[MoE Refactor] EPLB refactoring for FusedMoE (#41055)

Signed-off-by: Bill Nell <bnell@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
bnellnm
2026-05-12 14:16:31 -04:00
committed by GitHub
co-authored by Robert Shaw
parent 4d591db470
commit d9b4990783
17 changed files with 83 additions and 98 deletions
@@ -311,10 +311,6 @@ class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):
shared_experts: SharedExperts | None,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor:
if layer.enable_eplb:
raise NotImplementedError(
"EPLB not supported for `CompressedTensorsW4A8Fp8MoEMethod` yet."
)
assert self.moe_quant_config is not None
from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import (
@@ -307,7 +307,6 @@ class CompressedTensorsW4A8Int8MoEMethod(CompressedTensorsMoEMethod):
router_logits: torch.Tensor,
input_ids: torch.Tensor | None = None,
) -> torch.Tensor:
assert not layer.enable_eplb, "EPLB not supported for W4A8-int MoE yet."
assert layer.activation in (
MoEActivation.SILU,
MoEActivation.SWIGLUOAI,
@@ -2101,7 +2101,7 @@ class ModelOptMxFp8FusedMoE(FusedMoEMethodBase):
assert self.mxfp8_backend == Fp8MoeBackend.FLASHINFER_TRTLLM
if layer.enable_eplb:
if layer.eplb_state is not None:
raise NotImplementedError(
"EPLB is not supported for FlashInfer TRTLLM MXFP8 MoE backend."
)