[Bugfix] Fix memory-leak caused by dist._functional_collectives.reduce_scatter_tensor (#1380)

### What this PR does / why we need it?
In some cases, `dist._functional_collectives.reduce_scatter_tensor` can
cause its input tensor not to be released immediately after the current
layer ends. Instead, it will only be released when the GPU memory usage
of the current process reaches a certain threshold (approximately every
15 layers each time).

**Before Fix**

<img width="1441" alt="截屏2025-06-24 01 26 13"
src="https://github.com/user-attachments/assets/72d5dbb3-c8c8-4778-bf64-8db7bab8aff0"
/>

**After Fix**
<img width="1475" alt="截屏2025-06-24 01 23 43"
src="https://github.com/user-attachments/assets/6c69cfcd-a469-4ee5-b8c6-210aeb3a5bdf"
/>

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?


- vLLM version: v0.9.1
- vLLM main:
vllm-project/vllm@9ff2af6

---------

Signed-off-by: ApsarasX <apsarax@outlook.com>

Author: ApsarasX

vllm_ascend/distributed/communication_op.py

@@ -0,0 +1,25 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+#
+
+import torch
+from vllm.distributed.parallel_state import get_dp_group
+
+
+def data_parallel_reduce_scatter(input_: torch.Tensor,
+                                 dim: int = -1) -> torch.Tensor:
+    """Reduce-Scatter the input tensor across data parallel group."""
+    return get_dp_group().reduce_scatter(input_, dim)

vllm_ascend/ops/fused_moe.py

@@ -39,6 +39,8 @@
 
 import vllm_ascend.envs as envs_ascend
 from vllm_ascend.ascend_config import get_ascend_config
+from vllm_ascend.distributed.communication_op import \
+    data_parallel_reduce_scatter
 from vllm_ascend.distributed.parallel_state import get_ep_group, get_etp_group
 from vllm_ascend.ops.expert_load_balancer import ExpertLoadBalancer
 from vllm_ascend.utils import (FusedMoEState, dispose_tensor,
@@ -1342,11 +1344,8 @@ def forward(self,
                 final_hidden_states = final_hidden_states[start:end, :]
                 dispose_tensor(e_hidden_states)
             elif fused_moe_state == FusedMoEState.AllGather:
-                final_hidden_states = dist._functional_collectives.reduce_scatter_tensor(
-                    e_hidden_states,
-                    "sum",
-                    scatter_dim=0,
-                    group=get_dp_group().device_group)
+                final_hidden_states = data_parallel_reduce_scatter(
+                    e_hidden_states, dim=0)
                 final_hidden_states = final_hidden_states[:num_tokens]
                 dispose_tensor(e_hidden_states)
         else: