[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
@@ -9,6 +9,7 @@ import torch
from torch.nn.parameter import Parameter
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import RoutedExperts
from vllm.model_executor.layers.linear import (
LinearBase,
LinearMethodBase,
@@ -234,7 +235,6 @@ class INCConfig(QuantizationConfig):
self.extra_config = hf_to_vllm_mapper.apply_dict(self.extra_config)
def apply_awq_quant_layer(self, layer, prefix: str, backend: str = "auto"):
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.quantization.utils.marlin_utils import (
check_marlin_supported,
check_moe_marlin_supports_layer,
@@ -264,7 +264,7 @@ class INCConfig(QuantizationConfig):
AWQ_TYPE_MAP[weight_bits], group_size, not sym
)
if isinstance(layer, FusedMoE):
if isinstance(layer, RoutedExperts):
use_marlin = use_marlin and check_moe_marlin_supports_layer(
layer, group_size
)
@@ -298,7 +298,7 @@ class INCConfig(QuantizationConfig):
zero_point=not sym,
)
if isinstance(layer, FusedMoE):
if isinstance(layer, RoutedExperts):
if use_marlin:
return AWQMarlinMoEMethod(quant_args_marlin, layer.moe_config)
from vllm.model_executor.layers.quantization.moe_wna16 import MoeWNA16Config
@@ -320,7 +320,6 @@ class INCConfig(QuantizationConfig):
return None
def apply_gptq_quant_layer(self, layer, prefix: str, backend: str = "auto"):
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.quantization.utils.marlin_utils import (
check_marlin_supported,
check_moe_marlin_supports_layer,
@@ -349,7 +348,7 @@ class INCConfig(QuantizationConfig):
use_marlin = (weight_bits, sym) in GPTQ_TYPE_MAP and check_marlin_supported(
GPTQ_TYPE_MAP[(weight_bits, sym)], group_size, has_zp=not sym
)
if isinstance(layer, FusedMoE):
if isinstance(layer, RoutedExperts):
use_marlin = use_marlin and check_moe_marlin_supports_layer(
layer, group_size
)
@@ -385,7 +384,7 @@ class INCConfig(QuantizationConfig):
dynamic={},
)
if isinstance(layer, FusedMoE):
if isinstance(layer, RoutedExperts):
if use_marlin:
return GPTQMarlinMoEMethod(quant_args_marlin, layer.moe_config)
else: