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 13, 00:45 UTC
[Bugfix] Fix FusedMoEModularKernel for triton backend (#28913)
Signed-off-by: Xin Yang <xyangx@amazon.com>
Xin Yang committed
Nov 18, 2025 at 21:05 UTC
468a8d72bac181c1499320478940cec64363e107
1 file changed
+6
-4
vllm/model_executor/layers/quantization/mxfp4.py
+6
-4
@@ -755,8 +755,10 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
755
756
self.w13_weight = w13_weight
757
self.w2_weight = w2_weight
758
- layer.w13_weight = Parameter(w13_weight.storage.data, requires_grad=False)
759
- layer.w2_weight = Parameter(w2_weight.storage.data, requires_grad=False)
758
+ del layer.w13_weight
759
+ del layer.w2_weight
760
+ layer.w13_weight = w13_weight
761
+ layer.w2_weight = w2_weight
762
else:
763
raise ValueError(f"Unsupported backend: {self.mxfp4_backend}")
764
@@ -1065,8 +1067,8 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
1067
1068
return triton_kernel_moe_forward(
1069
hidden_states=x,
1068
- w1=self.w13_weight,
1069
- w2=self.w2_weight,
1070
+ w1=layer.w13_weight,
1071
+ w2=layer.w2_weight,
1072
gating_output=router_logits,
1073
topk=top_k,
1074
renormalize=renormalize,