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, 05:30 UTC
[Misc][DP] Guard mxfp4 implementation selection (#27484)
Signed-off-by: Varun Sundar Rabindranath <vsundarr@redhat.com> Co-authored-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
Varun Sundar Rabindranath committed
Oct 24, 2025 at 19:29 UTC
269c4db0a4f2635d0703f53f67decd4bb93b046b
1 file changed
+7
-2
vllm/model_executor/layers/quantization/mxfp4.py
+7
-2
@@ -794,7 +794,8 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
794
)
795
else:
796
raise NotImplementedError(
797
- "Incompatible Mxfp4 backend for EP batched experts format"
797
+ f"Incompatible Mxfp4 backend ({self.mxfp4_backend}) for "
798
+ "EP batched experts format"
799
)
800
else:
801
assert self.moe_quant_config is not None
@@ -813,8 +814,12 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
814
return TrtLlmGenExperts(self.moe, self.moe_quant_config, **kwargs)
815
elif self.mxfp4_backend == Mxfp4Backend.MARLIN:
816
return MarlinExperts(self.moe_quant_config)
816
- else:
817
+ elif self.mxfp4_backend == Mxfp4Backend.TRITON:
818
return OAITritonExperts(self.moe_quant_config)
819
+ else:
820
+ raise NotImplementedError(
821
+ f"Incompatible Mxfp4 backend ({self.mxfp4_backend}) for EP"
822
+ )
823
824
def _route_and_experts(
825
self,