[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
@@ -13,7 +13,6 @@ from vllm.model_executor.layers.fused_moe import (
FusedMoE,
FusedMoEConfig,
FusedMoEMethodBase,
FusedMoERouter,
FusedMoeWeightScaleSupported,
)
from vllm.model_executor.layers.fused_moe.config import (
@@ -351,15 +350,10 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
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]:
topk_weights, topk_ids = router.select_experts(
hidden_states=x,
router_logits=router_logits,
)
if self.rocm_aiter_moe_enabled:
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
rocm_aiter_fused_experts,
@@ -388,7 +382,6 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
None,
layer.w13_weight_scale,
layer.w2_weight_scale,
router_logits,
topk_weights,
topk_ids,
quant_type_id=scalar_types.float8_e4m3fn.id,
@@ -544,15 +537,10 @@ class QuarkW4A8Fp8MoEMethod(QuarkMoEMethod):
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]:
topk_weights, topk_ids = router.select_experts(
hidden_states=x,
router_logits=router_logits,
)
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
rocm_aiter_fused_experts,
)
@@ -753,15 +741,10 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
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]:
topk_weights, topk_ids = router.select_experts(
hidden_states=x,
router_logits=router_logits,
)
if not self.emulate:
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
rocm_aiter_fused_experts,