Updated benchmarks

benchmark/delete_transient_failures.py

@@ -0,0 +1,141 @@
+#!/usr/bin/env python3
+import argparse
+import glob
+import os
+import re
+
+RESULTS_DIR_DEFAULT = "results"
+
+# Same detection logic as your extractor
+HEADER_RE = re.compile(r"^\|\s*model\s*\|", re.IGNORECASE)
+SEP_RE    = re.compile(r"^\|\s*-+")
+
+LOAD_ERR    = re.compile(r"failed to load model|Device memory allocation.*failed|⚠️\s*Fail", re.IGNORECASE)
+HANG_ERR    = re.compile(r"GPU Hang|HW Exception", re.IGNORECASE)
+GENERIC_ERR = re.compile(r"error:|exit \d+|runtime error|⚠️\s*Runtime Error", re.IGNORECASE)
+
+
+def parse_table(text):
+    lines = text.splitlines()
+    rows = []
+    header = None
+    col_idx = {}
+
+    for line in lines:
+        if HEADER_RE.search(line):
+            header = [c.strip().lower() for c in line.strip().strip("|").split("|")]
+            for idx, name in enumerate(header):
+                col_idx[name] = idx
+            continue
+
+        if header and (SEP_RE.search(line) or not line.strip()):
+            continue
+
+        if header and line.startswith("|"):
+            parts = [c.strip() for c in line.strip().strip("|").split("|")]
+            if len(parts) < len(header):
+                continue
+            row = {}
+            for name, idx in col_idx.items():
+                row[name] = parts[idx]
+            rows.append(row)
+
+        if header and line.strip() == "" and rows:
+            break
+
+    return rows
+
+
+def detect_error(text):
+    if LOAD_ERR.search(text):
+        return True
+    if HANG_ERR.search(text):
+        return True
+    if GENERIC_ERR.search(text):
+        return True
+    return False
+
+
+def is_non_transient_vram_issue(text):
+    # Do NOT delete logs with this kind of Vulkan OOM
+    return (
+        "ggml_vulkan: Device memory allocation of size" in text
+        and "Requested buffer size exceeds device buffer size limit" in text
+    )
+
+
+def is_failed_run(text):
+    table_rows = parse_table(text)
+
+    has_pp = any(r.get("test", "").lower() == "pp512" for r in table_rows)
+    has_tg = any(r.get("test", "").lower() == "tg128" for r in table_rows)
+
+    if has_pp or has_tg:
+        return False
+
+    return detect_error(text)
+
+
+def main():
+    ap = argparse.ArgumentParser(
+        description="Delete transient-failure benchmark logs in results/"
+    )
+    ap.add_argument(
+        "--results-dir",
+        default=RESULTS_DIR_DEFAULT,
+        help="Directory containing *.log files (default: results)",
+    )
+    ap.add_argument(
+        "--dry-run",
+        action="store_true",
+        help="Only print what would be deleted",
+    )
+    args = ap.parse_args()
+
+    results_dir = args.results_dir
+    pattern = os.path.join(results_dir, "*.log")
+
+    to_delete = []
+    skipped_non_transient = []
+
+    for path in sorted(glob.glob(pattern)):
+        try:
+            with open(path, errors="ignore") as f:
+                text = f.read()
+        except OSError as e:
+            print(f"Could not read {path}: {e}")
+            continue
+
+        if not is_failed_run(text):
+            continue
+
+        if is_non_transient_vram_issue(text):
+            skipped_non_transient.append(path)
+            continue
+
+        to_delete.append(path)
+
+    if not to_delete and not skipped_non_transient:
+        print("No failed logs found.")
+        return
+
+    if skipped_non_transient:
+        print("Keeping logs with non transient VRAM issues:")
+        for p in skipped_non_transient:
+            print(f"  KEEP  {p}")
+
+    if to_delete:
+        print("Deleting logs with transient failures:")
+        for p in to_delete:
+            print(f"  DELETE {p}")
+            if not args.dry_run:
+                try:
+                    os.remove(p)
+                except OSError as e:
+                    print(f"    Failed to delete {p}: {e}")
+    else:
+        print("No logs to delete.")
+
+
+if __name__ == "__main__":
+    main()

benchmark/generate_results.json.py

@@ -34,26 +34,45 @@ NAME_B_RE = re.compile(r"(\d+(?:\.\d+)?)B")
 # Shard suffix in filenames
 SHARD_RE = re.compile(r"-000\d+-of-000\d+", re.IGNORECASE)
 
