update benchmarks

benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log

@@ -0,0 +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
+| 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 |
+
+build: 2405d59cb (8577)

benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log

@@ -0,0 +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(+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)

benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log

@@ -0,0 +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(+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)

benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm-7_2__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm6_4_4__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__rocm7-nightlies__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_amdvlk__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-BF16-00001-of-00002__vulkan_radv__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-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: 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm6_4_4__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__rocm7-nightlies__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_amdvlk__fa1__longctx32768.log

@@ -0,0 +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(+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)

benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/GLM-4.7-Flash-UD-Q8_K_XL__vulkan_radv__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm6_4_4__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__rocm7-nightlies__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1.log

@@ -0,0 +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(+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)

benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_amdvlk__fa1__longctx32768.log

@@ -0,0 +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(+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)

benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/Ministral-3-14B-Instruct-2512-BF16__vulkan_radv__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__rocm7-nightlies__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_amdvlk__fa1__longctx32768.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1.log

@@ -0,0 +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
+| 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)

benchmark/results/29-03-2026/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003__vulkan_radv__fa1__longctx32768.log

@@ -0,0 +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
+| 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)