "tps_std": 0.05,
       "error": false,
       "error_type": null,
@@ -10469,355 +8138,1027 @@
       "ngl": 99,
       "mmap": 0,
       "params_b": 34.66,
-      "file_size_gib": 20.81,
+      "file_size_gib": 35.8,
       "name_params_b": 34.66,
-      "quant": "Q4_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log",
+      "quant": "Q8_K_XL",
+      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "1a68ec937",
+        "number": "9193"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1222.77,
-      "tps_std": 2.88,
+      "tps_mean": 909.04,
+      "tps_std": 7.54,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log",
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 43.72,
-      "tps_std": 0.08,
+      "tps_mean": 24.64,
+      "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log",
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 653.49,
-      "tps_std": 0.18,
+      "tps_mean": 667.44,
+      "tps_std": 3.68,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 37.21,
-      "tps_std": 0.02,
+      "tps_mean": 21.59,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 414.31,
+      "tps_mean": 446.88,
+      "tps_std": 2.91,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 20.79,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 893.46,
+      "tps_std": 6.97,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 23.37,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 691.94,
+      "tps_std": 9.26,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 21.39,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 470.28,
+      "tps_std": 3.56,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 20.66,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 870.72,
+      "tps_std": 80.56,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 24.65,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 638.41,
+      "tps_std": 6.72,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 21.58,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 421.88,
+      "tps_std": 0.33,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 20.88,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 123.59,
+      "tps_std": 0.22,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 18.85,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 60.07,
+      "tps_std": 0.05,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 16.44,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 39.84,
+      "tps_std": 0.03,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 15.1,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 587.64,
+      "tps_std": 6.59,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 18.6,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 428.06,
+      "tps_std": 1.23,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 16.99,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 331.82,
       "tps_std": 0.61,
       "error": false,
       "error_type": null,
-      "backend": "ROCm",
+      "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-26B-A4B-it-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 32.3,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1043.12,
-      "tps_std": 46.1,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 44.53,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 544.04,
-      "tps_std": 1.79,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 37.7,
-      "tps_std": 0.21,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 414.74,
-      "tps_std": 1.74,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 32.74,
+      "tps_mean": 16.12,
       "tps_std": 0.02,
       "error": false,
       "error_type": null,
-      "backend": "ROCm",
+      "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log",
+      "params_b": 25.23,
+      "file_size_gib": 47.02,
+      "name_params_b": 25.23,
+      "quant": "BF16",
+      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1329.04,
+      "tps_std": 9.98,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 48.46,
+      "tps_std": 0.11,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 725.46,
+      "tps_std": 9.61,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 37.84,
+      "tps_std": 0.15,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 471.3,
+      "tps_std": 1.61,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 35.61,
+      "tps_std": 0.14,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "env": "rocm6_4_4",
       "env_base": "rocm6_4_4",
       "env_variant": null,
@@ -10825,27 +9166,27 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1049.22,
-      "tps_std": 7.25,
+      "tps_mean": 1334.91,
+      "tps_std": 9.9,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log",
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "env": "rocm6_4_4",
       "env_base": "rocm6_4_4",
       "env_variant": null,
@@ -10853,27 +9194,27 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 43.1,
-      "tps_std": 0.01,
+      "tps_mean": 47.74,
+      "tps_std": 0.07,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1.log",
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "env": "rocm6_4_4",
       "env_base": "rocm6_4_4",
       "env_variant": null,
@@ -10881,27 +9222,27 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 758.76,
-      "tps_std": 0.99,
+      "tps_mean": 761.5,
+      "tps_std": 12.52,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "env": "rocm6_4_4",
       "env_base": "rocm6_4_4",
       "env_variant": null,
@@ -10909,27 +9250,27 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 38.7,
-      "tps_std": 0.01,
+      "tps_mean": 37.29,
+      "tps_std": 0.26,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "env": "rocm6_4_4",
       "env_base": "rocm6_4_4",
       "env_variant": null,
@@ -10937,27 +9278,27 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 515.64,
-      "tps_std": 0.13,
+      "tps_mean": 499.49,
+      "tps_std": 1.7,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "env": "rocm6_4_4",
       "env_base": "rocm6_4_4",
       "env_variant": null,
@@ -10965,335 +9306,335 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 34.53,
+      "tps_mean": 35.18,
+      "tps_std": 0.24,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1194.14,
+      "tps_std": 6.99,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 48.47,
+      "tps_std": 0.12,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 674.07,
+      "tps_std": 6.75,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 37.68,
+      "tps_std": 0.38,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 437.2,
+      "tps_std": 0.38,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 35.52,
+      "tps_std": 0.34,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 765.18,
+      "tps_std": 2.13,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 52.04,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 101.11,
+      "tps_std": 0.1,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 36.34,
+      "tps_std": 0.06,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 54.46,
       "tps_std": 0.03,
       "error": false,
       "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1036.89,
-      "tps_std": 9.24,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 44.2,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 510.38,
-      "tps_std": 2.64,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 38.18,
-      "tps_std": 1.71,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 375.11,
-      "tps_std": 2.14,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 35.02,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 576.25,
-      "tps_std": 2.16,
-      "error": false,
-      "error_type": null,
       "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 45.25,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 273.77,
-      "tps_std": 0.3,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 36.71,
-      "tps_std": 0.08,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 175.02,
-      "tps_std": 1.28,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "env": "vulkan_amdvlk",
       "env_base": "vulkan_amdvlk",
       "env_variant": null,
@@ -11301,27 +9642,27 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 31.29,
+      "tps_mean": 30.34,
       "tps_std": 0.05,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "env": "vulkan_radv",
       "env_base": "vulkan_radv",
       "env_variant": null,
@@ -11329,27 +9670,27 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 893.79,
-      "tps_std": 4.74,
+      "tps_mean": 1324.66,
+      "tps_std": 8.47,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log",
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "env": "vulkan_radv",
       "env_base": "vulkan_radv",
       "env_variant": null,
@@ -11357,27 +9698,27 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 46.05,
-      "tps_std": 0.07,
+      "tps_mean": 54.67,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1.log",
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "env": "vulkan_radv",
       "env_base": "vulkan_radv",
       "env_variant": null,
@@ -11385,27 +9726,27 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 610.84,
-      "tps_std": 1.99,
+      "tps_mean": 682.46,
+      "tps_std": 2.1,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "env": "vulkan_radv",
       "env_base": "vulkan_radv",
       "env_variant": null,
@@ -11413,2163 +9754,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 39.67,
-      "tps_std": 0.15,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 468.83,
-      "tps_std": 1.17,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "model_clean": "Qwen3.6-35B-A3B-UD-Q8_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 35.46,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 34.66,
-      "file_size_gib": 35.8,
-      "name_params_b": 34.66,
-      "quant": "Q8_K_XL",
-      "log": "results/Qwen3.6-35B-A3B-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 836.74,
-      "tps_std": 6.15,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 22.95,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 619.85,
-      "tps_std": 6.1,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 20.38,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 408.04,
-      "tps_std": 3.58,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 19.52,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 823.65,
-      "tps_std": 83.54,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 22.67,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 628.14,
-      "tps_std": 5.88,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 20.15,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 403.29,
-      "tps_std": 1.83,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 19.35,
-      "tps_std": 0.19,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 796.06,
-      "tps_std": 126.76,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 22.68,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 628.2,
-      "tps_std": 5.71,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 20.12,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 426.61,
-      "tps_std": 3.98,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 19.55,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 839.13,
-      "tps_std": 7.61,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 20.75,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 646.67,
-      "tps_std": 11.22,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 19.86,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 441.14,
-      "tps_std": 2.12,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 17.76,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 844.42,
-      "tps_std": 19.66,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 22.69,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 546.66,
-      "tps_std": 5.96,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 20.11,
-      "tps_std": 0.07,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 345.7,
-      "tps_std": 0.13,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 19.16,
-      "tps_std": 0.47,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 117.96,
-      "tps_std": 0.28,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 16.18,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 58.94,
-      "tps_std": 0.07,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 15.26,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 39.57,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 14.1,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 376.52,
-      "tps_std": 7.37,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 14.74,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 351.87,
-      "tps_std": 1.49,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 13.71,
-      "tps_std": 0.06,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 281.32,
-      "tps_std": 0.7,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-26B-A4B-it-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 13.15,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 47.02,
-      "name_params_b": 25.23,
-      "quant": "BF16",
-      "log": "results/gemma-4-26B-A4B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1608.01,
-      "tps_std": 5.26,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 47.13,
-      "tps_std": 0.29,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 689.3,
-      "tps_std": 12.21,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 37.14,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 446.29,
-      "tps_std": 0.87,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 35.16,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1299.9,
-      "tps_std": 11.12,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 46.59,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 693.48,
-      "tps_std": 4.74,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 36.76,
-      "tps_std": 0.16,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 453.92,
-      "tps_std": 5.18,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 34.85,
-      "tps_std": 0.16,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1301.66,
-      "tps_std": 10.69,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 46.42,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 693.56,
-      "tps_std": 10.92,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 36.83,
-      "tps_std": 0.15,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 450.75,
-      "tps_std": 0.77,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 34.84,
-      "tps_std": 0.16,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1288.38,
-      "tps_std": 10.85,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 45.58,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 721.84,
-      "tps_std": 10.97,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 36.17,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 469.64,
-      "tps_std": 3.17,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 34.31,
-      "tps_std": 0.25,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1219.74,
-      "tps_std": 15.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 46.57,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 600.42,
-      "tps_std": 7.55,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 36.88,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 374.22,
-      "tps_std": 0.13,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 34.77,
-      "tps_std": 0.32,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 720.78,
-      "tps_std": 2.94,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 50.21,
-      "tps_std": 0.06,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 100.79,
+      "tps_mean": 42.4,
       "tps_std": 0.03,
       "error": false,
       "error_type": null,
@@ -13580,25 +9765,25 @@
       "file_size_gib": 15.9,
       "name_params_b": 25.23,
       "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
       "env_variant": null,
       "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 35.41,
-      "tps_std": 0.01,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 469.95,
+      "tps_std": 0.4,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -13608,24 +9793,612 @@
       "file_size_gib": 15.9,
       "name_params_b": 25.23,
       "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
       "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 37.46,
+      "tps_std": 0.07,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 15.9,
+      "name_params_b": 25.23,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1354.0,
+      "tps_std": 8.88,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 42.35,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 725.86,
+      "tps_std": 11.87,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 34.01,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 471.81,
+      "tps_std": 3.32,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 32.14,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1356.47,
+      "tps_std": 9.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 41.28,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 769.66,
+      "tps_std": 14.85,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 33.66,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
       "env_variant": null,
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 54.9,
+      "tps_mean": 504.88,
+      "tps_std": 0.76,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 32.09,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1342.08,
+      "tps_std": 8.02,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 42.32,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 692.34,
+      "tps_std": 8.31,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 34.0,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 447.95,
+      "tps_std": 0.85,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 32.16,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 669.11,
+      "tps_std": 1.51,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 43.93,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 99.13,
       "tps_std": 0.12,
       "error": false,
       "error_type": null,
@@ -13633,19 +10406,75 @@
       "ngl": 99,
       "mmap": 0,
       "params_b": 25.23,
-      "file_size_gib": 15.9,
+      "file_size_gib": 25.94,
       "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 32.25,
+      "tps_std": 0.02,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 53.86,
+      "tps_std": 0.04,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
       "env": "vulkan_amdvlk",
       "env_base": "vulkan_amdvlk",
       "env_variant": null,
@@ -13653,7 +10482,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 29.52,
+      "tps_mean": 27.52,
       "tps_std": 0.03,
       "error": false,
       "error_type": null,
@@ -13661,19 +10490,19 @@
       "ngl": 99,
       "mmap": 0,
       "params_b": 25.23,
-      "file_size_gib": 15.9,
+      "file_size_gib": 25.94,
       "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
       "env": "vulkan_radv",
       "env_base": "vulkan_radv",
       "env_variant": null,
@@ -13681,27 +10510,27 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1213.78,
-      "tps_std": 7.36,
+      "tps_mean": 1291.45,
+      "tps_std": 8.95,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
       "params_b": 25.23,
-      "file_size_gib": 15.9,
+      "file_size_gib": 25.94,
       "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1.log",
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
       "env": "vulkan_radv",
       "env_base": "vulkan_radv",
       "env_variant": null,
@@ -13709,1758 +10538,274 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 53.04,
-      "tps_std": 0.1,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 638.76,
-      "tps_std": 1.76,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 40.93,
-      "tps_std": 0.09,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 445.0,
-      "tps_std": 0.08,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q4_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 36.64,
-      "tps_std": 0.06,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 15.9,
-      "name_params_b": 25.23,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1524.44,
-      "tps_std": 7.11,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 41.68,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 689.36,
-      "tps_std": 8.08,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 33.67,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 436.86,
-      "tps_std": 3.98,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 31.55,
-      "tps_std": 0.72,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1272.69,