+# Long-context suffix in filenames (e.g., __longctx32768)
+LONGCTX_RE = re.compile(r"longctx(\d+)", re.IGNORECASE)
+
 # --- Helpers ---------------------------------------------------------------
 
 def clean_model_name(raw):
     base = SHARD_RE.sub("", raw)
     return base
 
-def parse_env_and_fa(basename):
+def parse_env_flags(basename):
-    # pattern: <model>__<env>[__fa1][__hblt0]
+    """
+    pattern: <model>__<env>[__fa1][__hblt0][__longctx32768]
+    Returns (env, fa, context_tag, context_tokens)
+    """
     parts = basename.split("__")
     if len(parts) < 2:
-        return None, False
+        return None, False, "default", None
 
     env = parts[1]
-    # scan any extra suffix segments
+    fa = False
-    suffixes = {p.lower() for p in parts[2:]}
+    context_tag = "default"
-    fa = ("fa1" in suffixes)
+    context_tokens = None
-    if "hblt0" in suffixes:
-        env = f"{env}-hblt0"
 
-    return env, fa
+    for raw_suffix in parts[2:]:
+        suffix = raw_suffix.lower()
+        if suffix == "fa1":
+            fa = True
+        elif suffix == "hblt0":
+            env = f"{env}-hblt0"
+        elif suffix.startswith("longctx"):
+            context_tag = suffix
+            m = LONGCTX_RE.search(suffix)
+            if m:
+                try:
+                    context_tokens = int(m.group(1))
+                except ValueError:
+                    context_tokens = None
+
+    return env, fa, context_tag, context_tokens
 
 def env_base_and_variant(env):
     # e.g. "rocm6_4_2-rocwmma" -> ("rocm6_4_2", "rocwmma")
@@ -135,7 +154,8 @@ for path in sorted(glob.glob(os.path.join(RESULTS_DIR, "*.log"))):
         continue
 
     model_raw, _rest = base.split("__", 1)
-    env, fa_from_name = parse_env_and_fa(base)
+    env, fa_from_name, context_tag, context_tokens = parse_env_flags(base)
+    if env:
         envs.add(env)
 
     model_clean = clean_model_name(model_raw)
@@ -215,6 +235,8 @@ for path in sorted(glob.glob(os.path.join(RESULTS_DIR, "*.log"))):
             "env_base": env_base,
             "env_variant": env_variant,         # e.g. "rocwmma"
             "fa": bool(fa_enabled),
+            "context": context_tag or "default",
+            "context_tokens": context_tokens,
             "test": test,                       # "pp512" | "tg128" | None (if error)
             "tps_mean": tps_mean,
             "tps_std": tps_std,

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma-improved__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
+| 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 |    0 |           pp512 |        135.87 ± 0.06 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        103.27 ± 0.47 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         20.46 ± 0.01 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         22.61 ± 0.00 |
 
-build: 11f0af55 (6736)
+build: 31df4608 (7038)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma-improved__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         19.09 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0xe277840) reason :GPU Hang
+✖ ! [rocm-7alpha-rocwmma-improved] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002 __fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
+| 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 |    0 |           pp512 |        160.81 ± 0.78 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |         87.62 ± 0.29 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         20.41 ± 0.01 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         22.57 ± 0.00 |
 
-build: 4807e8f9 (6609)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         13.99 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x3b2ac4f0) reason :GPU Hang
+✖ ! [rocm-7alpha-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002 __fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |         93.53 ± 0.08 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         22.66 ± 0.00 |
+
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm-7alpha__fa1__longctx32768.log

@@ -0,0 +1,6 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: 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: 0x2cfb74f0) on address 0x7fb6c65be000. Reason: Page not present or supervisor privilege.
+✖ ! [rocm-7alpha] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002 __fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma.log

