[Model] Removes redundant all-reduce operation in Qwen3MoeSparseMoeBlock (#23169)

Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>

vllm/model_executor/models/qwen3_moe.py

@@ -139,7 +139,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
                                 top_k=config.num_experts_per_tok,
                                 hidden_size=config.hidden_size,
                                 intermediate_size=config.moe_intermediate_size,
-                                reduce_results=False,
+                                reduce_results=True,
                                 renormalize=config.norm_topk_prob,
                                 quant_config=quant_config,
                                 prefix=f"{prefix}.experts",
@@ -163,10 +163,6 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
         final_hidden_states = self.experts(hidden_states=hidden_states,
                                            router_logits=router_logits)
 
-        if self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(  # noqa E501
-                final_hidden_states)
-
         return final_hidden_states.view(orig_shape)