[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 = []
@@ -13,7 +13,7 @@ import regex as re
# Matches per-expert weight names like ".experts.42.gate_proj.weight".
# Does NOT match 3D fused-expert names like ".experts.gate_proj.weight"
# (no numeric id) — those are intentionally left unfiltered so the full
# tensor is loaded and sliced later by FusedMoE.weight_loader.
# tensor is loaded and sliced later by RoutedExperts.weight_loader.
_EXPERT_ID_RE = re.compile(r"\.experts\.(\d+)\.")
@@ -13,7 +13,7 @@ import tempfile
import threading
import time
from collections import defaultdict
from collections.abc import Callable, Generator
from collections.abc import Callable, Generator, Iterable
from contextlib import contextmanager
from pathlib import Path
from typing import IO, Any
@@ -1633,3 +1633,108 @@ def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> str | None:
# If there were no matches, return the untouched param name
return name
def maybe_remap_moe_expert_param_name(
name: str,
params_dict: dict[str, torch.nn.Parameter],
) -> str:
"""
Remap MoE expert parameter names to account for routed_experts hierarchy.
This handles the transition from the old FusedMoE structure where weights
were directly in the experts module, to the new MoERunner → RoutedExperts
structure.
Checkpoint weights have names like:
layers.0.mlp.experts.w13_weight
But actual parameters are now:
layers.0.mlp.experts.routed_experts.w13_weight
This function inserts 'routed_experts.' into the path when needed.
Args:
name: Parameter name from checkpoint
params_dict: Dictionary of model parameters (from named_parameters())
Returns:
Remapped parameter name if routed_experts hierarchy exists,
otherwise the original name
"""
# Only remap if this looks like an expert parameter
if ".mlp.experts." not in name:
return name
# Skip if already has routed_experts
if ".mlp.experts.routed_experts." in name:
return name
# Expert parameter patterns to check
expert_param_suffixes = [
"w13_weight",
"w2_weight",
"w13_weight_scale",
"w2_weight_scale",
"w13_input_scale",
"w2_input_scale",
"w13_bias",
"w2_bias",
"w13_scale",
"w2_scale",
"w13_g_idx",
"w2_g_idx",
"w13_qweight",
"w2_qweight",
"w13_qzeros",
"w2_qzeros",
"w13_weight_shape",
"w2_weight_shape",
]
# Check if this is an expert weight parameter
is_expert_param = any(
f".{suffix}" in name or name.endswith(suffix)
for suffix in expert_param_suffixes
)
if not is_expert_param:
return name
# Try inserting routed_experts
new_name = name.replace(".mlp.experts.", ".mlp.experts.routed_experts.")
# Only use the new name if it exists in the model
if new_name in params_dict:
return new_name
# Otherwise return original name (old checkpoint format or different structure)
return name
def remap_moe_expert_weights(
weights: Iterable[tuple[str, torch.Tensor]],
params_dict: dict[str, torch.nn.Parameter],
) -> Generator[tuple[str, torch.Tensor], None, None]:
"""
Wrapper generator that remaps MoE expert parameter names for backward compatibility.
This allows models with custom weight loading to automatically handle both old
and new checkpoint formats without needing model-specific remapping code.
Usage:
params_dict = dict(model.named_parameters())
for name, weight in remap_moe_expert_weights(weights, params_dict):
# name is automatically remapped if needed
param = params_dict[name]
...
Args:
weights: Iterator of (name, tensor) tuples from checkpoint
params_dict: Dictionary of model parameters (from named_parameters())
Yields:
(remapped_name, tensor) tuples
"""
for name, weight in weights:
remapped_name = maybe_remap_moe_expert_param_name(name, params_dict)
yield (remapped_name, weight)