@@ -1,15 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-rocBLAS error: No hipBLASLt solution found
-This message will be only be displayed once, unless the ROCBLAS_VERBOSE_HIPBLASLT_ERROR environment variable is set.
-
-rocBLAS warning: hipBlasLT failed, falling back to tensile. 
-This message will be only be displayed once, unless the ROCBLAS_VERBOSE_TENSILE_ERROR environment variable is set.
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           pp512 |        121.16 ± 0.25 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         20.46 ± 0.00 |
-
-build: 11f0af55 (6736)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| 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 |  1 |    0 |           pp512 |        134.20 ± 0.28 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        135.10 ± 0.15 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.04 ± 0.01 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         21.72 ± 0.00 |
 
-build: 11f0af55 (6736)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         12.41 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x3c42f5c0) reason :GPU Hang
+✖ ! [rocm6_4_4-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002 __fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| 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 |  1 |    0 |           pp512 |        171.53 ± 0.45 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        142.39 ± 0.25 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.08 ± 0.00 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         21.70 ± 0.02 |
 
-build: 11f0af55 (6736)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         13.30 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x1feee5c0) reason :GPU Hang
+✖ ! [rocm6_4_4-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002 __hblt0__fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4.log

@@ -1,15 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-rocBLAS error: No hipBLASLt solution found
-This message will be only be displayed once, unless the ROCBLAS_VERBOSE_HIPBLASLT_ERROR environment variable is set.
-
-rocBLAS warning: hipBlasLT failed, falling back to tensile. 
-This message will be only be displayed once, unless the ROCBLAS_VERBOSE_TENSILE_ERROR environment variable is set.
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           pp512 |        126.30 ± 0.14 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         19.83 ± 0.00 |
-
-build: 4807e8f9 (6609)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| 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 |  1 |    0 |           pp512 |        136.00 ± 0.15 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        135.50 ± 0.06 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         20.93 ± 0.00 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         21.19 ± 0.00 |
 
-build: 4807e8f9 (6609)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         30.21 ± 0.00 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          9.28 ± 0.00 |
+
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| 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 |  1 |    0 |           pp512 |        132.96 ± 0.49 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        172.61 ± 0.32 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         20.99 ± 0.00 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         21.78 ± 0.00 |
 
-build: 4807e8f9 (6609)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         33.91 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x1d99e5c0) reason :GPU Hang
+✖ ! [rocm6_4_4] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002 __hblt0__fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1-rocwmma__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |         77.55 ± 0.60 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         21.70 ± 0.00 |
+
+build: ee8dd5c65 (7035)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1-rocwmma__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         16.62 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x34ae58c0) reason :GPU Hang
+✖ ! [rocm7.1-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002 __fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1-rocwmma__hblt0__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |        146.50 ± 0.48 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         21.72 ± 0.00 |
+
+build: ee8dd5c65 (7035)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1-rocwmma__hblt0__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         17.23 ± 0.00 |
+Memory access fault by GPU node-1 (Agent handle: 0x15b288c0) on address 0x7f4ebb38d000. Reason: Page not present or supervisor privilege.
+✖ ! [rocm7.1-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002 __hblt0__fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |         99.42 ± 0.19 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         21.70 ± 0.00 |
+
+build: ee8dd5c65 (7035)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1__fa1__longctx32768.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         29.83 ± 0.00 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.85 ± 0.00 |
+
+build: ee8dd5c65 (7035)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1__hblt0__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |        172.08 ± 0.59 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         21.73 ± 0.00 |
+
+build: ee8dd5c65 (7035)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7.1__hblt0__fa1__longctx32768.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         32.69 ± 0.00 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.85 ± 0.00 |
+
+build: ee8dd5c65 (7035)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1.log

@@ -2,9 +2,9 @@ ggml_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 |
+| 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 |  1 |    0 |           pp512 |         67.08 ± 0.15 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |         72.38 ± 0.12 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.07 ± 0.00 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         21.73 ± 0.01 |
 
-build: a3cb0474 (6735)
+build: bca95ca51 (7036)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_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 @ d32768 |         16.40 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x1678a6f0) reason :GPU Hang
+✖ ! [rocm7_rc-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002 __fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0.log

@@ -1,10 +0,0 @@
-ggml_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 | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           pp512 |        117.71 ± 0.66 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         20.53 ± 0.00 |
-
-build: a3cb0474 (6735)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1.log

@@ -2,9 +2,9 @@ ggml_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 |
+| 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 |  1 |    0 |           pp512 |        148.21 ± 0.25 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        131.85 ± 0.23 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.10 ± 0.00 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         21.71 ± 0.04 |
 
