[MoE Refactor] Remove MoE legacy code (#51078)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-08-05 14:35:36 -07:00
committed by GitHub
parent 4282fe52d5
commit 373fe8b83e
22 changed files with 3 additions and 295 deletions
@@ -8,7 +8,6 @@ from compressed_tensors.quantization import (
QuantizationArgs,
)
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import (
RoutedExperts,
@@ -310,13 +309,6 @@ class CompressedTensorsW4A16FlydslMoEMethod(CompressedTensorsMoEMethod):
block_shape=[0, self.group_size],
)
def select_gemm_impl(
self,
prepare_finalize: mk.FusedMoEPrepareAndFinalizeModular,
layer: torch.nn.Module,
) -> mk.FusedMoEExpertsModular:
raise NotImplementedError
def apply(
self,
layer: RoutedExperts,
@@ -4,7 +4,6 @@
import torch
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import (
FusedMoeWeightScaleSupported,
@@ -242,15 +241,6 @@ class CompressedTensorsW4A4Nvfp4MoEMethod(CompressedTensorsMoEMethod):
)
self.moe_kernel.fused_experts.process_weights_after_loading(layer)
def maybe_make_prepare_finalize(
self,
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
) -> mk.FusedMoEPrepareAndFinalizeModular | None:
raise ValueError(
f"{self.__class__.__name__} uses the new modular kernel initialization "
"logic. This function should not be called."
)
def get_fused_moe_quant_config(self, layer: torch.nn.Module) -> FusedMoEQuantConfig:
return make_nvfp4_moe_quant_config(
backend=self.nvfp4_backend,
@@ -8,7 +8,6 @@ from compressed_tensors.quantization import (
QuantizationStrategy,
)
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import (
@@ -341,15 +340,6 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
layer=layer,
)
def maybe_make_prepare_finalize(
self,
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
) -> mk.FusedMoEPrepareAndFinalizeModular | None:
raise ValueError(
f"{self.__class__.__name__} uses the new modular kernel initialization "
"logic. This function should not be called."
)
def get_fused_moe_quant_config(self, layer: torch.nn.Module) -> FusedMoEQuantConfig:
is_per_token = self.input_quant.strategy == QuantizationStrategy.TOKEN
return make_fp8_moe_quant_config(
@@ -7,7 +7,6 @@ from compressed_tensors.quantization import (
QuantizationStrategy,
)
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import (
FusedMoeWeightScaleSupported,
@@ -164,15 +163,6 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
)
self.moe_kernel.fused_experts.process_weights_after_loading(layer)
def maybe_make_prepare_finalize(
self,
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
) -> mk.FusedMoEPrepareAndFinalizeModular | None:
raise ValueError(
f"{self.__class__.__name__} uses the new modular kernel initialization "
"logic. This function should not be called."
)
def get_fused_moe_quant_config(self, layer: torch.nn.Module) -> FusedMoEQuantConfig:
return make_int8_moe_quant_config(
int8_backend=self.int8_backend,
@@ -3,7 +3,6 @@
import torch
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.model_executor.layers.fused_moe import (
FusedMoeWeightScaleSupported,
RoutedExperts,
@@ -159,15 +158,6 @@ class CompressedTensorsW8A8Mxfp8MoEMethod(CompressedTensorsMoEMethod):
layer=layer,
)
def maybe_make_prepare_finalize(
self,
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
) -> mk.FusedMoEPrepareAndFinalizeModular | None:
raise ValueError(
f"{self.__class__.__name__} uses the new modular kernel "
"initialization logic. This function should not be called."
)
def apply_monolithic(
self,
layer: RoutedExperts,
@@ -6,7 +6,6 @@ from typing import TYPE_CHECKING, Any
import torch
import vllm.envs as envs
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.config import get_current_vllm_config
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.logger import init_logger
@@ -731,15 +730,6 @@ class Fp8MoEMethod(FusedMoEMethodBase):
layer, w13, w2, w13_scale, w2_scale, w13_input_scale, w2_input_scale
)
def maybe_make_prepare_finalize(
self,
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
) -> mk.FusedMoEPrepareAndFinalizeModular | None:
raise ValueError(
f"{self.__class__.__name__} uses the new modular kernel initialization "
"logic. This function should not be called."
)
def get_fused_moe_quant_config(self, layer: RoutedExperts) -> FusedMoEQuantConfig:
w1_scale = getattr(layer, f"w13_{self.weight_scale_name}")
w2_scale = getattr(layer, f"w2_{self.weight_scale_name}")
@@ -8,7 +8,6 @@ import torch
from torch.nn.parameter import Parameter
import vllm.envs as envs
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.config import get_current_vllm_config
from vllm.logger import init_logger
from vllm.model_executor.kernels.linear import (
@@ -771,15 +770,6 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
activation_key=kFp8StaticTensorSym,
)
def maybe_make_prepare_finalize(
self,
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
) -> mk.FusedMoEPrepareAndFinalizeModular | None:
raise ValueError(
f"{self.__class__.__name__} uses the new modular kernel initialization "
"logic. This function should not be called."
)
def create_weights(
self,
layer: RoutedExperts,
@@ -1409,15 +1399,6 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
self.nvfp4_backend
)
def maybe_make_prepare_finalize(
self,
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
) -> mk.FusedMoEPrepareAndFinalizeModular | None:
raise ValueError(
f"{self.__class__.__name__} uses the new modular kernel initialization "
"logic. This function should not be called."
)
def uses_weight_scale_2_pattern(self) -> bool:
"""
FP4 variants use 'weight_scale_2' pattern for per-tensor weight scales.
@@ -2087,15 +2068,6 @@ class ModelOptMxFp8FusedMoE(FusedMoEMethodBase):
):
self._dequant_mxfp8_weights_to_bf16(layer)
def maybe_make_prepare_finalize(
self,
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
) -> mk.FusedMoEPrepareAndFinalizeModular | None:
raise ValueError(
f"{self.__class__.__name__} uses the new modular kernel initialization "
"logic. This function should not be called."
)
def get_fused_moe_quant_config(
self, layer: RoutedExperts
) -> FusedMoEQuantConfig | None:
@@ -5,7 +5,6 @@ from abc import abstractmethod
import torch
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.model_executor.layers.fused_moe import (
FusedMoEMethodBase,
RoutedExperts,
@@ -100,15 +99,6 @@ class OnlineMoEMethodBase(FusedMoEMethodBase):
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
pass
def maybe_make_prepare_finalize(
self,
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
) -> mk.FusedMoEPrepareAndFinalizeModular | None:
raise ValueError(
f"{self.__class__.__name__} uses the new modular kernel "
"initialization logic. This function should not be called."
)
@property
def supports_eplb(self) -> bool:
return True