A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 23:45 UTC
[NVIDIA] Explicitly disable shuffled weights for flashinfer blockscale moe fp8 kernels (#21411)
Signed-off-by: kaixih <kaixih@nvidia.com>
Kaixi Hou committed
Jul 26, 2025 at 07:10 UTC
de509ae8eb4467bd789234d909e5c725b8fba326
1 file changed
+1
vllm/model_executor/layers/fused_moe/fused_moe.py
+1
@@ -1127,6 +1127,7 @@ def flashinfer_fused_moe_blockscale_fp8(
1127
tile_tokens_dim=_get_tile_tokens_dim(x.shape[0], top_k,
1128
global_num_experts),
1129
routing_method_type=2, # DeepSeek-styled routing method
1130
+ use_shuffled_weight=False,
1131
)
1132
1133