[MoE] Add routing simulation override for MXFP4 quantized MoE (#33595)
Signed-off-by: Jaewon Lee <jaewon@meta.com>
This commit is contained in:
@@ -1109,6 +1109,12 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
|
|||||||
layer.eplb_state.logical_replica_count,
|
layer.eplb_state.logical_replica_count,
|
||||||
), "MXFP4 are not supported with this configuration."
|
), "MXFP4 are not supported with this configuration."
|
||||||
|
|
||||||
|
# Apply routing simulation strategy if specified.
|
||||||
|
# This applies to all monolithic backends (SM100_FI and TRITON).
|
||||||
|
routing_strategy = envs.VLLM_MOE_ROUTING_SIMULATION_STRATEGY
|
||||||
|
if routing_strategy == "uniform_random":
|
||||||
|
router_logits = torch.rand_like(router_logits)
|
||||||
|
|
||||||
if (
|
if (
|
||||||
self.mxfp4_backend == Mxfp4Backend.SM100_FI_MXFP4_MXFP8_TRTLLM
|
self.mxfp4_backend == Mxfp4Backend.SM100_FI_MXFP4_MXFP8_TRTLLM
|
||||||
or self.mxfp4_backend == Mxfp4Backend.SM100_FI_MXFP4_BF16
|
or self.mxfp4_backend == Mxfp4Backend.SM100_FI_MXFP4_BF16
|
||||||
|
|||||||
Reference in New Issue
Block a user