Register tensors with symmetric memory for qwen (#18643)

python/sglang/srt/models/qwen2_moe.py

@@ -324,7 +324,11 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
             final_hidden_states = self._forward_router_experts(hidden_states)
 
         if shared_output is not None:
-            final_hidden_states = final_hidden_states + shared_output
+            # In-place add is required to keep final_hidden_states in the
+            # symmetric memory pool (when --enable-symm-mem is used).
+            # An out-of-place add would allocate a new tensor outside symm
+            # memory, breaking subsequent symmetric collective operations.
+            final_hidden_states += shared_output
         if self.tp_size > 1 and not use_reduce_scatter:
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)