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 12, 22:15 UTC
[Bugfix] Move flashinfer kernel check into ```__init__``` function of ```FusedMoE``` (#29018)
Signed-off-by: Max Hu <hyoung2991@gmail.com>
Max Hu committed
Nov 19, 2025 at 16:54 UTC
cb0a7b4bea26657da989562a10055b7d0b59fd3a
1 file changed
+4
-1
vllm/model_executor/layers/fused_moe/layer.py
+4
-1
@@ -574,6 +574,9 @@ class FusedMoE(CustomOp):
574
is_act_and_mul=is_act_and_mul,
575
is_lora_enabled=vllm_config.lora_config is not None,
576
)
577
+ self.moe_config_use_flashinfer_cutlass_kernels = (
578
+ self.moe_config.use_flashinfer_cutlass_kernels
579
+ )
580
581
self.quant_config = quant_config
582
@@ -728,7 +731,7 @@ class FusedMoE(CustomOp):
731
return (
732
self.moe_quant_config is not None
733
and self.moe_quant_config.quant_dtype == "nvfp4"
731
- and self.moe_config.use_flashinfer_cutlass_kernels
734
+ and self.moe_config_use_flashinfer_cutlass_kernels
735
)
736
737
@property