update bs390 of qwen3-30b-tp2(bw)

[Qwen3-30B] MoE configs for BW1000 TP=2 (bs390)

update bs390 of qwen3-30b-tp2(bw)
[Qwen3-30B] MoE configs for BW1000 TP=2 (bs390)
4c92e64a · zhuwenwen · b0e99e82 · 4c92e64a
Commit 4c92e64a authored Nov 28, 2025 by zhuwenwen
Hide whitespace changes
Inline Side-by-side

Showing with 10 additions and 1 deletion

vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=gfx936_80cu_nn.json ...d_moe/configs/E=128,N=384,device_name=gfx936_80cu_nn.json +10 -1

No files found.
--- a/vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=gfx936_80cu_nn.json
+++ b/vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=gfx936_80cu_nn.json
@@ -99,7 +99,7 @@
        "num_ldmatrixes": 1
    },
    "256": {
-        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_M": 16,
        "BLOCK_SIZE_N": 128,
        "BLOCK_SIZE_K": 64,
        "GROUP_SIZE_M": 1,
@@ -107,6 +107,15 @@
        "num_stages": 2,
        "num_ldmatrixes": 1
    },
+    "390": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "num_ldmatrixes": 1
+    },
    "512": {
        "BLOCK_SIZE_M": 16,
        "BLOCK_SIZE_N": 256,