-build: a3cb0474 (6735)
+build: bca95ca51 (7036)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc-rocwmma__hblt0__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_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 @ d32768 |         17.15 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x106c26f0) reason :GPU Hang
+✖ ! [rocm7_rc-rocwmma] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002 __hblt0__fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc.log

@@ -1,10 +0,0 @@
-ggml_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 | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           pp512 |        119.33 ± 0.28 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         20.19 ± 0.00 |
-
-build: a3cb0474 (6735)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__fa1.log

@@ -2,9 +2,9 @@ ggml_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 |
+| 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 |  1 |    0 |           pp512 |         93.03 ± 0.12 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |         99.25 ± 0.11 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.00 ± 0.00 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         21.54 ± 0.01 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__fa1__longctx32768.log

@@ -0,0 +1,6 @@
+ggml_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
+Memory access fault by GPU node-1 (Agent handle: 0x3c42b6f0) on address 0x7f0a849aa000. Reason: Page not present or supervisor privilege.
+✖ ! [rocm7_rc] GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002 __fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__hblt0.log

@@ -1,10 +0,0 @@
-ggml_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 | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           pp512 |        136.81 ± 0.43 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         20.52 ± 0.00 |
-
-build: a3cb0474 (6735)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1.log

@@ -2,9 +2,9 @@ ggml_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 |
+| 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 |  1 |    0 |           pp512 |        148.95 ± 0.73 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        169.93 ± 0.33 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         21.04 ± 0.00 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         21.71 ± 0.00 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__rocm7_rc__hblt0__fa1__longctx32768.log

@@ -2,9 +2,9 @@ ggml_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 | mmap |            test |                  t/s |
+| 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 |    0 |           pp512 |        102.61 ± 0.20 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         33.65 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         20.54 ± 0.00 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          7.82 ± 0.00 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_amdvlk__fa1.log

@@ -2,7 +2,7 @@ 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 |        218.68 ± 0.54 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           pp512 |        217.91 ± 0.48 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         24.27 ± 0.01 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         24.50 ± 0.06 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-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
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |    0 |           pp512 |        216.84 ± 0.52 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         22.82 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |    0 |           tg128 |         24.15 ± 0.01 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          5.79 ± 0.00 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002__vulkan_radv__fa1.log

@@ -2,7 +2,7 @@ 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 |        236.02 ± 2.60 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           pp512 |        235.07 ± 0.58 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         24.51 ± 0.01 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         24.84 ± 0.01 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q4_K_XL-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
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |    0 |           pp512 |        223.39 ± 1.25 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         33.43 ± 0.00 |
-| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |    0 |           tg128 |         24.06 ± 0.03 |
+| glm4moe 106B.A12B Q4_K - Medium |  68.01 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |         10.49 ± 0.00 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma-improved__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
+| 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 |    0 |           pp512 |        133.48 ± 0.45 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |         92.38 ± 0.37 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         15.77 ± 0.11 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.64 ± 0.00 |
 
-build: 11f0af55 (6736)
+build: 31df4608 (7038)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma-improved__fa1__longctx32768.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
+| 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 |    0 |           pp512 |        140.68 ± 0.66 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         19.45 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         15.84 ± 0.01 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.86 ± 0.00 |
 
-build: 4807e8f9 (6609)
+build: 31df4608 (7038)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |         86.50 ± 0.17 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.61 ± 0.00 |
+
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha-rocwmma__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         14.06 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x97894f0) reason :GPU Hang
+✖ ! [rocm-7alpha-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003 __fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |         74.73 ± 0.27 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.66 ± 0.00 |
+
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm-7alpha__fa1__longctx32768.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         26.97 ± 0.00 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          6.58 ± 0.00 |
+
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma.log

@@ -1,15 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-rocBLAS error: No hipBLASLt solution found
-This message will be only be displayed once, unless the ROCBLAS_VERBOSE_HIPBLASLT_ERROR environment variable is set.
-
-rocBLAS warning: hipBlasLT failed, falling back to tensile. 
-This message will be only be displayed once, unless the ROCBLAS_VERBOSE_TENSILE_ERROR environment variable is set.
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           pp512 |        118.41 ± 0.20 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         15.75 ± 0.16 |
-
-build: 11f0af55 (6736)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| 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 |  1 |    0 |           pp512 |        128.51 ± 0.51 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        125.43 ± 0.26 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.13 ± 0.16 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.48 ± 0.14 |
 