-      "tps_std": 74.5,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 41.31,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 696.25,
-      "tps_std": 7.26,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 32.95,
-      "tps_std": 0.77,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 452.1,
-      "tps_std": 0.82,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 29.87,
-      "tps_std": 3.28,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1301.93,
-      "tps_std": 18.07,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 41.35,
+      "tps_mean": 45.48,
       "tps_std": 0.03,
       "error": false,
       "error_type": null,
-      "backend": "ROCm",
+      "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
       "params_b": 25.23,
       "file_size_gib": 25.94,
       "name_params_b": 25.23,
       "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
       "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 695.0,
-      "tps_std": 7.26,
+      "tps_mean": 672.93,
+      "tps_std": 0.29,
       "error": false,
       "error_type": null,
-      "backend": "ROCm",
+      "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
       "params_b": 25.23,
       "file_size_gib": 25.94,
       "name_params_b": 25.23,
       "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
       "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 33.41,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 447.63,
-      "tps_std": 3.37,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 31.76,
+      "tps_mean": 36.81,
       "tps_std": 0.04,
       "error": false,
       "error_type": null,
-      "backend": "ROCm",
+      "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
       "params_b": 25.23,
       "file_size_gib": 25.94,
       "name_params_b": 25.23,
       "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
       "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
       "env_variant": null,
       "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 465.71,
+      "tps_std": 0.32,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 32.9,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 25.23,
+      "file_size_gib": 25.94,
+      "name_params_b": 25.23,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-31B-it-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1303.32,
-      "tps_std": 8.77,
+      "tps_mean": 428.07,
+      "tps_std": 3.61,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log",
+      "params_b": 30.7,
+      "file_size_gib": 57.18,
+      "name_params_b": 30.7,
+      "quant": "BF16",
+      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
+      "model": "gemma-4-31B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-31B-it-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 38.81,
-      "tps_std": 0.01,
+      "tps_mean": 3.55,
+      "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log",
+      "params_b": 30.7,
+      "file_size_gib": 57.18,
+      "name_params_b": 30.7,
+      "quant": "BF16",
+      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
+      "model": "gemma-4-31B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-31B-it-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 727.92,
-      "tps_std": 8.62,
+      "tps_mean": 167.2,
+      "tps_std": 1.91,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "params_b": 30.7,
+      "file_size_gib": 57.18,
+      "name_params_b": 30.7,
+      "quant": "BF16",
+      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
+      "model": "gemma-4-31B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-31B-it-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 32.52,
+      "tps_mean": 3.23,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 57.18,
+      "name_params_b": 30.7,
+      "quant": "BF16",
+      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-31B-it-BF16",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 108.99,
       "tps_std": 0.35,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 451.48,
-      "tps_std": 1.29,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 30.76,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1294.4,
-      "tps_std": 31.54,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 41.09,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 596.78,
-      "tps_std": 7.98,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 33.35,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 363.83,
-      "tps_std": 2.85,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 31.68,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 589.2,
-      "tps_std": 1.79,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 43.18,
-      "tps_std": 0.07,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 97.7,
-      "tps_std": 0.1,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 31.82,
-      "tps_std": 0.05,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 54.04,
-      "tps_std": 0.08,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 26.65,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 944.96,
-      "tps_std": 19.24,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 43.58,
-      "tps_std": 1.06,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 579.88,
-      "tps_std": 0.09,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 35.52,
-      "tps_std": 0.07,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 415.94,
-      "tps_std": 0.55,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-26B-A4B-it-UD-Q8_K_XL",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 31.99,
-      "tps_std": 0.06,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 25.23,
-      "file_size_gib": 25.94,
-      "name_params_b": 25.23,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-26B-A4B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 380.58,
-      "tps_std": 1.49,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
       "params_b": 30.7,
       "file_size_gib": 57.18,
       "name_params_b": 30.7,
       "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log",
+      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 3.49,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 150.63,
-      "tps_std": 1.53,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 3.17,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 96.84,
-      "tps_std": 0.57,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 3.07,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 377.93,
-      "tps_std": 2.48,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 3.44,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 153.11,
-      "tps_std": 1.83,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 3.18,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 98.23,
-      "tps_std": 0.26,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 3.08,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 376.47,
-      "tps_std": 1.77,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 3.44,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 152.54,
-      "tps_std": 2.43,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 3.18,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15472,9 +10817,9 @@
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 98.53,
-      "tps_std": 0.11,
+      "test": "tg32 @ d65536",
+      "tps_mean": 3.13,
+      "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -15487,36 +10832,8 @@
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 3.08,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15529,63 +10846,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 360.84,
-      "tps_std": 1.6,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 3.3,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 154.86,
+      "tps_mean": 410.15,
       "tps_std": 2.16,
       "error": false,
       "error_type": null,
@@ -15596,11 +10857,67 @@
       "file_size_gib": 57.18,
       "name_params_b": 30.7,
       "quant": "BF16",
+      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-31B-it-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 3.44,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 57.18,
+      "name_params_b": 30.7,
+      "quant": "BF16",
+      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-31B-it-BF16",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 172.49,
+      "tps_std": 2.42,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 57.18,
+      "name_params_b": 30.7,
+      "quant": "BF16",
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15613,8 +10930,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 3.14,
-      "tps_std": 0.01,
+      "tps_mean": 3.21,
+      "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -15627,8 +10944,8 @@
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15641,8 +10958,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 101.08,
-      "tps_std": 0.13,
+      "tps_mean": 114.16,
+      "tps_std": 0.63,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -15655,8 +10972,8 @@
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15669,8 +10986,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 2.94,
-      "tps_std": 0.02,
+      "tps_mean": 3.12,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -15683,8 +11000,8 @@
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm6_4_4__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15697,8 +11014,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 373.84,
-      "tps_std": 0.93,
+      "tps_mean": 423.05,
+      "tps_std": 6.41,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -15711,8 +11028,8 @@
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15725,7 +11042,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 3.41,
+      "tps_mean": 3.53,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -15739,8 +11056,8 @@
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15753,8 +11070,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 139.53,
-      "tps_std": 1.98,
+      "tps_mean": 158.77,
+      "tps_std": 2.07,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -15767,8 +11084,8 @@
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15781,8 +11098,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 3.16,
-      "tps_std": 0.05,
+      "tps_mean": 3.25,
+      "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -15795,8 +11112,8 @@
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15809,8 +11126,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 83.27,
-      "tps_std": 0.13,
+      "tps_mean": 102.83,
+      "tps_std": 0.08,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -15823,8 +11140,8 @@
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15837,7 +11154,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 3.04,
+      "tps_mean": 3.14,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -15851,8 +11168,8 @@
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15940,8 +11257,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 79.53,
-      "tps_std": 0.26,
+      "tps_mean": 135.94,
+      "tps_std": 0.46,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -15954,8 +11271,8 @@
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15982,8 +11299,8 @@
       "log": "results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -15996,850 +11313,514 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 60.29,
-      "tps_std": 0.37,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 3.24,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 45.96,
-      "tps_std": 1.09,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-BF16-00001-of-00002",
-      "model_clean": "gemma-4-31B-it-BF16",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 3.06,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 57.18,
-      "name_params_b": 30.7,
-      "quant": "BF16",
-      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 293.27,
-      "tps_std": 0.37,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 10.5,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 136.84,
-      "tps_std": 1.27,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 8.11,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 91.86,
-      "tps_std": 0.33,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 7.51,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 310.54,
-      "tps_std": 0.73,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 10.52,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 140.71,
-      "tps_std": 1.29,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 8.11,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 93.57,
-      "tps_std": 0.2,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 7.52,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 309.16,
-      "tps_std": 0.31,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 10.49,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 139.29,
-      "tps_std": 1.61,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 8.1,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 92.8,
-      "tps_std": 0.18,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 7.51,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 309.22,
-      "tps_std": 0.13,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 10.02,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 143.91,
-      "tps_std": 1.64,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 7.81,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 96.98,
-      "tps_std": 0.37,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 7.26,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 315.72,
-      "tps_std": 0.3,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 10.51,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 17.46,
-      "name_params_b": 30.7,
-      "quant": "Q4_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q4_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 130.8,
+      "tps_mean": 92.4,
       "tps_std": 1.45,
       "error": false,
       "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 57.18,
+      "name_params_b": 30.7,
+      "quant": "BF16",
+      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-31B-it-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 3.25,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 57.18,
+      "name_params_b": 30.7,
+      "quant": "BF16",
+      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-31B-it-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 67.88,
+      "tps_std": 1.23,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 57.18,
+      "name_params_b": 30.7,
+      "quant": "BF16",
+      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-BF16-00001-of-00002",
+      "model_clean": "gemma-4-31B-it-BF16",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 3.08,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 57.18,
+      "name_params_b": 30.7,
+      "quant": "BF16",
+      "log": "results/gemma-4-31B-it-BF16-00001-of-00002__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 322.69,
+      "tps_std": 0.81,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 10.55,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 150.43,
+      "tps_std": 1.59,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 8.13,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 101.59,
+      "tps_std": 0.41,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 7.54,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 317.49,
+      "tps_std": 0.94,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 10.2,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 154.73,
+      "tps_std": 1.21,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 7.95,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 106.22,
+      "tps_std": 0.22,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 7.38,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 313.83,
+      "tps_std": 2.55,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 10.55,
+      "tps_std": 0.01,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 17.46,
+      "name_params_b": 30.7,
+      "quant": "Q4_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q4_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q4_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 142.53,
+      "tps_std": 1.4,
+      "error": false,
+      "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
@@ -16850,8 +11831,8 @@
       "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -16864,8 +11845,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 8.2,
-      "tps_std": 0.0,
+      "tps_mean": 8.21,
+      "tps_std": 0.02,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -16878,8 +11859,8 @@
       "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -16892,8 +11873,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 84.96,
-      "tps_std": 0.07,
+      "tps_mean": 95.32,
+      "tps_std": 0.1,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -16906,8 +11887,8 @@
       "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -16920,7 +11901,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 7.56,
+      "tps_mean": 7.58,
       "tps_std": 0.02,
       "error": false,
       "error_type": null,
@@ -16934,8 +11915,8 @@
       "log": "results/gemma-4-31B-it-UD-Q4_K_XL__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -16948,7 +11929,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 69.12,
+      "tps_mean": 70.48,
       "tps_std": 0.02,
       "error": false,
       "error_type": null,
@@ -16962,8 +11943,8 @@
       "log": "results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -16976,7 +11957,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 10.71,
+      "tps_mean": 11.32,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -16990,8 +11971,8 @@
       "log": "results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -17054,8 +12035,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 244.46,
-      "tps_std": 0.27,
+      "tps_mean": 301.53,
+      "tps_std": 0.16,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -17068,8 +12049,8 @@
       "log": "results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -17082,8 +12063,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 11.02,
-      "tps_std": 0.01,
+      "tps_mean": 11.45,
+      "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -17096,8 +12077,8 @@
       "log": "results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -17110,7 +12091,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 124.32,
+      "tps_mean": 146.08,
       "tps_std": 1.04,
       "error": false,
       "error_type": null,
@@ -17124,8 +12105,8 @@
       "log": "results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -17138,7 +12119,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 8.78,
+      "tps_mean": 9.09,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -17152,8 +12133,8 @@
       "log": "results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -17166,8 +12147,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 76.03,
-      "tps_std": 1.52,
+      "tps_mean": 86.61,
+      "tps_std": 2.84,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -17180,8 +12161,8 @@
       "log": "results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -17194,7 +12175,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 7.78,
+      "tps_mean": 7.88,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -17208,22 +12189,22 @@
       "log": "results/gemma-4-31B-it-UD-Q4_K_XL__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-31B-it-UD-Q8_K_XL",
       "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 282.42,
-      "tps_std": 0.4,
+      "tps_mean": 326.41,
+      "tps_std": 1.5,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -17233,24 +12214,24 @@
       "file_size_gib": 32.6,
       "name_params_b": 30.7,
       "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-31B-it-UD-Q8_K_XL",
       "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 6.16,
+      "tps_mean": 6.17,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -17261,25 +12242,25 @@
       "file_size_gib": 32.6,
       "name_params_b": 30.7,
       "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1.log",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-31B-it-UD-Q8_K_XL",
       "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 132.6,
-      "tps_std": 1.37,
+      "tps_mean": 150.02,
+      "tps_std": 1.17,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -17289,19 +12270,19 @@
       "file_size_gib": 32.6,
       "name_params_b": 30.7,
       "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-31B-it-UD-Q8_K_XL",
       "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
@@ -17317,25 +12298,25 @@
       "file_size_gib": 32.6,
       "name_params_b": 30.7,
