[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
@@ -9,14 +9,15 @@ import regex as re
import torch
from vllm import envs
from vllm.model_executor.layers.fused_moe import (
FusedMoEMethodBase,
RoutedExperts,
SharedExperts,
)
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.layer import (
FusedMoE,
FusedMoEMethodBase,
)
from vllm.model_executor.layers.fused_moe.unquantized_fused_moe_method import (
UnquantizedFusedMoEMethod,
)
@@ -334,20 +335,20 @@ class HummingConfig(QuantizationConfig):
self, layer: torch.nn.Module, prefix: str
) -> "QuantizeMethodBase | None":
layer_type = "other"
if isinstance(layer, FusedMoE):
if isinstance(layer, RoutedExperts):
layer_type = "moe"
elif isinstance(layer, LinearBase):
layer_type = "linear"
quant_config = self.get_quant_config_for_layer(prefix, layer_type)
if quant_config is None:
if isinstance(layer, FusedMoE):
if isinstance(layer, RoutedExperts):
return UnquantizedFusedMoEMethod(layer.moe_config)
elif isinstance(layer, LinearBase):
return UnquantizedLinearMethod()
elif isinstance(layer, LinearBase):
return HummingLinearMethod(quant_config)
elif isinstance(layer, FusedMoE):
elif isinstance(layer, RoutedExperts):
return HummingMoEMethod(quant_config, layer.moe_config)
return None
@@ -764,7 +765,7 @@ class HummingMoEMethod(FusedMoEMethodBase):
return get_humming_moe_quant_config(layer)
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
def process_weights_after_loading(self, layer: RoutedExperts) -> None:
if getattr(self, "processed", False):
return
self.processed = True
@@ -898,10 +899,11 @@ class HummingMoEMethod(FusedMoEMethodBase):
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 | tuple[torch.Tensor, torch.Tensor]:
workspace1, workspace2, output = self.experts.make_workspaces(