[DP][V1] Fix rank set in DP scenario & Bump torch-npu version to 2.5.1.post1.dev20250528 (vllm-project#1235)

1. Fix rank set in DP scenario. The new poc version of torch-npu support
setting `ASCEND_RT_VISIBLE_DEVICES` dynamically, thus we could use the
rank set in `DPEngineCoreProc` directly instead of calculating local
rank across dp by hand in the patched `_init_data_parallel`

Closes: vllm-project#1170

2. Bump torch-npu version to 2.5.1.post1.dev20250528

Closes: vllm-project#1242
Closes: vllm-project#1232

CI passed with new added test.

---------

Signed-off-by: MengqingCao <cmq0113@163.com>
Signed-off-by: Icey <1790571317@qq.com>
Co-authored-by: Icey <1790571317@qq.com>

Author: 2 people

pyproject.toml

@@ -12,7 +12,7 @@ requires = [
     "scipy",
     "setuptools>=64",
     "setuptools-scm>=8",
-    "torch-npu==2.5.1",
+    "torch-npu==2.5.1.post1.dev20250528",
     "torch>=2.5.1",
     "torchvision<0.21.0",
     "wheel",

tests/multicard/test_data_parallel.py

@@ -0,0 +1,66 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# Copyright 2023 The vLLM team.
+#
+# 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.
+#
+"""
+Compare the outputs of vLLM with and without aclgraph.
+
+Run `pytest tests/multicard/test_data_parallel.py`.
+"""
+
+import os
+
+import pytest
+
+from tests.conftest import VllmRunner
+from tests.model_utils import check_outputs_equal
+
+MODELS = ["Qwen/Qwen2.5-0.5B-Instruct"]
+
+
+@pytest.mark.skipif(os.getenv("VLLM_USE_V1") == "0",
+                    reason="Data parallel only support on v1")
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("max_tokens", [32])
+def test_data_parallel_correctness(
+    model: str,
+    max_tokens: int,
+) -> None:
+    example_prompts = [
+        "Hello, my name is", "The president of the United States is",
+        "The capital of France is", "The future of AI is"
+    ]
+
+    with VllmRunner(model_name=model,
+                    max_model_len=1024,
+                    max_num_seqs=16,
+                    data_parallel_size=2,
+                    distributed_executor_backend="mp") as vllm_model:
+        vllm_dp_outputs = vllm_model.generate_greedy(example_prompts,
+                                                     max_tokens)
+
+    with VllmRunner(
+            model_name=model,
+            max_model_len=1024,
+            max_num_seqs=16,
+    ) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=vllm_outputs,
+        outputs_1_lst=vllm_dp_outputs,
+        name_0="vllm_outputs",
+        name_1="vllm_dp_outputs",
+    )

vllm_ascend/patch/__init__.py

@@ -47,16 +47,7 @@
 #    Related PR (if no, explain why):
 #    Future Plan:
 #       Remove those patch when vllm merged them
-#   2. `vllm.v1.engine.core.DPEngineCoreProc._init_data_parallel`
-#    Why:
-#       There is some bug for ASCEND_RT_VISIBLE_DEVICES usage.
-#    How：
-#       The ASCEND_RT_VISIBLE_DEVICES related code is dropped.
-#    Related PR (if no, explain why):
-#       No, this is a bug for vllm ascend
-#    Future Plan:
-#       Remove this patch once ASCEND_RT_VISIBLE_DEVICES bug is fixed.
-#   3. `vllm.config.ParallelConfig.get_next_dp_init_port`
+#   2. `vllm.config.ParallelConfig.get_next_dp_init_port`
 #    Why:
 #       vllm doesn't support get port from environment.
 #    How：
@@ -65,7 +56,7 @@
 #       Need a PR to vllm to support get port from environment.
 #    Future Plan:
 #       Remove those patch when vllm merged them
-#   4. `vllm.config.ParallelConfig.ParallelConfig.stateless_init_dp_group`
+#   3. `vllm.config.ParallelConfig.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