[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
@@ -7,15 +7,14 @@ from vllm.config import get_current_vllm_config
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import (
FusedMoE,
FusedMoEConfig,
FusedMoEMethodBase,
)
from vllm.model_executor.layers.fused_moe import modular_kernel as mk
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEParallelConfig,
FusedMoEQuantConfig,
RoutedExperts,
SharedExperts,
)
from vllm.model_executor.layers.fused_moe import modular_kernel as mk
from vllm.model_executor.layers.fused_moe.oracle.mxfp4 import (
TRITON_BACKENDS,
Mxfp4MoeBackend,
@@ -87,7 +86,7 @@ class Mxfp4Config(QuantizationConfig):
"UnquantizedLinearMethod.",
)
return UnquantizedLinearMethod()
elif isinstance(layer, FusedMoE):
elif isinstance(layer, RoutedExperts):
return GptOssMxfp4MoEMethod(layer.moe_config)
elif isinstance(layer, Attention):
logger.debug_once(
@@ -178,7 +177,7 @@ class GptOssMxfp4MoEMethod(FusedMoEMethodBase):
def create_weights(
self,
layer: torch.nn.Module,
layer: RoutedExperts,
num_experts: int,
hidden_size: int,
intermediate_size_per_partition: int,
@@ -272,7 +271,7 @@ class GptOssMxfp4MoEMethod(FusedMoEMethodBase):
def _setup_kernel(
self,
layer: FusedMoE,
layer: RoutedExperts,
w13: torch.Tensor,
w2: torch.Tensor,
w13_scale: torch.Tensor,
@@ -370,11 +369,10 @@ class GptOssMxfp4MoEMethod(FusedMoEMethodBase):
mxfp4_backend=self.mxfp4_backend,
experts_cls=self.experts_cls,
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
layer=layer,
)
def process_weights_after_loading(self, layer):
def process_weights_after_loading(self, layer: RoutedExperts) -> None:
w13 = layer.w13_weight
w2 = layer.w2_weight
w13_scale = layer.w13_weight_scale
@@ -388,7 +386,7 @@ class GptOssMxfp4MoEMethod(FusedMoEMethodBase):
self._setup_kernel(layer, w13, w2, w13_scale, w2_scale, w13_bias, w2_bias)
def get_fused_moe_quant_config(
self, layer: torch.nn.Module
self, layer: RoutedExperts
) -> FusedMoEQuantConfig | None:
w1_scale = layer.w13_weight_scale
w2_scale = layer.w2_weight_scale
@@ -416,7 +414,7 @@ class GptOssMxfp4MoEMethod(FusedMoEMethodBase):
def select_gemm_impl(
self,
prepare_finalize: mk.FusedMoEPrepareAndFinalize,
layer: torch.nn.Module,
layer: RoutedExperts,
) -> mk.FusedMoEExpertsModular:
raise ValueError(
f"{self.__class__.__name__} uses the new modular kernel "
@@ -425,10 +423,11 @@ class GptOssMxfp4MoEMethod(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:
assert not self.is_monolithic
@@ -443,12 +442,13 @@ class GptOssMxfp4MoEMethod(FusedMoEMethodBase):
global_num_experts=layer.global_num_experts,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
expert_map=layer.expert_map,
shared_experts=shared_experts,
shared_experts_input=shared_experts_input,
)
def apply_monolithic(
self,
layer: FusedMoE,
layer: RoutedExperts,
x: torch.Tensor,
router_logits: torch.Tensor,
input_ids: torch.Tensor | None = None,
@@ -511,7 +511,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
def create_weights(
self,
layer: torch.nn.Module,
layer: RoutedExperts,
num_experts: int,
hidden_size: int,
intermediate_size_per_partition: int,
@@ -605,7 +605,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
def _setup_kernel(
self,
layer: FusedMoE,
layer: RoutedExperts,
w13: torch.Tensor,
w2: torch.Tensor,
w13_scale: torch.Tensor,
@@ -703,7 +703,6 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
mxfp4_backend=self.mxfp4_backend,
experts_cls=self.experts_cls,
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
layer=layer,
)
@@ -721,7 +720,8 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
self._setup_kernel(layer, w13, w2, w13_scale, w2_scale, w13_bias, w2_bias)
def get_fused_moe_quant_config(
self, layer: torch.nn.Module
self,
layer: RoutedExperts,
) -> FusedMoEQuantConfig | None:
w1_scale = layer.w13_weight_scale
w2_scale = layer.w2_weight_scale
@@ -748,7 +748,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
def select_gemm_impl(
self,
prepare_finalize: mk.FusedMoEPrepareAndFinalize,
layer: torch.nn.Module,
layer: RoutedExperts,
) -> mk.FusedMoEExpertsModular:
raise ValueError(
f"{self.__class__.__name__} uses the new modular kernel "
@@ -757,10 +757,11 @@ class Mxfp4MoEMethod(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:
assert not self.is_monolithic
@@ -775,12 +776,13 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
global_num_experts=layer.global_num_experts,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
expert_map=layer.expert_map,
shared_experts=shared_experts,
shared_experts_input=shared_experts_input,
)
def apply_monolithic(
self,
layer: FusedMoE,
layer: RoutedExperts,
x: torch.Tensor,
router_logits: torch.Tensor,
input_ids: torch.Tensor | None = None,