[MoE] Remove inplace fused experts mechanism (#43727)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
This commit is contained in:
Yongye Zhu
2026-05-27 20:00:19 -07:00
committed by GitHub
parent 05eec7120e
commit 2d2c660104
44 changed files with 11 additions and 237 deletions
@@ -497,7 +497,6 @@ class BitsAndBytesMoEMethod(FusedMoEMethodBase):
w2=w2,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=not self.moe.disable_inplace,
activation=layer.activation,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
global_num_experts=layer.global_num_experts,
@@ -256,7 +256,6 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
layer.w2_weight_packed,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=not self.moe.disable_inplace,
activation=layer.activation,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
global_num_experts=layer.global_num_experts,
@@ -573,5 +573,4 @@ class CompressedTensorsWNA16MarlinMoEMethod(CompressedTensorsMoEMethod):
workspace=layer.workspace,
input_dtype=self.marlin_input_dtype,
is_k_full=self.is_k_full,
inplace=not self.moe.disable_inplace,
)
@@ -377,7 +377,6 @@ class MoeWNA16Method(FusedMoEMethodBase):
layer.w2_qweight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=not self.moe.disable_inplace,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,
@@ -497,7 +497,6 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
apply_router_weight_on_input=layer.apply_router_weight_on_input,
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,
inplace=not self.moe.disable_inplace,
)
else:
from vllm.model_executor.layers.fused_moe import fused_experts
@@ -508,7 +507,6 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
w2=layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=not self.moe.disable_inplace,
activation=layer.activation,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
global_num_experts=layer.global_num_experts,
@@ -809,7 +807,6 @@ class QuarkW8A8Int8MoEMethod(QuarkMoEMethod):
w2=layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=not self.moe.disable_inplace,
activation=layer.activation,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
global_num_experts=layer.global_num_experts,