-build: 11f0af55 (6736)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__fa1__longctx32768.log

@@ -0,0 +1,6 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: 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: 0x3a41b5c0) reason :GPU Hang
+✖ ! [rocm6_4_4-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003 __fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| 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 |  1 |    0 |           pp512 |        143.55 ± 0.54 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        140.41 ± 0.79 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.17 ± 0.06 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.52 ± 0.05 |
 
-build: 11f0af55 (6736)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4-rocwmma__hblt0__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         13.20 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x2b2915c0) reason :GPU Hang
+✖ ! [rocm6_4_4-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003 __hblt0__fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4.log

@@ -1,15 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-rocBLAS error: No hipBLASLt solution found
-This message will be only be displayed once, unless the ROCBLAS_VERBOSE_HIPBLASLT_ERROR environment variable is set.
-
-rocBLAS warning: hipBlasLT failed, falling back to tensile. 
-This message will be only be displayed once, unless the ROCBLAS_VERBOSE_TENSILE_ERROR environment variable is set.
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           pp512 |        121.76 ± 0.14 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         15.69 ± 0.00 |
-
-build: 4807e8f9 (6609)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| 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 |  1 |    0 |           pp512 |        129.77 ± 0.12 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        130.63 ± 0.70 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.14 ± 0.00 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.44 ± 0.00 |
 
-build: 4807e8f9 (6609)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__fa1__longctx32768.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         26.75 ± 0.00 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          8.10 ± 0.00 |
+
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| 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 |  1 |    0 |           pp512 |        134.52 ± 0.53 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        145.79 ± 0.11 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.08 ± 0.21 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.57 ± 0.06 |
 
-build: 4807e8f9 (6609)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm6_4_4__hblt0__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         33.65 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x14c455c0) reason :GPU Hang
+✖ ! [rocm6_4_4] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003 __hblt0__fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1-rocwmma__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |         69.31 ± 0.07 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.50 ± 0.09 |
+
+build: ee8dd5c65 (7035)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1-rocwmma__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         17.07 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x231278c0) reason :GPU Hang
+✖ ! [rocm7.1-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003 __fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1-rocwmma__hblt0__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |        136.65 ± 0.08 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.46 ± 0.16 |
+
+build: ee8dd5c65 (7035)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1-rocwmma__hblt0__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         17.05 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x3b6528c0) reason :GPU Hang
+✖ ! [rocm7.1-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003 __hblt0__fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |         94.32 ± 0.20 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.25 ± 0.53 |
+
+build: ee8dd5c65 (7035)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1__fa1__longctx32768.log

@@ -0,0 +1,6 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: 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: 0x3cdc48c0) on address 0x7f1399b6e000. Reason: Page not present or supervisor privilege.
+✖ ! [rocm7.1] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003 __fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1__hblt0__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |        130.72 ± 0.76 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.54 ± 0.04 |
+
+build: ee8dd5c65 (7035)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7.1__hblt0__fa1__longctx32768.log

@@ -0,0 +1,6 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: 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: 0x123888c0) reason :GPU Hang
+✖ ! [rocm7.1] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003 __hblt0__fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma.log

@@ -1,10 +0,0 @@
-ggml_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 | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           pp512 |         97.09 ± 0.15 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         15.89 ± 0.00 |
-
-build: a3cb0474 (6735)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1.log

@@ -2,9 +2,9 @@ ggml_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 |
+| 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 |  1 |    0 |           pp512 |         69.91 ± 0.44 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        114.56 ± 0.18 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.13 ± 0.11 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.58 ± 0.00 |
 
-build: a3cb0474 (6735)
+build: bca95ca51 (7036)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__fa1__longctx32768.log

@@ -2,9 +2,9 @@ ggml_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 | mmap |            test |                  t/s |
+| 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 |    0 |           pp512 |         91.95 ± 0.23 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         16.46 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         15.80 ± 0.00 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          1.65 ± 0.00 |
 
-build: a3cb0474 (6735)
+build: bca95ca51 (7036)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0.log

