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 7, 11:30 UTC
fix(mxfp4): return is_monolithic=False when LoRA is enabled for Triton backend (#35382)
Signed-off-by: Seungho Yoon <yoonsnowdev@gmail.com>
Seungho Yoon committed
Mar 1, 2026 at 23:59 UTC
5a435507d877f4eb16802095037d5c56e767c589
1 file changed
+2
vllm/model_executor/layers/quantization/mxfp4.py
+2
@@ -1001,6 +1001,8 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
1001
1002
@property
1003
def is_monolithic(self) -> bool:
1004
+ if self.moe.is_lora_enabled:
1005
+ return False
1006
return (
1007
self.mxfp4_backend == Mxfp4Backend.SM100_FI_MXFP4_MXFP8_TRTLLM
1008
or self.mxfp4_backend == Mxfp4Backend.SM100_FI_MXFP4_BF16