       "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx32768.log",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-31B-it-UD-Q8_K_XL",
       "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 90.15,
-      "tps_std": 0.22,
+      "tps_mean": 100.93,
+      "tps_std": 0.14,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -17345,24 +12326,24 @@
       "file_size_gib": 32.6,
       "name_params_b": 30.7,
       "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-31B-it-UD-Q8_K_XL",
       "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 5.0,
+      "tps_mean": 5.01,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -17373,25 +12354,25 @@
       "file_size_gib": 32.6,
       "name_params_b": 30.7,
       "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2-pr21344__fa1__longctx65536.log",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-31B-it-UD-Q8_K_XL",
       "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 308.17,
-      "tps_std": 0.62,
+      "tps_mean": 325.63,
+      "tps_std": 1.72,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -17401,501 +12382,193 @@
       "file_size_gib": 32.6,
       "name_params_b": 30.7,
       "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1.log",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gemma-4-31B-it-UD-Q8_K_XL",
       "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 6.14,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 32.6,
+      "name_params_b": 30.7,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 155.9,
+      "tps_std": 2.5,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 32.6,
+      "name_params_b": 30.7,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 5.27,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 32.6,
+      "name_params_b": 30.7,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 106.02,
+      "tps_std": 0.86,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 32.6,
+      "name_params_b": 30.7,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 5.01,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 32.6,
+      "name_params_b": 30.7,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 333.66,
+      "tps_std": 1.84,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 30.7,
+      "file_size_gib": 32.6,
+      "name_params_b": 30.7,
+      "quant": "Q8_K_XL",
+      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gemma-4-31B-it-UD-Q8_K_XL",
+      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
       "tps_mean": 6.16,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 138.85,
-      "tps_std": 1.58,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 5.26,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 93.92,
-      "tps_std": 0.29,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 5.0,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 309.61,
-      "tps_std": 0.81,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 6.15,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 139.33,
-      "tps_std": 1.73,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 5.25,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 93.3,
-      "tps_std": 0.54,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 5.0,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 303.4,
-      "tps_std": 0.23,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 6.07,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 141.37,
-      "tps_std": 1.98,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 5.18,
-      "tps_std": 0.08,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 93.78,
-      "tps_std": 0.36,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 4.95,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 319.75,
-      "tps_std": 0.26,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -17908,36 +12581,8 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gemma-4-31B-it-UD-Q8_K_XL",
-      "model_clean": "gemma-4-31B-it-UD-Q8_K_XL",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 6.12,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 30.7,
-      "file_size_gib": 32.6,
-      "name_params_b": 30.7,
-      "quant": "Q8_K_XL",
-      "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -17950,8 +12595,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 128.0,
-      "tps_std": 1.3,
+      "tps_mean": 143.8,
+      "tps_std": 1.56,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -17964,8 +12609,8 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -17979,7 +12624,7 @@
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
       "tps_mean": 5.3,
-      "tps_std": 0.0,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -17992,8 +12637,8 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18006,8 +12651,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 81.58,
-      "tps_std": 0.14,
+      "tps_mean": 96.12,
+      "tps_std": 0.1,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -18020,8 +12665,8 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18034,8 +12679,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 4.99,
-      "tps_std": 0.05,
+      "tps_mean": 5.03,
+      "tps_std": 0.01,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -18048,8 +12693,8 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18062,8 +12707,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 72.26,
-      "tps_std": 0.02,
+      "tps_mean": 73.59,
+      "tps_std": 0.03,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -18076,8 +12721,8 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18090,7 +12735,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 6.28,
+      "tps_mean": 6.21,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -18104,8 +12749,8 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18168,8 +12813,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 209.2,
-      "tps_std": 6.85,
+      "tps_mean": 272.1,
+      "tps_std": 0.28,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -18182,8 +12827,8 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18196,7 +12841,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 6.28,
+      "tps_mean": 6.3,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -18210,8 +12855,8 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18224,8 +12869,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 112.72,
-      "tps_std": 1.26,
+      "tps_mean": 141.06,
+      "tps_std": 0.76,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -18238,8 +12883,8 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18252,7 +12897,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 5.49,
+      "tps_mean": 5.52,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -18266,8 +12911,8 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18280,8 +12925,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 71.68,
-      "tps_std": 0.77,
+      "tps_mean": 88.45,
+      "tps_std": 3.85,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -18294,8 +12939,8 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18308,7 +12953,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 5.02,
+      "tps_mean": 5.05,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -18322,22 +12967,22 @@
       "log": "results/gemma-4-31B-it-UD-Q8_K_XL__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gpt-oss-120b-mxfp4-00001-of-00003",
       "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 945.82,
-      "tps_std": 60.17,
+      "tps_mean": 625.58,
+      "tps_std": 109.36,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -18347,192 +12992,24 @@
       "file_size_gib": 59.02,
       "name_params_b": 116.83,
       "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1.log",
+      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gpt-oss-120b-mxfp4-00001-of-00003",
       "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_2-pr21344",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 51.25,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 311.38,
-      "tps_std": 1.36,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 31.94,
-      "tps_std": 7.16,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 175.99,
-      "tps_std": 0.26,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 27.44,
-      "tps_std": 0.31,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 634.64,
-      "tps_std": 3.8,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 50.7,
+      "tps_mean": 51.81,
       "tps_std": 0.04,
       "error": false,
       "error_type": null,
@@ -18543,179 +13020,11 @@
       "file_size_gib": 59.02,
       "name_params_b": 116.83,
       "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 304.96,
-      "tps_std": 1.71,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 35.84,
-      "tps_std": 0.12,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 175.32,
-      "tps_std": 0.5,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 24.52,
-      "tps_std": 5.66,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 635.33,
-      "tps_std": 4.03,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 50.99,
-      "tps_std": 0.05,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18728,8 +13037,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 302.31,
-      "tps_std": 0.46,
+      "tps_mean": 589.25,
+      "tps_std": 0.34,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -18742,8 +13051,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18756,8 +13065,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 35.94,
-      "tps_std": 0.12,
+      "tps_mean": 36.32,
+      "tps_std": 0.13,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -18770,8 +13079,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18784,8 +13093,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 175.25,
-      "tps_std": 0.49,
+      "tps_mean": 371.47,
+      "tps_std": 2.43,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -18798,8 +13107,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18812,8 +13121,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 27.78,
-      "tps_std": 0.06,
+      "tps_mean": 26.11,
+      "tps_std": 3.5,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -18826,8 +13135,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18840,8 +13149,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 648.77,
-      "tps_std": 4.37,
+      "tps_mean": 675.76,
+      "tps_std": 4.24,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -18854,8 +13163,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18868,8 +13177,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 49.76,
-      "tps_std": 0.02,
+      "tps_mean": 51.02,
+      "tps_std": 0.07,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -18882,8 +13191,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18896,8 +13205,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 439.58,
-      "tps_std": 0.57,
+      "tps_mean": 603.27,
+      "tps_std": 1.54,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -18910,8 +13219,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -18924,147 +13233,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 39.77,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 273.87,
-      "tps_std": 0.3,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 32.17,
-      "tps_std": 0.38,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 641.05,
-      "tps_std": 2.79,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 50.73,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 305.81,
+      "tps_mean": 40.02,
       "tps_std": 0.24,
       "error": false,
       "error_type": null,
@@ -19075,11 +13244,123 @@
       "file_size_gib": 59.02,
       "name_params_b": 116.83,
       "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log",
+      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
+      "model_clean": "gpt-oss-120b-mxfp4",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 379.11,
+      "tps_std": 1.07,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 116.83,
+      "file_size_gib": 59.02,
+      "name_params_b": 116.83,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
+      "model_clean": "gpt-oss-120b-mxfp4",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 32.99,
+      "tps_std": 0.14,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 116.83,
+      "file_size_gib": 59.02,
+      "name_params_b": 116.83,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm6_4_4__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
+      "model_clean": "gpt-oss-120b-mxfp4",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 684.87,
+      "tps_std": 3.91,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 116.83,
+      "file_size_gib": 59.02,
+      "name_params_b": 116.83,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
+      "model_clean": "gpt-oss-120b-mxfp4",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 51.48,
+      "tps_std": 0.07,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 116.83,
+      "file_size_gib": 59.02,
+      "name_params_b": 116.83,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19091,9 +13372,9 @@
       "fa": true,
       "context": "longctx32768",
       "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 36.03,
-      "tps_std": 0.02,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 653.09,
+      "tps_std": 1.74,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -19106,8 +13387,36 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
+      "model_clean": "gpt-oss-120b-mxfp4",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 36.34,
+      "tps_std": 0.22,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 116.83,
+      "file_size_gib": 59.02,
+      "name_params_b": 116.83,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19120,8 +13429,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 177.27,
-      "tps_std": 0.45,
+      "tps_mean": 412.83,
+      "tps_std": 4.5,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -19134,8 +13443,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19148,7 +13457,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 27.69,
+      "tps_mean": 28.15,
       "tps_std": 0.13,
       "error": false,
       "error_type": null,
@@ -19162,8 +13471,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19176,8 +13485,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 575.01,
-      "tps_std": 3.11,
+      "tps_mean": 664.67,
+      "tps_std": 3.1,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -19190,8 +13499,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19204,8 +13513,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 51.2,
-      "tps_std": 0.06,
+      "tps_mean": 51.42,
+      "tps_std": 0.05,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -19218,8 +13527,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19232,8 +13541,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 214.22,
-      "tps_std": 0.5,
+      "tps_mean": 234.02,
+      "tps_std": 0.46,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -19246,8 +13555,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19260,7 +13569,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 34.54,
+      "tps_mean": 35.87,
       "tps_std": 0.04,
       "error": false,
       "error_type": null,
@@ -19274,8 +13583,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19288,8 +13597,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 129.83,
-      "tps_std": 0.04,
+      "tps_mean": 143.42,
+      "tps_std": 0.11,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -19302,8 +13611,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19316,8 +13625,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 26.17,
-      "tps_std": 0.07,
+      "tps_mean": 27.87,
+      "tps_std": 0.02,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -19330,8 +13639,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_amdvlk__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19344,8 +13653,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 635.12,
-      "tps_std": 4.23,
+      "tps_mean": 719.91,
+      "tps_std": 4.36,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -19358,8 +13667,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19372,7 +13681,7 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 56.31,
+      "tps_mean": 56.61,
       "tps_std": 0.08,
       "error": false,
       "error_type": null,
@@ -19386,8 +13695,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19400,8 +13709,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 290.19,
-      "tps_std": 1.03,
+      "tps_mean": 307.5,
+      "tps_std": 0.37,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -19414,8 +13723,8 @@
       "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -19428,260 +13737,92 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 42.28,
-      "tps_std": 0.06,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 139.61,
-      "tps_std": 0.56,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
-      "model_clean": "gpt-oss-120b-mxfp4",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 35.0,
-      "tps_std": 0.05,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 116.83,
-      "file_size_gib": 59.02,
-      "name_params_b": 116.83,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 2029.48,
-      "tps_std": 6.68,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 72.53,
-      "tps_std": 0.11,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 494.19,
-      "tps_std": 1.26,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 51.81,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 277.16,
-      "tps_std": 1.25,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 40.39,
+      "tps_mean": 43.01,
       "tps_std": 0.04,
       "error": false,
       "error_type": null,
-      "backend": "ROCm",
+      "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
+      "params_b": 116.83,
+      "file_size_gib": 59.02,
+      "name_params_b": 116.83,
       "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_2-pr21344__fa1__longctx65536.log",
+      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
+      "model_clean": "gpt-oss-120b-mxfp4",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 145.92,
+      "tps_std": 0.15,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 116.83,
+      "file_size_gib": 59.02,
+      "name_params_b": 116.83,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-120b-mxfp4-00001-of-00003",
+      "model_clean": "gpt-oss-120b-mxfp4",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 34.97,
+      "tps_std": 0.1,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 116.83,
+      "file_size_gib": 59.02,
+      "name_params_b": 116.83,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-120b-mxfp4-00001-of-00003__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gpt-oss-20b-mxfp4",
       "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_2",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1631.69,
-      "tps_std": 17.03,
+      "tps_mean": 1786.17,
+      "tps_std": 19.42,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -19691,24 +13832,192 @@
       "file_size_gib": 11.27,
       "name_params_b": 20.91,
       "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1.log",
+      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "f53577432",
-        "number": "8942"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gpt-oss-20b-mxfp4",
       "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_2",
+      "env": "rocm-7_2_3",
       "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "env_variant": "7_2_3",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 72.53,
+      "tps_mean": 73.05,
+      "tps_std": 0.06,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 1003.11,
+      "tps_std": 2.13,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 52.17,
+      "tps_std": 0.17,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 601.63,
+      "tps_std": 1.09,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 40.68,
+      "tps_std": 0.09,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1787.57,
+      "tps_std": 15.77,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 72.68,
       "tps_std": 0.1,
       "error": false,
       "error_type": null,
@@ -19719,347 +14028,11 @@
       "file_size_gib": 11.27,
       "name_params_b": 20.91,
       "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 490.56,
-      "tps_std": 2.76,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 51.69,
-      "tps_std": 0.16,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 275.76,
-      "tps_std": 1.52,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 40.22,
-      "tps_std": 0.34,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_2__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1636.3,
-      "tps_std": 15.57,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 72.58,