@@ -1,10 +0,0 @@
-ggml_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 | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           pp512 |        128.74 ± 0.30 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         15.77 ± 0.20 |
-
-build: a3cb0474 (6735)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1.log

@@ -2,9 +2,9 @@ ggml_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 |
+| 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 |  1 |    0 |           pp512 |        133.32 ± 0.82 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        159.14 ± 0.64 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.10 ± 0.20 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.44 ± 0.20 |
 
-build: a3cb0474 (6735)
+build: bca95ca51 (7036)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc-rocwmma__hblt0__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_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 @ d32768 |         17.46 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x2566c6f0) reason :GPU Hang
+✖ ! [rocm7_rc-rocwmma] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003 __hblt0__fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__fa1.log

@@ -2,9 +2,9 @@ ggml_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 |
+| 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 |  1 |    0 |           pp512 |        125.81 ± 0.29 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |         96.45 ± 0.26 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.20 ± 0.00 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.51 ± 0.00 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__fa1__longctx32768.log

@@ -0,0 +1,6 @@
+ggml_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
+Memory access fault by GPU node-1 (Agent handle: 0x19b4f6f0) on address 0x7f5ea34ff000. Reason: Page not present or supervisor privilege.
+✖ ! [rocm7_rc] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003 __fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__hblt0.log

@@ -1,10 +0,0 @@
-ggml_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 | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           pp512 |        142.12 ± 0.60 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |    0 |           tg128 |         15.89 ± 0.00 |
-
-build: a3cb0474 (6735)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1.log

@@ -2,9 +2,9 @@ ggml_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 |
+| 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 |  1 |    0 |           pp512 |        139.60 ± 0.48 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        130.86 ± 0.36 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |  1 |    0 |           tg128 |         16.10 ± 0.17 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |         16.53 ± 0.06 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__rocm7_rc__hblt0__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_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 @ d32768 |         36.62 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x2d1506f0) reason :GPU Hang
+✖ ! [rocm7_rc] GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003 __hblt0__fa1__longctx32768 failed (exit 134)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_amdvlk__fa1.log

@@ -2,7 +2,7 @@ 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 |        262.18 ± 1.19 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           pp512 |        261.54 ± 1.01 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         17.30 ± 0.01 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         17.42 ± 0.00 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_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
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |    0 |           pp512 |        260.51 ± 1.03 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         23.19 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |    0 |           tg128 |         17.26 ± 0.01 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          5.25 ± 0.00 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_K_XL-00001-of-00003__vulkan_radv__fa1.log

@@ -2,7 +2,7 @@ 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 |        233.21 ± 6.28 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           pp512 |        233.87 ± 0.08 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         17.65 ± 0.01 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |           tg128 |         17.74 ± 0.01 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/GLM-4.5-Air-UD-Q6_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
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
+| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |    0 |           pp512 |        222.31 ± 0.71 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 | pp2048 @ d32768 |         33.31 ± 0.00 |
-| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |    0 |           tg128 |         17.43 ± 0.01 |
+| glm4moe 106B.A12B Q6_K         |  94.57 GiB |   110.47 B | Vulkan     |  99 |  1 |    0 |   tg32 @ d32768 |          9.00 ± 0.00 |
 
-build: a3cb0474 (6735)
+build: 1c398dc9e (7034)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma-improved__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |    0 |           pp512 |         98.82 ± 0.24 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        103.85 ± 0.10 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |    0 |           tg128 |          2.78 ± 0.00 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
 
-build: 11f0af55 (6736)
+build: 31df4608 (7038)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma-improved__fa1__longctx32768.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |    0 |           pp512 |         98.12 ± 0.10 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         19.88 ± 0.00 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |    0 |           tg128 |          2.78 ± 0.00 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.46 ± 0.00 |
 
-build: 4807e8f9 (6609)
+build: 31df4608 (7038)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |        103.65 ± 0.07 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
+
+build: 1c398dc9e (7034)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha-rocwmma__fa1__longctx32768.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         12.76 ± 0.00 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          0.80 ± 0.00 |
+
+build: 1c398dc9e (7034)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |        105.64 ± 0.24 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
+
+build: 1c398dc9e (7034)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm-7alpha__fa1__longctx32768.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         33.13 ± 0.00 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.31 ± 0.00 |
+
+build: 1c398dc9e (7034)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma.log

