[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
@@ -21,7 +21,7 @@ from vllm.utils.flashinfer import (
)
if TYPE_CHECKING:
from vllm.model_executor.layers.fused_moe.layer import FusedMoE
from vllm.model_executor.layers.fused_moe import RoutedExperts
from vllm.model_executor.layers.fused_moe.oracle.nvfp4 import (
NvFp4MoeBackend,
)
@@ -80,7 +80,7 @@ def interleave_linear_and_gate(
def prepare_nvfp4_moe_layer_for_flashinfer_cutedsl(
layer: "FusedMoE",
layer: "RoutedExperts",
w13: torch.Tensor,
w13_scale: torch.Tensor,
w13_scale_2: torch.Tensor,
@@ -286,7 +286,7 @@ def prepare_static_weights_for_trtllm_fp4_moe(
def prepare_nvfp4_moe_layer_for_fi_or_cutlass(
backend: "NvFp4MoeBackend",
layer: "FusedMoE",
layer: "RoutedExperts",
w13: torch.Tensor,
w13_scale: torch.Tensor,
w13_scale_2: torch.Tensor,
@@ -12,7 +12,7 @@ from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig,
FusedMoEQuantDesc,
)
from vllm.model_executor.layers.fused_moe.layer import FusedMoE
from vllm.model_executor.layers.fused_moe.routed_experts import RoutedExperts
from vllm.model_executor.layers.linear import LinearBase
from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
@@ -82,7 +82,7 @@ def prepare_humming_layer(layer: LinearBase, quant_config: dict):
HummingMethod.transform_humming_layer(layer)
def prepare_humming_moe_layer(layer: FusedMoE, quant_config: dict):
def prepare_humming_moe_layer(layer: RoutedExperts, quant_config: dict):
weight_schema = BaseWeightSchema.from_config(quant_config)
input_quant_config = envs.VLLM_HUMMING_INPUT_QUANT_CONFIG or {}
if humming_is_layer_skipped(input_quant_config, layer.layer_name):
@@ -164,7 +164,7 @@ def prepare_humming_moe_layer(layer: FusedMoE, quant_config: dict):
layer.register_buffer("locks", locks)
def get_humming_moe_quant_config(layer: FusedMoE):
def get_humming_moe_quant_config(layer: RoutedExperts):
input_schema = layer.input_schemas["w13"]
weight_schema = layer.weight_schemas["w13"]
@@ -8,6 +8,7 @@ import torch
import vllm.envs as envs
from vllm import _custom_ops as ops
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import RoutedExperts
from vllm.model_executor.layers.linear import LinearBase
from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
from vllm.model_executor.layers.quantization.utils.int8_utils import (
@@ -226,7 +227,7 @@ def check_marlin_supports_layer(layer: LinearBase, group_size: int) -> bool:
)[0]
def check_moe_marlin_supports_layer(layer: LinearBase, group_size: int) -> bool:
def check_moe_marlin_supports_layer(layer: RoutedExperts, group_size: int) -> bool:
if current_platform.is_rocm():
return False
hidden_size = layer.hidden_size