[Bugfix] Avoid unnecessary reduce-scatter call in prepare_mlp (#9169)

python/sglang/srt/layers/dp_attention.py

@@ -292,6 +292,10 @@ def _dp_gather_via_all_gather(
     forward_batch: ForwardBatch,
     is_partial: bool,
 ):
+    if get_attention_tp_size() == 1:
+        get_tp_group().all_gather_into_tensor(global_tokens, local_tokens)
+        return
+
     if not is_partial:
         if get_attention_tp_rank() != 0:
             local_tokens.fill_(0)