[MoE Refactor] FusedMoE/MoERunner inversion refactor (#41184)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-06-08 10:42:58 -04:00
committed by GitHub
parent 753e9d55e6
commit dc68bd8c41
90 changed files with 2738 additions and 2031 deletions
@@ -22,7 +22,7 @@ from vllm.distributed import (
)
from vllm.logger import init_logger
from vllm.lora.utils import is_moe_model
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.fused_moe import RoutedExperts
from vllm.model_executor.layers.linear import (
LinearBase,
MergedColumnParallelLinear,
@@ -464,13 +464,13 @@ class BitsAndBytesModelLoader(BaseModelLoader):
self.target_modules.append(name)
if module.disable_tp:
self.tp_disabled_modules.append(name)
elif isinstance(module, FusedMoE) and hasattr(
elif isinstance(module, RoutedExperts) and hasattr(
module.quant_method, "quant_config"
):
# TODO: support FusedMoE with prequant and 8bit.
if self.pre_quant and self.load_8bit:
raise ValueError(
"Prequant BitsAndBytes 8bit models with FusedMoE "
"Prequant BitsAndBytes 8bit models with RoutedExperts "
"is not supported yet."
)
# Get the corresponding weight name using module name and
@@ -508,7 +508,7 @@ class BitsAndBytesModelLoader(BaseModelLoader):
# dimension (dim=-1)
elif isinstance(module, (RowParallelLinear,)):
self.column_sharded_weights_modules.append(name)
elif isinstance(module, FusedMoE):
elif isinstance(module, RoutedExperts):
expert_mapping = self.expert_params_mapping
for exp in expert_mapping:
if exp[-1] == "w2":
@@ -629,7 +629,7 @@ class BitsAndBytesModelLoader(BaseModelLoader):
expert_mapping = self.expert_params_mapping
expert_qs_dict = {}
for name, module in model.named_modules():
if not isinstance(module, FusedMoE):
if not isinstance(module, RoutedExperts):
continue
w1_states_lst = []
w2_states_lst = []