[fix]: disable cutlass block scaled group gemm for EP (#20781)

Signed-off-by: Duncan Moss <djm.moss@gmail.com>
2025-07-10 19:39:18 -07:00
parent 0cf893cae1
commit 5923ab9524
3 changed files with 34 additions and 9 deletions
--- a/vllm/model_executor/layers/fused_moe/fused_moe.py
+++ b/vllm/model_executor/layers/fused_moe/fused_moe.py
@@ -1192,8 +1192,9 @@ def fused_experts(
            apply_router_weight_on_input=apply_router_weight_on_input,
        )
    elif (allow_cutlass_block_scaled_grouped_gemm and use_fp8_w8a8
-          and _valid_cutlass_block_scaled_grouped_gemm(w1, w2)):
-        assert apply_router_weight_on_input is False
+          and _valid_cutlass_block_scaled_grouped_gemm(
+              w1, w2, inplace, activation, apply_router_weight_on_input,
+              expert_map)):
        return run_cutlass_block_scaled_fused_experts(
            a=hidden_states,
            w1=w1,