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 6, 19:15 UTC
[MoE] Add routing simulation override for MXFP4 quantized MoE (#33595)
Signed-off-by: Jaewon Lee <jaewon@meta.com>
Jaewon committed
Mar 12, 2026 at 17:30 UTC
aaa3092f5137870d7a30e17bdbcd3f8268fa4c29
1 file changed
+6
vllm/model_executor/layers/quantization/mxfp4.py
+6
@@ -1109,6 +1109,12 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
1109
layer.eplb_state.logical_replica_count,
1110
), "MXFP4 are not supported with this configuration."
1111
1112
+ # Apply routing simulation strategy if specified.
1113
+ # This applies to all monolithic backends (SM100_FI and TRITON).
1114
+ routing_strategy = envs.VLLM_MOE_ROUTING_SIMULATION_STRATEGY
1115
+ if routing_strategy == "uniform_random":
1116
+ router_logits = torch.rand_like(router_logits)
1117
+
1118
if (
1119
self.mxfp4_backend == Mxfp4Backend.SM100_FI_MXFP4_MXFP8_TRTLLM
1120
or self.mxfp4_backend == Mxfp4Backend.SM100_FI_MXFP4_BF16