[MoE Refactor] Move select_experts from FusedMoEQuantMethod -> FusedMoE (#31996)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-01-22 18:21:35 -05:00
committed by GitHub
parent fc56f4a071
commit dc917cceb8
22 changed files with 498 additions and 533 deletions
@@ -6,7 +6,6 @@ from typing import Any, Optional
import torch
from vllm.distributed import get_tensor_model_parallel_rank, get_tp_group
from vllm.model_executor.layers.fused_moe import FusedMoERouter
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig,
int4_w4a16_moe_quant_config,
@@ -365,17 +364,13 @@ class MoeWNA16Method(FusedMoEMethodBase):
def apply(
self,
layer: FusedMoE,
router: FusedMoERouter,
x: torch.Tensor,
router_logits: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
from vllm.model_executor.layers.fused_moe import fused_experts
assert layer.activation == "silu", "Only SiLU activation is supported."
topk_weights, topk_ids = router.select_experts(
hidden_states=x,
router_logits=router_logits,
)
return fused_experts(
x,