-      "tps_std": 0.05,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 487.06,
-      "tps_std": 3.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 51.73,
-      "tps_std": 0.18,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 276.26,
-      "tps_std": 0.34,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 40.32,
-      "tps_std": 0.08,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm-7_2_3__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1694.78,
-      "tps_std": 17.2,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
       "log": "results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 72.19,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -20072,8 +14045,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 730.43,
-      "tps_std": 1.15,
+      "tps_mean": 1027.43,
+      "tps_std": 1.91,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -20086,8 +14059,8 @@
       "log": "results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -20100,962 +14073,598 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 57.22,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 440.23,
-      "tps_std": 0.27,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm6_4_4",
-      "env_base": "rocm6_4_4",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 47.51,
-      "tps_std": 0.29,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1635.73,
-      "tps_std": 10.16,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 72.47,
-      "tps_std": 0.03,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 493.84,
-      "tps_std": 1.85,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 52.05,
-      "tps_std": 0.04,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 277.7,
-      "tps_std": 0.33,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "rocm7-nightlies",
-      "env_base": "rocm7",
-      "env_variant": "nightlies",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 40.28,
-      "tps_std": 0.41,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1306.44,
-      "tps_std": 9.57,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 73.63,
-      "tps_std": 0.06,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 365.33,
-      "tps_std": 0.21,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 49.83,
-      "tps_std": 0.07,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 210.72,
-      "tps_std": 0.07,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "vulkan_amdvlk",
-      "env_base": "vulkan_amdvlk",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 38.53,
-      "tps_std": 0.01,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1574.45,
-      "tps_std": 16.63,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 79.03,
-      "tps_std": 0.17,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 545.65,
+      "tps_mean": 57.29,
       "tps_std": 0.37,
       "error": false,
       "error_type": null,
-      "backend": "Vulkan",
+      "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
       "params_b": 20.91,
       "file_size_gib": 11.27,
       "name_params_b": 20.91,
       "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log",
+      "log": "results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gpt-oss-20b-mxfp4",
       "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
-      "env_variant": null,
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 60.7,
-      "tps_std": 0.02,
-      "error": false,
-      "error_type": null,
-      "backend": "Vulkan",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 20.91,
-      "file_size_gib": 11.27,
-      "name_params_b": 20.91,
-      "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
-      }
-    },
-    {
-      "model": "gpt-oss-20b-mxfp4",
-      "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
       "env_variant": null,
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 229.28,
-      "tps_std": 0.91,
+      "tps_mean": 613.02,
+      "tps_std": 1.23,
       "error": false,
       "error_type": null,
-      "backend": "Vulkan",
+      "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
       "params_b": 20.91,
       "file_size_gib": 11.27,
       "name_params_b": 20.91,
       "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx65536.log",
+      "log": "results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
       "model": "gpt-oss-20b-mxfp4",
       "model_clean": "gpt-oss-20b-mxfp4",
-      "env": "vulkan_radv",
-      "env_base": "vulkan_radv",
+      "env": "rocm6_4_4",
+      "env_base": "rocm6_4_4",
       "env_variant": null,
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 50.14,
-      "tps_std": 0.11,
+      "tps_mean": 47.67,
+      "tps_std": 0.25,
       "error": false,
       "error_type": null,
-      "backend": "Vulkan",
+      "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
       "params_b": 20.91,
       "file_size_gib": 11.27,
       "name_params_b": 20.91,
       "quant": "MXFP4",
-      "log": "results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx65536.log",
+      "log": "results/gpt-oss-20b-mxfp4__rocm6_4_4__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1512.85,
-      "tps_std": 4.84,
+      "tps_mean": 1812.57,
+      "tps_std": 16.28,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1.log",
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 50.54,
-      "tps_std": 0.15,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 163.39,
-      "tps_std": 1.12,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 5.65,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 72.65,
-      "tps_std": 0.35,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_2-pr21344",
-      "env_base": "rocm",
-      "env_variant": "7_2_2-pr21344",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "tg32 @ d65536",
-      "tps_mean": 3.0,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_2-pr21344__fa1__longctx65536.log",
-      "rpc": false,
-      "build": {
-        "hash": "7957de9dc",
-        "number": "8645"
-      }
-    },
-    {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "pp512",
-      "tps_mean": 1542.98,
-      "tps_std": 7.94,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "default",
-      "context_tokens": null,
-      "test": "tg128",
-      "tps_mean": 50.52,
-      "tps_std": 0.12,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_2__fa1.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "pp2048 @ d32768",
-      "tps_mean": 162.64,
-      "tps_std": 0.31,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx32768",
-      "context_tokens": 32768,
-      "test": "tg32 @ d32768",
-      "tps_mean": 5.64,
-      "tps_std": 0.0,
-      "error": false,
-      "error_type": null,
-      "backend": "ROCm",
-      "ngl": 99,
-      "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_2__fa1__longctx32768.log",
-      "rpc": false,
-      "build": {
-        "hash": "f53577432",
-        "number": "8942"
-      }
-    },
-    {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
-      "fa": true,
-      "context": "longctx65536",
-      "context_tokens": 65536,
-      "test": "pp2048 @ d65536",
-      "tps_mean": 72.92,
+      "tps_mean": 73.03,
       "tps_std": 0.24,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_2__fa1__longctx65536.log",
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_2",
-      "env_base": "rocm",
-      "env_variant": "7_2_2",
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 1124.14,
+      "tps_std": 2.36,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 52.23,
+      "tps_std": 0.24,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 679.48,
+      "tps_std": 9.08,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "rocm7-nightlies",
+      "env_base": "rocm7",
+      "env_variant": "nightlies",
       "fa": true,
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 3.0,
-      "tps_std": 0.0,
+      "tps_mean": 40.79,
+      "tps_std": 0.14,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_2__fa1__longctx65536.log",
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1543.35,
-      "tps_std": 8.0,
+      "tps_mean": 1376.47,
+      "tps_std": 10.73,
       "error": false,
       "error_type": null,
-      "backend": "ROCm",
+      "backend": "Vulkan",
       "ngl": 99,
       "mmap": 0,
-      "params_b": 6.74,
-      "file_size_gib": 3.56,
-      "name_params_b": 6.74,
-      "quant": "Q4_0",
-      "log": "results/llama-2-7b.Q4_0__rocm-7_2_3__fa1.log",
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
-      "model": "llama-2-7b.Q4_0",
-      "model_clean": "llama-2-7b.Q4_0",
-      "env": "rocm-7_2_3",
-      "env_base": "rocm",
-      "env_variant": "7_2_3",
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
       "fa": true,
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 50.5,
+      "tps_mean": 74.15,
+      "tps_std": 0.04,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 388.64,
+      "tps_std": 0.03,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 51.77,
       "tps_std": 0.12,
       "error": false,
       "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 228.5,
+      "tps_std": 0.15,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "vulkan_amdvlk",
+      "env_base": "vulkan_amdvlk",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 40.52,
+      "tps_std": 0.0,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__vulkan_amdvlk__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1692.49,
+      "tps_std": 16.61,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 79.78,
+      "tps_std": 0.23,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__vulkan_radv__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "pp2048 @ d32768",
+      "tps_mean": 555.73,
+      "tps_std": 1.5,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx32768",
+      "context_tokens": 32768,
+      "test": "tg32 @ d32768",
+      "tps_mean": 61.12,
+      "tps_std": 0.08,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx32768.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "pp2048 @ d65536",
+      "tps_mean": 237.74,
+      "tps_std": 1.25,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "gpt-oss-20b-mxfp4",
+      "model_clean": "gpt-oss-20b-mxfp4",
+      "env": "vulkan_radv",
+      "env_base": "vulkan_radv",
+      "env_variant": null,
+      "fa": true,
+      "context": "longctx65536",
+      "context_tokens": 65536,
+      "test": "tg32 @ d65536",
+      "tps_mean": 50.13,
+      "tps_std": 0.03,
+      "error": false,
+      "error_type": null,
+      "backend": "Vulkan",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 20.91,
+      "file_size_gib": 11.27,
+      "name_params_b": 20.91,
+      "quant": "MXFP4",
+      "log": "results/gpt-oss-20b-mxfp4__vulkan_radv__fa1__longctx65536.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "llama-2-7b.Q4_0",
+      "model_clean": "llama-2-7b.Q4_0",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "pp512",
+      "tps_mean": 1545.36,
+      "tps_std": 22.91,
+      "error": false,
+      "error_type": null,
       "backend": "ROCm",
       "ngl": 99,
       "mmap": 0,
@@ -21066,8 +14675,36 @@
       "log": "results/llama-2-7b.Q4_0__rocm-7_2_3__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
+      }
+    },
+    {
+      "model": "llama-2-7b.Q4_0",
+      "model_clean": "llama-2-7b.Q4_0",
+      "env": "rocm-7_2_3",
+      "env_base": "rocm",
+      "env_variant": "7_2_3",
+      "fa": true,
+      "context": "default",
+      "context_tokens": null,
+      "test": "tg128",
+      "tps_mean": 50.59,
+      "tps_std": 0.03,
+      "error": false,
+      "error_type": null,
+      "backend": "ROCm",
+      "ngl": 99,
+      "mmap": 0,
+      "params_b": 6.74,
+      "file_size_gib": 3.56,
+      "name_params_b": 6.74,
+      "quant": "Q4_0",
+      "log": "results/llama-2-7b.Q4_0__rocm-7_2_3__fa1.log",
+      "rpc": false,
+      "build": {
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21080,8 +14717,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 159.17,
-      "tps_std": 0.71,
+      "tps_mean": 159.95,
+      "tps_std": 1.29,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -21094,8 +14731,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21108,7 +14745,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 5.64,
+      "tps_mean": 5.6,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -21122,8 +14759,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm-7_2_3__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21136,8 +14773,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 72.03,
-      "tps_std": 0.23,
+      "tps_mean": 71.44,
+      "tps_std": 0.29,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -21150,8 +14787,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21178,8 +14815,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm-7_2_3__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "8e1f9d083",
-        "number": "9112"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21192,8 +14829,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1524.35,
-      "tps_std": 2.98,
+      "tps_mean": 1532.82,
+      "tps_std": 13.52,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -21206,8 +14843,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21220,8 +14857,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 50.78,
-      "tps_std": 0.02,
+      "tps_mean": 50.83,
+      "tps_std": 0.04,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -21234,8 +14871,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm6_4_4__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21248,8 +14885,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 196.44,
-      "tps_std": 1.93,
+      "tps_mean": 187.91,
+      "tps_std": 1.55,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -21262,8 +14899,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21276,7 +14913,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 6.96,
+      "tps_mean": 6.94,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -21290,8 +14927,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21304,8 +14941,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 83.57,
-      "tps_std": 0.98,
+      "tps_mean": 82.46,
+      "tps_std": 0.59,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -21318,8 +14955,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21332,8 +14969,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 3.74,
-      "tps_std": 0.02,
+      "tps_mean": 3.76,
+      "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -21346,8 +14983,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm6_4_4__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21360,8 +14997,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1513.21,
-      "tps_std": 2.79,
+      "tps_mean": 1516.11,
+      "tps_std": 11.19,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -21374,8 +15011,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21388,8 +15025,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 51.41,
-      "tps_std": 0.13,
+      "tps_mean": 51.27,
+      "tps_std": 0.05,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -21402,8 +15039,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm7-nightlies__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21416,8 +15053,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 189.28,
-      "tps_std": 0.4,
+      "tps_mean": 204.17,
+      "tps_std": 1.96,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -21430,8 +15067,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21444,7 +15081,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 5.62,
+      "tps_mean": 7.0,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -21458,8 +15095,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21472,8 +15109,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 85.33,
-      "tps_std": 0.3,
+      "tps_mean": 94.78,
+      "tps_std": 1.11,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -21486,8 +15123,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21500,8 +15137,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 2.9,
-      "tps_std": 0.03,
+      "tps_mean": 3.79,
+      "tps_std": 0.0,
       "error": false,
       "error_type": null,
       "backend": "ROCm",
@@ -21514,8 +15151,8 @@
       "log": "results/llama-2-7b.Q4_0__rocm7-nightlies__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21528,8 +15165,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 326.38,
-      "tps_std": 0.44,
+      "tps_mean": 347.25,
+      "tps_std": 0.61,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -21542,8 +15179,8 @@
       "log": "results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21556,8 +15193,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 55.44,
-      "tps_std": 0.11,
+      "tps_mean": 55.69,
+      "tps_std": 0.15,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -21570,8 +15207,8 @@
       "log": "results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21584,8 +15221,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 145.14,
-      "tps_std": 0.2,
+      "tps_mean": 155.9,
+      "tps_std": 0.09,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -21598,8 +15235,8 @@
       "log": "results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21612,7 +15249,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 10.21,
+      "tps_mean": 9.42,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -21626,8 +15263,8 @@
       "log": "results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21640,8 +15277,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 89.9,
-      "tps_std": 0.68,
+      "tps_mean": 102.05,
+      "tps_std": 0.15,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -21654,8 +15291,8 @@
       "log": "results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21668,7 +15305,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 4.61,
+      "tps_mean": 5.12,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -21682,8 +15319,8 @@
       "log": "results/llama-2-7b.Q4_0__vulkan_amdvlk__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21696,8 +15333,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "pp512",
-      "tps_mean": 1311.79,
-      "tps_std": 0.38,
+      "tps_mean": 1337.7,
+      "tps_std": 0.75,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -21710,8 +15347,8 @@
       "log": "results/llama-2-7b.Q4_0__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21724,8 +15361,8 @@
       "context": "default",
       "context_tokens": null,
       "test": "tg128",
-      "tps_mean": 55.59,
-      "tps_std": 0.03,
+      "tps_mean": 55.73,
+      "tps_std": 0.04,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -21738,8 +15375,8 @@
       "log": "results/llama-2-7b.Q4_0__vulkan_radv__fa1.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21752,8 +15389,8 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "pp2048 @ d32768",
-      "tps_mean": 210.35,
-      "tps_std": 0.6,
+      "tps_mean": 216.79,
+      "tps_std": 1.07,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -21766,8 +15403,8 @@
       "log": "results/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21780,7 +15417,7 @@
       "context": "longctx32768",
       "context_tokens": 32768,
       "test": "tg32 @ d32768",
-      "tps_mean": 8.53,
+      "tps_mean": 8.47,
       "tps_std": 0.01,
       "error": false,
       "error_type": null,
@@ -21794,8 +15431,8 @@
       "log": "results/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx32768.log",
       "rpc": false,
       "build": {
-        "hash": "3f8752b55",
-        "number": "8743"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21808,8 +15445,8 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "pp2048 @ d65536",
-      "tps_mean": 104.4,
-      "tps_std": 1.78,
+      "tps_mean": 79.73,
+      "tps_std": 0.65,
       "error": false,
       "error_type": null,
       "backend": "Vulkan",
@@ -21822,8 +15459,8 @@
       "log": "results/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     },
     {
@@ -21836,7 +15473,7 @@
       "context": "longctx65536",
       "context_tokens": 65536,
       "test": "tg32 @ d65536",
-      "tps_mean": 4.6,
+      "tps_mean": 4.58,
       "tps_std": 0.0,
       "error": false,
       "error_type": null,
@@ -21850,8 +15487,8 @@
       "log": "results/llama-2-7b.Q4_0__vulkan_radv__fa1__longctx65536.log",
       "rpc": false,
       "build": {
-        "hash": "ab6120cde",
-        "number": "8997"
+        "hash": "0253fb21f",
+        "number": "9187"
       }
     }
   ]
diff --git a/toolboxes/Dockerfile.rocm-6.4.4 b/toolboxes/Dockerfile.rocm-6.4.4
index c15f2cb..54c5c1e 100644
--- a/toolboxes/Dockerfile.rocm-6.4.4
+++ b/toolboxes/Dockerfile.rocm-6.4.4
@@ -1,24 +1,32 @@
 # build stage
 FROM registry.fedoraproject.org/fedora:43 AS builder
 
