[MoE Refactor] Rename FusedMoE to FusedMoEFactory (#44941)

Signed-off-by: Bill Nell <bnell@redhat.com>
Signed-off-by: bnellnm <49004751+bnellnm@users.noreply.github.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
bnellnm
2026-07-31 11:28:18 +01:00
committed by GitHub
co-authored by Harry Mellor
parent c911120a6f
commit 6e311c6e20
80 changed files with 243 additions and 234 deletions
+12 -7
View File
@@ -42,7 +42,11 @@ from vllm.distributed.device_communicators.all_reduce_utils import (
from vllm.distributed.eplb.eplb_communicator import create_eplb_communicator
from vllm.distributed.eplb.rebalance_execute import rearrange_expert_weights_inplace
from vllm.forward_context import set_forward_context
from vllm.model_executor.layers.fused_moe import FusedMoE, MoERunner, fused_experts
from vllm.model_executor.layers.fused_moe import (
FusedMoEFactory,
MoERunner,
fused_experts,
)
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
from vllm.model_executor.layers.fused_moe.router.router_factory import (
@@ -80,7 +84,7 @@ NUM_EXPERTS = [8, 64]
TOP_KS = [2, 6]
# dp_size, tp_size, use_ep
# Note: DP+TP is not yet supported in the FusedMoE layer.
# Note: DP+TP is not yet supported in the FusedMoEFactory layer.
PARALLEL_COMBOS = [
[1, 2, False],
[1, 4, False],
@@ -1023,7 +1027,7 @@ def make_fused_moe_layer(
kwargs["routed_input_transform"] = routed_input_transform
kwargs["routed_output_transform"] = routed_output_transform
layer = FusedMoE(
layer = FusedMoEFactory(
num_experts=global_num_experts,
top_k=top_k,
hidden_size=hidden_size,
@@ -1267,7 +1271,7 @@ def _test_body_eplb(
):
output_before = sp_wrapper(moe_layer)(hidden_states, router_logits)
# Create a fresh FusedMoE layer with enable_eplb=True
# Create a fresh MoERunner layer with enable_eplb=True
# Delete the original layer's registration so the constructor can
# re-use the same "from_forward_context" prefix
cc = vllm_config.compilation_config
@@ -1401,7 +1405,8 @@ def _run_one_config(
- When is_sequence_parallel=True (EP + sequence splitting):
* ep_size: Number of expert parallel ranks (equals dp_size * tp_size)
* tp_size: Number of ranks to split sequence across (becomes sp_size in FusedMoE)
* tp_size: Number of ranks to split sequence across (becomes sp_size in
MoERunner)
* Weights are chunked by ep_size (experts) but NOT by tp_size
* Input sequences are chunked by tp_size (via sp_wrapper)
"""
@@ -1476,8 +1481,8 @@ def _run_one_config(
torch.accelerator.empty_cache()
with set_current_vllm_config(vllm_config):
# Chunk weights for EP BEFORE creating FusedMoE
# FusedMoE uses EP-chunked weights and handles reductions internally
# Chunk weights for EP BEFORE creating MoERunner.
# MoERunner uses EP-chunked weights and handles reductions internally.
if ep_size > 1:
# Split experts across ranks (dimension 0 is the expert dimension)
# When EP is enabled, use EP group rank and ep_size for chunking