[MoE Refactor] Separate Router into OO Classes (#30623)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-01-18 11:40:49 -05:00
committed by GitHub
parent 2f03035a61
commit 327a02d8db
45 changed files with 1750 additions and 684 deletions
@@ -10,12 +10,12 @@ from torch.nn import Parameter
import vllm.model_executor.layers.fused_moe # noqa
from vllm import _custom_ops as ops
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import FusedMoERouter
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.fused_marlin_moe import fused_marlin_moe
from vllm.model_executor.layers.fused_moe.fused_moe_router import FusedMoERouter
from vllm.model_executor.layers.fused_moe.layer import (
FusedMoE,
FusedMoEMethodBase,
@@ -6,11 +6,11 @@ from typing import Any, Union
import torch
from packaging import version
from vllm.model_executor.layers.fused_moe import FusedMoERouter
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.fused_moe_router import FusedMoERouter
from vllm.model_executor.layers.fused_moe.layer import (
FusedMoE,
FusedMoEMethodBase,
@@ -22,6 +22,7 @@ from vllm.model_executor.layers.fused_moe import (
FusedMoEConfig,
FusedMoEMethodBase,
FusedMoEPermuteExpertsUnpermute,
FusedMoERouter,
FusedMoeWeightScaleSupported,
UnquantizedFusedMoEMethod,
)
@@ -40,7 +41,6 @@ from vllm.model_executor.layers.fused_moe.fused_marlin_moe import (
MarlinExperts,
fused_marlin_moe,
)
from vllm.model_executor.layers.fused_moe.fused_moe_router import FusedMoERouter
from vllm.model_executor.layers.fused_moe.oracle.fp8 import (
Fp8MoeBackend,
convert_to_fp8_moe_kernel_format,
@@ -10,12 +10,12 @@ from vllm.model_executor.layers.fused_moe import (
FusedMoE,
FusedMoEConfig,
FusedMoEMethodBase,
FusedMoERouter,
)
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig,
int8_w8a16_moe_quant_config,
)
from vllm.model_executor.layers.fused_moe.fused_moe_router import FusedMoERouter
from vllm.model_executor.layers.linear import LinearBase, UnquantizedLinearMethod
from vllm.model_executor.layers.quantization import QuantizationMethods
from vllm.model_executor.layers.quantization.base_config import (
@@ -23,13 +23,13 @@ from vllm.model_executor.layers.fused_moe import (
FusedMoEMethodBase,
FusedMoEPermuteExpertsUnpermute,
FusedMoEPrepareAndFinalize,
FusedMoERouter,
FusedMoeWeightScaleSupported,
)
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig,
RoutingMethodType,
)
from vllm.model_executor.layers.fused_moe.fused_moe_router import FusedMoERouter
from vllm.model_executor.layers.fused_moe.layer import UnquantizedFusedMoEMethod
from vllm.model_executor.layers.fused_moe.oracle.fp8 import (
Fp8MoeBackend,
@@ -12,11 +12,11 @@ from torch.nn.parameter import Parameter, UninitializedParameter
from vllm import _custom_ops as ops
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import FusedMoERouter
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.fused_moe_router import FusedMoERouter
from vllm.model_executor.layers.fused_moe.layer import (
FusedMoE,
FusedMoEMethodBase,
@@ -10,12 +10,12 @@ from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
import vllm.model_executor.layers.fused_moe # noqa
from vllm import _custom_ops as ops
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import FusedMoERouter
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.fused_marlin_moe import fused_marlin_moe
from vllm.model_executor.layers.fused_moe.fused_moe_router import FusedMoERouter
from vllm.model_executor.layers.fused_moe.layer import (
FusedMoE,
FusedMoEMethodBase,
@@ -8,10 +8,10 @@ from packaging import version
from torch.nn import Module
from vllm._ipex_ops import ipex_ops as ops
from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
from vllm.model_executor.layers.fused_moe.fused_moe_router import (
from vllm.model_executor.layers.fused_moe import (
FusedMoERouter,
)
from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
from vllm.model_executor.layers.linear import (
LinearBase,
LinearMethodBase,
@@ -13,11 +13,11 @@ import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm._custom_ops import cutlass_scaled_fp4_mm, scaled_fp4_quant
from vllm.attention.layer import Attention
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import FusedMoERouter
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.fused_moe_router import FusedMoERouter
from vllm.model_executor.layers.fused_moe.layer import (
FusedMoE,
FusedMoEMethodBase,
@@ -6,12 +6,12 @@ from typing import Any, Optional
import torch
from vllm.distributed import get_tensor_model_parallel_rank, get_tp_group
from vllm.model_executor.layers.fused_moe import FusedMoERouter
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig,
int4_w4a16_moe_quant_config,
int8_w8a16_moe_quant_config,
)
from vllm.model_executor.layers.fused_moe.fused_moe_router import FusedMoERouter
from vllm.model_executor.layers.fused_moe.layer import (
FusedMoE,
FusedMoEConfig,
@@ -14,6 +14,7 @@ from vllm.model_executor.layers.fused_moe import (
FusedMoE,
FusedMoEConfig,
FusedMoEMethodBase,
FusedMoERouter,
)
from vllm.model_executor.layers.fused_moe import modular_kernel as mk
from vllm.model_executor.layers.fused_moe.config import (
@@ -27,7 +28,6 @@ from vllm.model_executor.layers.fused_moe.fused_marlin_moe import (
MarlinExperts,
fused_marlin_moe,
)
from vllm.model_executor.layers.fused_moe.fused_moe_router import FusedMoERouter
from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import (
OAITritonExperts,
UnfusedOAITritonExperts,
@@ -936,9 +936,9 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
layer.apply_router_weight_on_input,
layer.scoring_func,
layer.activation,
layer.expert_load_view,
layer.logical_to_physical_map,
layer.logical_replica_count,
layer.eplb_state.expert_load_view,
layer.eplb_state.logical_to_physical_map,
layer.eplb_state.logical_replica_count,
), "MXFP4 are not supported with this configuration."
if (
@@ -548,7 +548,7 @@ class QuarkW4A8Fp8MoEMethod(QuarkMoEMethod):
x: torch.Tensor,
router_logits: torch.Tensor,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
topk_weights, topk_ids = layer.select_experts(
topk_weights, topk_ids = router.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -10,12 +10,12 @@ import torch
from torch.nn.parameter import Parameter
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import FusedMoERouter
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.fused_marlin_moe import fused_marlin_moe
from vllm.model_executor.layers.fused_moe.fused_moe_router import FusedMoERouter
from vllm.model_executor.layers.fused_moe.layer import (
FusedMoE,
FusedMoEMethodBase,