@@ -1,15 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-hipBLASLt error: Heuristic Fetch Failed!
-This message will be only be displayed once, unless the ROCBLAS_VERBOSE_HIPBLASLT_ERROR environment variable is set.
-
-rocBLAS warning: hipBlasLT failed, falling back to tensile. 
-This message will be only be displayed once, unless the ROCBLAS_VERBOSE_TENSILE_ERROR environment variable is set.
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |    0 |           pp512 |         98.00 ± 0.05 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |    0 |           tg128 |          2.77 ± 0.01 |
-
-build: 11f0af55 (6736)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        103.65 ± 0.17 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        102.56 ± 0.06 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |          2.78 ± 0.00 |
 
-build: 11f0af55 (6736)
+build: 1c398dc9e (7034)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         11.74 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x1ed645c0) reason :GPU Hang
+✖ ! [rocm6_4_4-rocwmma] Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002 __fa1__longctx32768 failed (exit 134)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        102.51 ± 0.05 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        103.23 ± 0.06 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |          2.78 ± 0.00 |
 
-build: 11f0af55 (6736)
+build: 1c398dc9e (7034)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4-rocwmma__hblt0__fa1__longctx32768.log

@@ -0,0 +1,9 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         11.69 ± 0.00 |
+HW Exception by GPU node-1 (Agent handle: 0x147425c0) reason :GPU Hang
+✖ ! [rocm6_4_4-rocwmma] Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002 __hblt0__fa1__longctx32768 failed (exit 134)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4.log

@@ -1,15 +0,0 @@
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
-ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
-ggml_cuda_init: found 1 ROCm devices:
-  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-hipBLASLt error: Heuristic Fetch Failed!
-This message will be only be displayed once, unless the ROCBLAS_VERBOSE_HIPBLASLT_ERROR environment variable is set.
-
-rocBLAS warning: hipBlasLT failed, falling back to tensile. 
-This message will be only be displayed once, unless the ROCBLAS_VERBOSE_TENSILE_ERROR environment variable is set.
-| model                          |       size |     params | backend    | ngl | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |    0 |           pp512 |         97.85 ± 0.06 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |    0 |           tg128 |          2.76 ± 0.00 |
-
-build: 4807e8f9 (6609)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        102.47 ± 0.08 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        105.28 ± 0.10 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |          2.77 ± 0.00 |
 
-build: 4807e8f9 (6609)
+build: 1c398dc9e (7034)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__fa1__longctx32768.log

@@ -0,0 +1,6 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: 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: 0x2dba35c0) reason :GPU Hang
+✖ ! [rocm6_4_4] Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002 __fa1__longctx32768 failed (exit 134)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1.log

@@ -2,9 +2,9 @@ ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 ROCm devices:
   Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
-| model                          |       size |     params | backend    | ngl | fa | mmap |            test |                  t/s |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ---: | --------------: | -------------------: |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           pp512 |        104.23 ± 0.24 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           pp512 |        104.32 ± 0.29 |
-| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |          2.78 ± 0.00 |
 
-build: 4807e8f9 (6609)
+build: 1c398dc9e (7034)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm6_4_4__hblt0__fa1__longctx32768.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         34.35 ± 0.00 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          2.40 ± 0.00 |
+
+build: 1c398dc9e (7034)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1-rocwmma__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |        102.80 ± 0.11 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
+
+build: ee8dd5c65 (7035)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1-rocwmma__fa1__longctx32768.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         15.25 ± 0.00 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          1.07 ± 0.00 |
+
+build: ee8dd5c65 (7035)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1-rocwmma__hblt0__fa1.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |           pp512 |        102.92 ± 0.14 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |           tg128 |          2.79 ± 0.00 |
+
+build: ee8dd5c65 (7035)

benchmark/results/Llama-3.3-70B-Instruct-UD-Q8_K_XL-00001-of-00002__rocm7.1-rocwmma__hblt0__fa1__longctx32768.log

@@ -0,0 +1,10 @@
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 ROCm devices:
+  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 |         15.47 ± 0.00 |
+| llama 70B Q8_0                 |  75.65 GiB |    70.55 B | ROCm       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |          1.07 ± 0.00 |
+
+build: ee8dd5c65 (7035)