fix

sufubao · sufubao · commit 6836bf11511e · 2025-08-25T04:09:29.000+08:00
diff --git a/lightllm/common/fused_moe/moe_silu_and_mul.py b/lightllm/common/fused_moe/moe_silu_and_mul.py
@@ -72,7 +72,7 @@ def _silu_and_mul_kernel_fast(
         for bm in [32, 64, 128, 256]
         for bn in [32, 64, 128, 256]
     ],
-    default_config={"BLOCK_M": 128, "BLOCK_N": 128, "num_warps": 4, "num_stages": 1},
+    default_config={"BLOCK_M": 128, "BLOCK_N": 128, "num_warps": 4, "NUM_STAGES": 1},
     static_key_func=lambda input, output: f"N={input.shape[-1] // 2},out_dtype={output.dtype}",
     run_key_func=lambda input: str(nearest_power_of_2(input.shape[0])),
 )

Original file line number	Diff line number	Diff line change
`@@ -72,7 +72,7 @@ def _silu_and_mul_kernel_fast(`
`72`	`72`	`for bm in [32, 64, 128, 256]`
`73`	`73`	`for bn in [32, 64, 128, 256]`
`74`	`74`	`],`
`75`		`- default_config={"BLOCK_M": 128, "BLOCK_N": 128, "num_warps": 4, "num_stages": 1},`
	`75`	`+ default_config={"BLOCK_M": 128, "BLOCK_N": 128, "num_warps": 4, "NUM_STAGES": 1},`
`76`	`76`	`static_key_func=lambda input, output: f"N={input.shape[-1] // 2},out_dtype={output.dtype}",`
`77`	`77`	`run_key_func=lambda input: str(nearest_power_of_2(input.shape[0])),`
`78`	`78`	`)`