+# we want to keep ALL dnf downloaded packages in dnf cache,
+# this is not the default, hackily add it to the config,
+# assume there is not much in it. should be improved
+RUN echo 'keepcache=True' >> /etc/dnf/dnf.conf
+
 # deps + rocm toolchain (native fedora packages)
-RUN dnf -y --nodocs --setopt=install_weak_deps=False \
+RUN --mount=type=cache,target=/var/cache/libdnf5 \
+  dnf -y --nodocs --setopt=install_weak_deps=False \
   install \
   make gcc cmake lld clang clang-devel compiler-rt libcurl-devel \
   rocminfo radeontop 'rocm-*' 'rocblas-*' hipblas 'hipblas-*' \
-  git vim rsync sudo tar xz patch \
-  && dnf clean all && rm -rf /var/cache/dnf/*
+  git vim rsync sudo tar xz patch
 
 # llama.cpp
 WORKDIR /opt/llama.cpp
 ARG REPO=https://github.com/ggerganov/llama.cpp.git
 ARG BRANCH=master
-RUN git clone -b ${BRANCH} --single-branch --recursive ${REPO} .
+RUN --mount=type=cache,id=rocm-644,target=/opt/llama.cpp \
+  test -d .git || git clone -b ${BRANCH} --single-branch --recursive ${REPO} .
 
