From e74331a1ede31c69ec0b1b97bd407d38742caa9c Mon Sep 17 00:00:00 2001
From: Pleaplusone <38376071+ganyi1996ppo@users.noreply.github.com>
Date: Wed, 23 Apr 2025 15:47:51 +0800
Subject: [PATCH] Add dp initialize patch with hccl backend (#626)

<!--  Thanks for sending a pull request!

BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html

-->
### What this PR does / why we need it?
<!--
- Please clarify what changes you are proposing. The purpose of this
section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR.

- Please clarify why the changes are needed. For instance, the use case
and bug description.

- Fixes #
-->
Add dp stateless process group initialization path with hccl backend as
vllm-ascend patch.
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->

### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->

---------

Signed-off-by: ganyi <pleaplusone.gy@gmail.com>
---
 vllm_ascend/patch/__init__.py                     | 10 +++++++++-
 .../platform/patch_common/patch_distributed.py    | 15 +++++++++++++++
 2 files changed, 24 insertions(+), 1 deletion(-)

diff --git a/vllm_ascend/patch/__init__.py b/vllm_ascend/patch/__init__.py
index d4286bd..93c0b0b 100644
--- a/vllm_ascend/patch/__init__.py
+++ b/vllm_ascend/patch/__init__.py
@@ -87,7 +87,15 @@
 #    Future Plan:
 #       Its a workaround in vllm-ascend to enable multi-node dp inference, maybe removed if vllm have better plan
 #       on multi-node dp inference implementation
-#
+#   4. `ParallelConfig.stateless_init_dp_group`
+#    Why:
+#       vLLM use gloo backend by default to initialize stateless dp process gourp, but we want to use hccl here to
+#       get better performance
+#    How：
+#       adopt nccl backend to init process group
+#    Related PR (if no, explain why): no related PR, we want add this ability into vllm
+#    Future Plan:
+#       Remove those patch when vllm merged them
 # * Worker Patch:
 # ===============
 # ** File: worker/patch_0_8_4/patch_metrics.py **
diff --git a/vllm_ascend/patch/platform/patch_common/patch_distributed.py b/vllm_ascend/patch/platform/patch_common/patch_distributed.py
index 1b356a9..ce43836 100644
--- a/vllm_ascend/patch/platform/patch_common/patch_distributed.py
+++ b/vllm_ascend/patch/platform/patch_common/patch_distributed.py
@@ -152,6 +152,21 @@ def parallel_config_get_dp_port(self) -> int:
     return port
 
 
+def ascend_stateless_init_dp_group(self) -> "ProcessGroup":
+    from vllm.distributed.utils import \
+        stateless_init_torch_distributed_process_group
+
+    dp_group = stateless_init_torch_distributed_process_group(
+        self.data_parallel_master_ip,
+        self.get_next_dp_init_port(),
+        self.data_parallel_rank,
+        self.data_parallel_size,
+        backend="hccl")
+
+    return dp_group
+
+
 vllm.distributed.parallel_state.destroy_model_parallel = ascend_destroy_model_parallel
 vllm.distributed.stateless_init_torch_distributed_process_group = ascend_stateless_init_torch_distributed_process_group
 ParallelConfig.get_next_dp_init_port = parallel_config_get_dp_port
+ParallelConfig.stateless_init_dp_group = ascend_stateless_init_dp_group