From de509ae8eb4467bd789234d909e5c725b8fba326 Mon Sep 17 00:00:00 2001 From: Kaixi Hou Date: Sat, 26 Jul 2025 07:10:36 -0700 Subject: [PATCH] [NVIDIA] Explicitly disable shuffled weights for flashinfer blockscale moe fp8 kernels (#21411) Signed-off-by: kaixih --- vllm/model_executor/layers/fused_moe/fused_moe.py | 1 + 1 file changed, 1 insertion(+) diff --git a/vllm/model_executor/layers/fused_moe/fused_moe.py b/vllm/model_executor/layers/fused_moe/fused_moe.py index c412f695ae766..1985e8612da35 100644 --- a/vllm/model_executor/layers/fused_moe/fused_moe.py +++ b/vllm/model_executor/layers/fused_moe/fused_moe.py @@ -1127,6 +1127,7 @@ def flashinfer_fused_moe_blockscale_fp8( tile_tokens_dim=_get_tile_tokens_dim(x.shape[0], top_k, global_num_experts), routing_method_type=2, # DeepSeek-styled routing method + use_shuffled_weight=False, )