 COPY llama-grammar.patch /tmp/llama-grammar.patch
 
 # build + install
-RUN git clean -xdf \
+RUN --mount=type=cache,id=rocm-644,target=/opt/llama.cpp \
+  git reset --hard \
+  && git pull \
   && git submodule update --recursive \
   && patch -p1 < /tmp/llama-grammar.patch \
   && HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
@@ -32,30 +40,33 @@ RUN git clean -xdf \
   && cmake --install build --config Release
 
 # libs
-RUN find /opt/llama.cpp/build -type f -name 'lib*.so*' -exec cp {} /usr/lib64/ \; \
+RUN --mount=type=cache,id=rocm-644,target=/opt/llama.cpp \
+  mkdir -p /usr/local/lib64 \
+  && find /opt/llama.cpp/build -type f -name 'lib*.so*' -exec cp {} /usr/local/lib64/ \; \
   && ldconfig
 
 # runtime stage
 FROM registry.fedoraproject.org/fedora-minimal:43
+RUN echo 'keepcache=True' >> /etc/dnf/dnf.conf
 
 # runtime deps (native fedora rocm + utilities)
-RUN microdnf -y --nodocs --setopt=install_weak_deps=0 \
+RUN --mount=type=cache,target=/var/cache/libdnf5 \
+  microdnf -y --nodocs --setopt=install_weak_deps=0 \
   install \
   bash ca-certificates libatomic libstdc++ libgcc libgomp sudo \
   rocminfo radeontop 'rocm-*' 'rocblas-*' hipblas 'hipblas-*' \
-  procps-ng \
-  && microdnf clean all && rm -rf /var/cache/dnf/*
+  procps-ng
 
 # copy
 COPY --from=builder /usr/local/ /usr/local/
 COPY --from=builder /usr/include/rocwmma /usr/include/rocwmma
-COPY --from=builder /opt/llama.cpp/build/bin/rpc-* /usr/local/bin/
 
 # ld
 RUN echo "/usr/local/lib"  > /etc/ld.so.conf.d/local.conf \
   && echo "/usr/local/lib64" >> /etc/ld.so.conf.d/local.conf \
   && ldconfig \
   && cp -n /usr/local/lib/libllama*.so* /usr/lib64/ 2>/dev/null || true \
+  && cp -n /usr/local/lib64/libllama*.so* /usr/lib64/ 2>/dev/null || true \
   && ldconfig
 
 # helper
diff --git a/toolboxes/Dockerfile.rocm-7.2.3 b/toolboxes/Dockerfile.rocm-7.2.3
index 57e63a5..73389ff 100644
--- a/toolboxes/Dockerfile.rocm-7.2.3
+++ b/toolboxes/Dockerfile.rocm-7.2.3
@@ -14,16 +14,21 @@ gpgkey=https://repo.radeon.com/rocm/rocm.gpg.key
 REPO
 EOF
 
+# we want to keep ALL dnf downloaded packages in dnf cache,
+# this is not the default, hackily add it to the config,
+# assume there is not much in it. should be improved
+RUN echo 'keepcache=True' >> /etc/dnf/dnf.conf
+
 # deps
-RUN dnf -y --nodocs --setopt=install_weak_deps=False \
+RUN --mount=type=cache,id=dnf-723,target=/var/cache/libdnf5 \
+  dnf -y --nodocs --setopt=install_weak_deps=False \
   --exclude='*sdk*' --exclude='*samples*' --exclude='*-doc*' --exclude='*-docs*' \
   install \
   make gcc cmake lld clang clang-devel compiler-rt libcurl-devel ninja-build \
   rocm-llvm rocm-device-libs hip-runtime-amd hip-devel \
   rocblas rocblas-devel hipblas hipblas-devel rocm-cmake libomp-devel libomp \
   rocminfo radeontop \
-  git-core vim sudo rsync patch \
-  && dnf clean all && rm -rf /var/cache/dnf/*
+  git-core vim sudo rsync patch
 
 # rocm env
 ENV ROCM_PATH=/opt/rocm \
@@ -36,12 +41,15 @@ ENV ROCM_PATH=/opt/rocm \
 WORKDIR /opt/llama.cpp
 ARG REPO=https://github.com/ggerganov/llama.cpp.git
 ARG BRANCH=master
-RUN git clone -b ${BRANCH} --single-branch --recursive ${REPO} .
+RUN --mount=type=cache,id=rocm-723,target=/opt/llama.cpp \
+  test -d .git || git clone -b ${BRANCH} --single-branch --recursive ${REPO} .
 
