[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
-1
View File
@@ -213,7 +213,6 @@ class ArcticMoE(nn.Module):
self.w2s,
topk_weights,
topk_ids,
inplace=True,
)
if self.reduce_results and self.tp_size > 1:
final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
+1 -1
View File
@@ -326,7 +326,7 @@ class NomicMoE(nn.Module):
hidden_states, router_logits, self.top_k, renormalize=False
)
final_hidden_states = torch.ops.vllm.outplace_fused_experts(
final_hidden_states = torch.ops.vllm.fused_experts(
hidden_states=hidden_states,
w1=self.w1,
w2=self.w2,
+1 -1
View File
@@ -181,7 +181,7 @@ class MiniCPMMoE(nn.Module):
)
final_hidden_states = fused_experts(
hidden_states, self.ws, self.w2s, topk_weights, topk_ids, inplace=False
hidden_states, self.ws, self.w2s, topk_weights, topk_ids
)
if self.tp_size > 1: