A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Model] Removes redundant all-reduce operation in Qwen3MoeSparseMoeBlock (#23169)
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
yiz-liu committed
Aug 20, 2025 at 00:18 UTC
4f510bc2a175a6eebfdceafa7013da8a3865eb38
1 file changed
+1
-5
vllm/model_executor/models/qwen3_moe.py
+1
-5
@@ -139,7 +139,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
139
top_k=config.num_experts_per_tok,
140
hidden_size=config.hidden_size,
141
intermediate_size=config.moe_intermediate_size,
142
- reduce_results=False,
142
+ reduce_results=True,
143
renormalize=config.norm_topk_prob,
144
quant_config=quant_config,
145
prefix=f"{prefix}.experts",
@@ -163,10 +163,6 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
163
final_hidden_states = self.experts(hidden_states=hidden_states,
164
router_logits=router_logits)
165
166
- if self.tp_size > 1:
167
- final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel( # noqa E501
168
- final_hidden_states)
169
-
166
return final_hidden_states.view(orig_shape)
167
168