[MoE Refactor] Introduce RoutedExperts alias for FusedMoE and don't store SharedExperts in MK (#40735)

Signed-off-by: Bill Nell <bnell@redhat.com>
Signed-off-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Signed-off-by: Robert Shaw <robertgshaw2@gmail.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Robert Shaw <robertgshaw2@gmail.com>
This commit is contained in:
bnellnm
2026-05-12 13:37:44 -04:00
committed by GitHub
co-authored by Robert Shaw Robert Shaw
parent 6ff7405b81
commit 4d591db470
53 changed files with 369 additions and 332 deletions
@@ -6,13 +6,12 @@ from typing import Any, Union
import torch
from packaging import version
from vllm.model_executor.layers.fused_moe.config import (
from vllm.model_executor.layers.fused_moe import (
FusedMoEConfig,
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.layer import (
FusedMoE,
FusedMoEMethodBase,
FusedMoEQuantConfig,
RoutedExperts,
SharedExperts,
)
from vllm.model_executor.layers.linear import (
LinearBase,
@@ -164,7 +163,7 @@ class BitsAndBytesConfig(QuantizationConfig):
if is_layer_skipped_bnb(prefix, self.llm_int8_skip_modules):
return UnquantizedLinearMethod()
return BitsAndBytesLinearMethod(self)
elif isinstance(layer, FusedMoE):
elif isinstance(layer, RoutedExperts):
return BitsAndBytesMoEMethod(self, layer.moe_config)
return None
@@ -451,7 +450,7 @@ class BitsAndBytesMoEMethod(FusedMoEMethodBase):
def create_weights(
self,
layer: torch.nn.Module,
layer: RoutedExperts,
num_experts: int,
hidden_size: int,
intermediate_size_per_partition: int,
@@ -472,16 +471,17 @@ class BitsAndBytesMoEMethod(FusedMoEMethodBase):
)
def get_fused_moe_quant_config(
self, layer: torch.nn.Module
self, layer: RoutedExperts
) -> FusedMoEQuantConfig | None:
return None
def apply(
self,
layer: FusedMoE,
layer: RoutedExperts,
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts: SharedExperts | None,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe import fused_experts