 COPY llama-grammar.patch /tmp/llama-grammar.patch
 
 # build
-RUN git clean -xdf \
+RUN --mount=type=cache,id=rocm-723,target=/opt/llama.cpp \
+  git reset --hard \
+  && git pull \
   && git submodule update --recursive \
   && patch -p1 < /tmp/llama-grammar.patch \
   && cmake -S . -B build \
@@ -59,7 +67,9 @@ RUN git clean -xdf \
   && cmake --install build --config Release
 
 # libs
-RUN find /opt/llama.cpp/build -type f -name 'lib*.so*' -exec cp {} /usr/lib64/ \; \
+RUN --mount=type=cache,id=rocm-723,target=/opt/llama.cpp \
+  mkdir -p /usr/local/lib64 \
+  && find /opt/llama.cpp/build -type f -name 'lib*.so*' -exec cp {} /usr/local/lib64/ \; \
   && ldconfig
 
 # helper
@@ -67,7 +77,8 @@ COPY gguf-vram-estimator.py /usr/local/bin/gguf-vram-estimator.py
 RUN chmod +x /usr/local/bin/gguf-vram-estimator.py
 
 # runtime stage
-FROM registry.fedoraproject.org/fedora-minimal:43
+FROM registry.fedoraproject.org/fedora-minimal:43 AS runtime
+RUN echo 'keepcache=True' >> /etc/dnf/dnf.conf
 
 # rocm 7.2.3 repo
 RUN <<'EOF'
@@ -83,23 +94,23 @@ REPO
 EOF
 
 # runtime deps
-RUN microdnf -y --nodocs --setopt=install_weak_deps=0 \
+RUN --mount=type=cache,id=dnf-723,target=/var/cache/libdnf5 \
+  microdnf -y --nodocs --setopt=install_weak_deps=0 \
   --exclude='*sdk*' --exclude='*samples*' --exclude='*-doc*' --exclude='*-docs*' \
   install \
   bash ca-certificates libatomic libstdc++ libgcc libgomp sudo \
   hip-runtime-amd rocblas hipblas \
-  rocminfo radeontop procps-ng \
-  && microdnf clean all && rm -rf /var/cache/dnf/*
+  rocminfo radeontop procps-ng
 
 # copy
 COPY --from=builder /usr/local/ /usr/local/
-COPY --from=builder /opt/llama.cpp/build/bin/rpc-* /usr/local/bin/
 
 # ld
 RUN echo "/usr/local/lib"  > /etc/ld.so.conf.d/local.conf \
   && echo "/usr/local/lib64" >> /etc/ld.so.conf.d/local.conf \
   && ldconfig \
   && cp -n /usr/local/lib/libllama*.so* /usr/lib64/ 2>/dev/null || true \
+  && cp -n /usr/local/lib64/libllama*.so* /usr/lib64/ 2>/dev/null || true \
   && ldconfig
 
 # helper
diff --git a/toolboxes/Dockerfile.rocm7-nightlies b/toolboxes/Dockerfile.rocm7-nightlies
index 922c871..ce48205 100644
--- a/toolboxes/Dockerfile.rocm7-nightlies
+++ b/toolboxes/Dockerfile.rocm7-nightlies
@@ -1,16 +1,29 @@
 # build
 FROM registry.fedoraproject.org/fedora:43 AS builder
 
-RUN dnf -y --nodocs --setopt=install_weak_deps=False install \
+# we want to keep ALL dnf downloaded packages in dnf cache,
+# this is not the default, hackily add it to the config,
+# assume there is not much in it. should be improved
+RUN echo 'keepcache=True' >> /etc/dnf/dnf.conf
+
+# we keep the dnf cache as a docker mount, which is shared
+# between layers and also between the various dockerfiles.
+# We can thus reuse the already-downloaded dnf packages
+# on subsequent runs, and the cache will NOT be part of
+# the resulting image, so we don't need to dnf clean.
+RUN --mount=type=cache,id=dnf-nightlies,target=/var/cache/libdnf5 \
+  dnf -y --nodocs --setopt=install_weak_deps=False install \
   make gcc cmake lld clang clang-devel compiler-rt libcurl-devel \
-  radeontop git vim patch curl ninja-build tar xz aria2c \
-  && dnf clean all && rm -rf /var/cache/dnf/*
+  radeontop git vim patch curl ninja-build tar xz aria2c
 
 # find & fetch the latest Linux 7.x.x tarball (gfx1151)
 WORKDIR /tmp
 ARG ROCM_MAJOR_VER=7
 ARG GFX=gfx1151
-RUN set -euo pipefail; \
+# We cache the downloaded tarball to avoid redownloading
+# if no newer version is present.
+RUN --mount=type=cache,id=rocm-nightlies-tmp,target=/tmp \
+  set -euo pipefail; \
   BASE="https://therock-nightly-tarball.s3.amazonaws.com"; \
   PREFIX="therock-dist-linux-${GFX}-${ROCM_MAJOR_VER}"; \
   KEY="$(curl -s "${BASE}?list-type=2&prefix=${PREFIX}" \
@@ -18,9 +31,15 @@ RUN set -euo pipefail; \
   | grep -o "therock-dist-linux-${GFX}-${ROCM_MAJOR_VER}\..*\.tar\.gz" \
   | sort -V | tail -n1)"; \
   echo "Latest tarball: ${KEY}"; \
-  aria2c -x 16 -s 16 -j 16 --file-allocation=none "${BASE}/${KEY}" -o therock.tar.gz
-RUN mkdir -p /opt/rocm-7.0 \
-  && tar xzf therock.tar.gz -C /opt/rocm-7.0 --strip-components=1
+  FILENAME="therock-${KEY}.tar.gz"; \
+  if [ -f "$FILENAME" ]; then \
+    echo "no new build available, using cached tarball"; \
+  else \
+    echo "downloading..."; \
+    aria2c -x 16 -s 16 -j 16 --file-allocation=none "${BASE}/${KEY}" -o "$FILENAME"; \
+  fi; \
+  mkdir -p /opt/rocm-7.0 \
+  && tar xzf "$FILENAME" -C /opt/rocm-7.0 --strip-components=1
 
 ENV ROCM_PATH=/opt/rocm-7.0 \
   HIP_PLATFORM=amd \
@@ -33,7 +52,7 @@ ENV ROCM_PATH=/opt/rocm-7.0 \
   LD_LIBRARY_PATH=/opt/rocm-7.0/lib:/opt/rocm-7.0/lib64:/opt/rocm-7.0/llvm/lib \
   LIBRARY_PATH=/opt/rocm-7.0/lib:/opt/rocm-7.0/lib64 \
   CPATH=/opt/rocm-7.0/include \
-  PKG_CONFIG_PATH=/opt/rocm-7.0/lib/pkgconfig 
+  PKG_CONFIG_PATH=/opt/rocm-7.0/lib/pkgconfig
 
 RUN printf '%s\n' \
   'export ROCM_PATH=/opt/rocm-7.0' \
@@ -55,13 +74,17 @@ RUN printf '%s\n' \
 WORKDIR /opt/llama.cpp
 ARG REPO=https://github.com/ggerganov/llama.cpp.git
 ARG BRANCH=master
-RUN git clone -b ${BRANCH} --single-branch --recursive ${REPO} . \
-  && git clean -xdf \
-  && git submodule update --recursive
+RUN --mount=type=cache,id=rocm-nightlies,target=/opt/llama.cpp \
+  test -d .git || git clone -b ${BRANCH} --single-branch --recursive ${REPO} .
 
 COPY llama-grammar.patch /tmp/llama-grammar.patch
 
-RUN patch -p1 < /tmp/llama-grammar.patch \
+# build
+RUN --mount=type=cache,id=rocm-nightlies,target=/opt/llama.cpp \
+  git reset --hard \
+  && git pull \
+  && git submodule update --recursive \
+  && patch -p1 < /tmp/llama-grammar.patch \
   && cmake -S . -B build \
   -DGGML_HIP=ON \
   -DAMDGPU_TARGETS=gfx1151 \
@@ -71,21 +94,29 @@ RUN patch -p1 < /tmp/llama-grammar.patch \
   && cmake --build build --config Release -- -j$(nproc) \
   && cmake --install build --config Release
 
+# libs
+RUN --mount=type=cache,id=rocm-nightlies,target=/opt/llama.cpp \
+  mkdir -p /usr/local/lib64 \
+  && find /opt/llama.cpp/build -type f -name 'lib*.so*' -exec cp {} /usr/local/lib64/ \; \
+  && ldconfig
+
 # keep bin; drop headers/docs/static libs (retain llama.cpp for rpc binaries)
-RUN find /opt/rocm-7.0 -type f -name '*.a' -delete \
+# TODO: combine with above nightly tarball download
+RUN --mount=type=cache,id=rocm-nightlies,target=/opt/llama.cpp \
+  find /opt/rocm-7.0 -type f -name '*.a' -delete \
   && rm -rf /opt/rocm-7.0/include /opt/rocm-7.0/share \
   /opt/rocm-7.0/llvm/include /opt/rocm-7.0/llvm/share
 
 # runtime
 FROM registry.fedoraproject.org/fedora-minimal:43
+RUN echo 'keepcache=True' >> /etc/dnf/dnf.conf
 
-RUN microdnf -y --nodocs --setopt=install_weak_deps=0 install \
-  bash ca-certificates libatomic libstdc++ libgcc radeontop vim procps-ng \
-  && microdnf clean all && rm -rf /var/cache/dnf/*
+RUN --mount=type=cache,id=dnf-nightlies,target=/var/cache/libdnf5 \
+  microdnf -y --nodocs --setopt=install_weak_deps=0 install \
+  bash ca-certificates libatomic libstdc++ libgcc radeontop vim procps-ng
 
 COPY --from=builder /opt/rocm-7.0 /opt/rocm-7.0
 COPY --from=builder /usr/local/ /usr/local/
-COPY --from=builder /opt/llama.cpp/build/bin/rpc-* /usr/local/bin/
 
 COPY gguf-vram-estimator.py /usr/local/bin/
 RUN chmod +x /usr/local/bin/gguf-vram-estimator.py
@@ -123,6 +154,9 @@ RUN printf '%s\n' \
 # make /usr/local libs visible without touching env
 RUN echo "/usr/local/lib"  > /etc/ld.so.conf.d/local.conf \
   && echo "/usr/local/lib64" >> /etc/ld.so.conf.d/local.conf \
+  && ldconfig \
+  && cp -n /usr/local/lib/libllama*.so* /usr/lib64/ 2>/dev/null || true \
+  && cp -n /usr/local/lib64/libllama*.so* /usr/lib64/ 2>/dev/null || true \
   && ldconfig
 
