[Bugfix] Fix Qwen3-coder moe tuned config (#24072)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>

benchmarks/kernels/benchmark_moe.py

@@ -678,7 +678,11 @@ def main(args: argparse.Namespace):
         is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
         search_space = get_configs_compute_bound(is_fp16, block_quant_shape)
         print(f"Start tuning over {len(search_space)} configurations...")
-
+        if use_deep_gemm:
+            raise ValueError(
+                "Tuning with --use-deep-gemm is not supported as it only tunes Triton "
+                "kernels. Please remove the flag."
+            )
         start = time.time()
         configs = _distribute(
             "tune",

vllm/model_executor/layers/fused_moe/configs/E=20,N=2560,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -18,18 +18,18 @@
     "4": {
         "BLOCK_SIZE_M": 16,
         "BLOCK_SIZE_N": 128,
-        "BLOCK_SIZE_K": 128,
-        "GROUP_SIZE_M": 1,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
         "num_warps": 4,
         "num_stages": 3
     },
     "8": {
         "BLOCK_SIZE_M": 16,
         "BLOCK_SIZE_N": 128,
-        "BLOCK_SIZE_K": 256,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 64,
         "num_warps": 4,
-        "num_stages": 4
+        "num_stages": 3
     },
     "16": {
         "BLOCK_SIZE_M": 16,
@@ -58,7 +58,7 @@
     "48": {
         "BLOCK_SIZE_M": 64,
         "BLOCK_SIZE_N": 128,
-        "BLOCK_SIZE_K": 128,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 64,
         "num_warps": 4,
         "num_stages": 4
@@ -74,73 +74,73 @@
     "96": {
         "BLOCK_SIZE_M": 64,
         "BLOCK_SIZE_N": 128,
-        "BLOCK_SIZE_K": 128,
-        "GROUP_SIZE_M": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
         "num_warps": 4,
-        "num_stages": 3
+        "num_stages": 4
     },
     "128": {
-        "BLOCK_SIZE_M": 128,
-        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
         "BLOCK_SIZE_K": 128,
-        "GROUP_SIZE_M": 1,
+        "GROUP_SIZE_M": 64,
         "num_warps": 4,
-        "num_stages": 2
+        "num_stages": 4
     },
     "256": {
-        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_M": 64,
         "BLOCK_SIZE_N": 128,
         "BLOCK_SIZE_K": 256,
-        "GROUP_SIZE_M": 64,
+        "GROUP_SIZE_M": 1,
         "num_warps": 4,
-        "num_stages": 3
+        "num_stages": 4
     },
     "512": {
-        "BLOCK_SIZE_M": 256,
-        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
         "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 64,
-        "num_warps": 8,
+        "num_warps": 4,
         "num_stages": 4
     },
     "1024": {
-        "BLOCK_SIZE_M": 256,
-        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
         "BLOCK_SIZE_K": 256,
-        "GROUP_SIZE_M": 16,
+        "GROUP_SIZE_M": 64,
         "num_warps": 4,
         "num_stages": 4
     },
     "1536": {
         "BLOCK_SIZE_M": 64,
-        "BLOCK_SIZE_N": 256,
-        "BLOCK_SIZE_K": 128,
-        "GROUP_SIZE_M": 1,
-        "num_warps": 8,
-        "num_stages": 4
-    },
-    "2048": {
-        "BLOCK_SIZE_M": 32,
         "BLOCK_SIZE_N": 128,
         "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 16,
-        "num_warps": 8,
-        "num_stages": 5
+        "num_warps": 4,
+        "num_stages": 3
     },
-    "3072": {
-        "BLOCK_SIZE_M": 128,
+    "2048": {
+        "BLOCK_SIZE_M": 64,
         "BLOCK_SIZE_N": 128,
         "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 64,
         "num_warps": 4,
-        "num_stages": 4
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
     },
     "4096": {
-        "BLOCK_SIZE_M": 128,
-        "BLOCK_SIZE_N": 256,
-        "BLOCK_SIZE_K": 256,
-        "GROUP_SIZE_M": 64,
-        "num_warps": 8,
-        "num_stages": 5
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
     }
 }