[MoE Refactor] Remove SharedFusedMoE class (#35782)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-04-21 18:12:12 -04:00
committed by GitHub
parent 1842447c09
commit 5e584ce9ec
33 changed files with 112 additions and 141 deletions
+3 -7
View File
@@ -37,7 +37,7 @@ from vllm.distributed.parallel_state import (
get_eplb_group,
)
from vllm.forward_context import set_forward_context
from vllm.model_executor.layers.fused_moe import FusedMoE, SharedFusedMoE, fused_experts
from vllm.model_executor.layers.fused_moe import FusedMoE, 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 (
@@ -858,11 +858,7 @@ def make_fused_moe_layer(
quant_config, qw = make_quant_config(quantization, w1, w2, global_num_experts)
kwargs = dict()
if shared_experts is None:
builder = FusedMoE
else:
builder = SharedFusedMoE
kwargs["shared_experts"] = shared_experts
kwargs["shared_experts"] = shared_experts
# Add gate and routed_input_transform if provided
if gate is not None:
@@ -872,7 +868,7 @@ def make_fused_moe_layer(
kwargs["routed_input_transform"] = routed_input_transform
kwargs["routed_output_transform"] = routed_output_transform
layer = builder(
layer = FusedMoE(
num_experts=global_num_experts,
top_k=top_k,
hidden_size=hidden_size,
@@ -1,9 +1,9 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Tests for SharedFusedMoE with routed_input_transform.
Tests for FusedMoE with routed_input_transform.
Verifies that applying routed_input_transform inside SharedFusedMoE
Verifies that applying routed_input_transform inside FusedMoE
produces the same results as applying the transform manually outside.
"""
@@ -13,7 +13,7 @@ import torch.nn as nn
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.forward_context import set_forward_context
from vllm.model_executor.layers.fused_moe.shared_fused_moe import SharedFusedMoE
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.platforms import current_platform
from vllm.utils.torch_utils import is_torch_equal_or_newer, set_random_seed
@@ -133,9 +133,9 @@ def test_routed_input_transform_inside_vs_outside(
workspace_init,
monkeypatch,
):
"""Compare SharedFusedMoE with transform inside vs manually applying outside.
Method A (inside): SharedFusedMoE with routed_input_transform
Method B (outside): Manually transform, then SharedFusedMoE without transform
"""Compare FusedMoE with transform inside vs manually applying outside.
Method A (inside): FusedMoE with routed_input_transform
Method B (outside): Manually transform, then FusedMoE without transform
"""
if current_platform.is_rocm():
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1" if use_rocm_aiter else "0")
@@ -157,8 +157,8 @@ def test_routed_input_transform_inside_vs_outside(
routed_transform = SimpleLinear(hidden_size, latent_size, dtype)
with set_current_vllm_config(vllm_config):
# Method A: SharedFusedMoE WITH routed_input_transform
moe_with_transform = SharedFusedMoE(
# Method A: FusedMoE WITH routed_input_transform
moe_with_transform = FusedMoE(
shared_experts=shared_experts,
routed_input_transform=routed_transform,
num_experts=num_experts,
@@ -173,9 +173,9 @@ def test_routed_input_transform_inside_vs_outside(
prefix="moe_with_transform",
)
# Method B: SharedFusedMoE WITHOUT routed_input_transform
# Method B: FusedMoE WITHOUT routed_input_transform
# Note: shared_experts=None because when transform is done outside,
moe_without_transform = SharedFusedMoE(
moe_without_transform = FusedMoE(
shared_experts=None,
routed_input_transform=None,
num_experts=num_experts,