 CMD ["/bin/bash"]
diff --git a/toolboxes/Dockerfile.vulkan-amdvlk b/toolboxes/Dockerfile.vulkan-amdvlk
index eb90012..fbe835e 100644
--- a/toolboxes/Dockerfile.vulkan-amdvlk
+++ b/toolboxes/Dockerfile.vulkan-amdvlk
@@ -1,30 +1,40 @@
 # build stage
 FROM registry.fedoraproject.org/fedora:43 AS builder
+# we want to keep ALL dnf downloaded packages in dnf cache,
+# this is not the default, hackily add it to the config,
+# assume there is not much in it. should be improved
+RUN echo 'keepcache=True' >> /etc/dnf/dnf.conf
 
 # deps
-RUN dnf -y --nodocs --setopt=install_weak_deps=False install \
+RUN --mount=type=cache,id=dnf-amdvlk,target=/var/cache/libdnf5 \
+  dnf -y --nodocs --setopt=install_weak_deps=False install \
   git vim \
   make gcc cmake ninja-build lld clang clang-devel compiler-rt libcurl-devel \
   vulkan-loader-devel vulkaninfo mesa-vulkan-drivers \
-  spirv-headers-devel radeontop glslc wget patch \
-  && dnf clean all && rm -rf /var/cache/dnf/*
+  spirv-headers-devel radeontop glslc wget patch
 
 # amdvlk
-RUN curl -L -o /tmp/amdvlk-2025.Q2.1.x86_64.rpm \
+RUN \
+  --mount=type=cache,id=vulkan-radv-tmp,target=/tmp \
+  --mount=type=cache,id=dnf-amdvlk,target=/var/cache/libdnf5 \
+  test -f /tmp/amdvlk-2025.Q2.1.x86_64.rpm || \
+  curl -L -o /tmp/amdvlk-2025.Q2.1.x86_64.rpm \
   https://github.com/GPUOpen-Drivers/AMDVLK/releases/download/v-2025.Q2.1/amdvlk-2025.Q2.1.x86_64.rpm \
-  && dnf -y install /tmp/amdvlk-*.rpm \
-  && rm -f /tmp/amdvlk-*.rpm
+  && dnf -y install /tmp/amdvlk-*.rpm
 
 # llama.cpp
 WORKDIR /opt/llama.cpp
 ARG REPO=https://github.com/ggerganov/llama.cpp.git
 ARG BRANCH=master
-RUN git clone -b ${BRANCH} --single-branch --recursive ${REPO} .
+RUN --mount=type=cache,id=vulkan-radv,target=/opt/llama.cpp \
+  test -d .git || git clone -b ${BRANCH} --single-branch --recursive ${REPO} .
 
 COPY llama-grammar.patch /tmp/llama-grammar.patch
 
 # build
-RUN git clean -xdf \
+RUN --mount=type=cache,id=vulkan-radv,target=/opt/llama.cpp \
+  git reset --hard \
+  && git pull \
   && git submodule update --recursive \
   && patch -p1 < /tmp/llama-grammar.patch \
   && cmake -S . -B build -G Ninja \
@@ -35,11 +45,12 @@ RUN git clean -xdf \
   -DLLAMA_BUILD_TESTS=OFF \
   -DLLAMA_BUILD_EXAMPLES=ON \
   -DLLAMA_BUILD_SERVER=ON \
-  && cmake --build build --config Release \
+  && cmake --build build --config Release -- -j$(nproc) \
   && cmake --install build --config Release
 
 # libs
-RUN find /opt/llama.cpp/build -type f -name 'lib*.so*' -exec cp {} /usr/lib64/ \; \
+RUN --mount=type=cache,id=vulkan-radv,target=/opt/llama.cpp \
+  find /opt/llama.cpp/build -type f -name 'lib*.so*' -exec cp {} /usr/lib64/ \; \
   && ldconfig
 
 # helper
@@ -49,23 +60,25 @@ RUN chmod +x /usr/local/bin/gguf-vram-estimator.py
 
 # runtime stage
 FROM registry.fedoraproject.org/fedora-minimal:43
+RUN echo 'keepcache=True' >> /etc/dnf/dnf.conf
 
 # runtime deps
-RUN microdnf -y --nodocs --setopt=install_weak_deps=0 install \
+RUN --mount=type=cache,id=dnf-amdvlk,target=/var/cache/libdnf5 \
+  microdnf -y --nodocs --setopt=install_weak_deps=0 install \
   bash ca-certificates libatomic libstdc++ libgcc \
-  vulkan-loader vulkan-loader-devel vulkaninfo mesa-vulkan-drivers radeontop procps-ng \
-  && microdnf clean all && rm -rf /var/cache/dnf/*
+  vulkan-loader vulkan-loader-devel vulkaninfo mesa-vulkan-drivers radeontop procps-ng
 
-# amdvlk
-RUN curl -L -o /tmp/amdvlk-2025.Q2.1.x86_64.rpm \
+RUN \
+  --mount=type=cache,id=vulkan-radv-tmp,target=/tmp \
+  --mount=type=cache,id=dnf-amdvlk,target=/var/cache/libdnf5 \
+  test -f /tmp/amdvlk-2025.Q2.1.x86_64.rpm || \
+  curl -L -o /tmp/amdvlk-2025.Q2.1.x86_64.rpm \
   https://github.com/GPUOpen-Drivers/AMDVLK/releases/download/v-2025.Q2.1/amdvlk-2025.Q2.1.x86_64.rpm \
-  && microdnf -y install /tmp/amdvlk-*.rpm \
-  && rm -f /tmp/amdvlk-*.rpm
+  && dnf -y install /tmp/amdvlk-*.rpm
 
 # copy
 COPY --from=builder /usr/ /usr/
 COPY --from=builder /usr/local/ /usr/local/
-COPY --from=builder /opt/llama.cpp/build/bin/rpc-* /usr/local/bin/
 
 # ld
 RUN echo "/usr/local/lib"  > /etc/ld.so.conf.d/local.conf \
diff --git a/toolboxes/Dockerfile.vulkan-radv b/toolboxes/Dockerfile.vulkan-radv
index 096fad9..a9d1e17 100644
--- a/toolboxes/Dockerfile.vulkan-radv
+++ b/toolboxes/Dockerfile.vulkan-radv
@@ -1,24 +1,31 @@
 # build stage
 FROM registry.fedoraproject.org/fedora:43 AS builder
+# we want to keep ALL dnf downloaded packages in dnf cache,
+# this is not the default, hackily add it to the config,
+# assume there is not much in it. should be improved
+RUN echo 'keepcache=True' >> /etc/dnf/dnf.conf
 
 # deps
-RUN dnf -y --nodocs --setopt=install_weak_deps=False install \
+RUN --mount=type=cache,id=dnf-radv,target=/var/cache/libdnf5 \
+  dnf -y --nodocs --setopt=install_weak_deps=False install \
   git vim \
   make gcc cmake ninja-build lld clang clang-devel compiler-rt libcurl-devel \
   vulkan-loader-devel vulkaninfo mesa-vulkan-drivers \
-  spirv-headers-devel radeontop glslc patch \
-  && dnf clean all && rm -rf /var/cache/dnf/*
+  spirv-headers-devel radeontop glslc patch
 
 # llama.cpp
 WORKDIR /opt/llama.cpp
 ARG REPO=https://github.com/ggerganov/llama.cpp.git
 ARG BRANCH=master
-RUN git clone -b ${BRANCH} --single-branch --recursive ${REPO} .
+RUN --mount=type=cache,id=vulkan-radv,target=/opt/llama.cpp \
+  test -d .git || git clone -b ${BRANCH} --single-branch --recursive ${REPO} .
 
 COPY llama-grammar.patch /tmp/llama-grammar.patch
 
 # build
-RUN git clean -xdf \
+RUN --mount=type=cache,id=vulkan-radv,target=/opt/llama.cpp \
+  git reset --hard \
+  && git pull \
   && git submodule update --recursive \
   && patch -p1 < /tmp/llama-grammar.patch \
   && cmake -S . -B build -G Ninja \
@@ -29,11 +36,12 @@ RUN git clean -xdf \
   -DLLAMA_BUILD_TESTS=OFF \
   -DLLAMA_BUILD_EXAMPLES=ON \
   -DLLAMA_BUILD_SERVER=ON \
-  && cmake --build build --config Release \
+  && cmake --build build --config Release -- -j$(nproc) \
   && cmake --install build --config Release
 
 # libs
-RUN find /opt/llama.cpp/build -type f -name 'lib*.so*' -exec cp {} /usr/lib64/ \; \
+RUN --mount=type=cache,id=vulkan-radv,target=/opt/llama.cpp \
+  find /opt/llama.cpp/build -type f -name 'lib*.so*' -exec cp {} /usr/lib64/ \; \
   && ldconfig
 
 # helper
@@ -43,17 +51,17 @@ RUN chmod +x /usr/local/bin/gguf-vram-estimator.py
 
 # runtime stage
 FROM registry.fedoraproject.org/fedora-minimal:43
+RUN echo 'keepcache=True' >> /etc/dnf/dnf.conf
 
 # runtime deps
-RUN microdnf -y --nodocs --setopt=install_weak_deps=0 install \
+RUN --mount=type=cache,id=dnf-radv,target=/var/cache/libdnf5 \
+  microdnf -y --nodocs --setopt=install_weak_deps=0 install \
   bash ca-certificates libatomic libstdc++ libgcc \
-  vulkan-loader vulkan-loader-devel vulkaninfo mesa-vulkan-drivers radeontop procps-ng \
-  && microdnf clean all && rm -rf /var/cache/dnf/*
+  vulkan-loader vulkan-loader-devel vulkaninfo mesa-vulkan-drivers radeontop procps-ng
 
 # copy
 COPY --from=builder /usr/ /usr/
 COPY --from=builder /usr/local/ /usr/local/
-COPY --from=builder /opt/llama.cpp/build/bin/rpc-* /usr/local/bin/
 
 # ld
 RUN echo "/usr/local/lib"  > /etc/ld.so.conf.d/local.conf \