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

This reverts commit 96fa7ff.

Author: Yikun

.github/workflows/accuracy_test.yaml

@@ -173,8 +173,6 @@ jobs:
 
       - name: Install vllm-project/vllm-ascend
         working-directory: ./vllm-ascend
-        env:
-          PIP_EXTRA_INDEX_URL: https://mirrors.huaweicloud.com/ascend/repos/pypi
         run: |
           pip install -r requirements-dev.txt
           pip install -e .

.github/workflows/image_openeuler.yml

@@ -19,12 +19,6 @@ on:
       - '.github/workflows/image_openeuler.yml'
       - 'Dockerfile.openEuler'
       - 'vllm_ascend/**'
-      - 'setup.py'
-      - 'pyproject.toml'
-      - 'requirements.txt'
-      - 'cmake/**'
-      - 'CMakeLists.txt'
-      - 'csrc/**'
   push:
     # Publish image when tagging, the Dockerfile in tag will be build as tag image
     branches:

.github/workflows/image_ubuntu.yml

@@ -19,12 +19,6 @@ on:
       - '.github/workflows/image_ubuntu.yml'
       - 'Dockerfile'
       - 'vllm_ascend/**'
-      - 'setup.py'
-      - 'pyproject.toml'
-      - 'requirements.txt'
-      - 'cmake/**'
-      - 'CMakeLists.txt'
-      - 'csrc/**'
   push:
     # Publish image when tagging, the Dockerfile in tag will be build as tag image
     branches:

.github/workflows/nightly_benchmarks.yaml

@@ -115,8 +115,6 @@ jobs:
           VLLM_TARGET_DEVICE=empty pip install -e .
 
       - name: Install vllm-project/vllm-ascend
-        env:
-          PIP_EXTRA_INDEX_URL: https://mirrors.huaweicloud.com/ascend/repos/pypi
         run: |
           pip install -e .
           pip install -r benchmarks/requirements-bench.txt

.github/workflows/vllm_ascend_test.yaml

@@ -165,7 +165,6 @@ jobs:
 
       - name: Install vllm-project/vllm-ascend
         run: |
-          export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi
           export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/x86_64-linux/devlib
           python3 -m pip install -r requirements-dev.txt --extra-index https://download.pytorch.org/whl/cpu/
           python3 -m pip install -v . --extra-index https://download.pytorch.org/whl/cpu/
@@ -239,8 +238,6 @@ jobs:
           VLLM_TARGET_DEVICE=empty pip install -e .
 
       - name: Install vllm-project/vllm-ascend
-        env:
-          PIP_EXTRA_INDEX_URL: https://mirrors.huaweicloud.com/ascend/repos/pypi
         run: |
           pip install -r requirements-dev.txt
           pip install -v -e .
@@ -338,8 +335,6 @@ jobs:
           VLLM_TARGET_DEVICE=empty pip install -e .
 
       - name: Install vllm-project/vllm-ascend
-        env:
-          PIP_EXTRA_INDEX_URL: https://mirrors.huaweicloud.com/ascend/repos/pypi
         run: |
           pip install -r requirements-dev.txt
           pip install -v -e .

.github/workflows/vllm_ascend_test_long_term.yaml

@@ -88,8 +88,6 @@ jobs:
           VLLM_TARGET_DEVICE=empty pip install -e .
 
       - name: Install vllm-project/vllm-ascend
-        env:
-          PIP_EXTRA_INDEX_URL: https://mirrors.huaweicloud.com/ascend/repos/pypi
         run: |
           pip install -r requirements-dev.txt
           pip install -v -e .

.github/workflows/vllm_ascend_test_pd.yaml

@@ -97,8 +97,6 @@ jobs:
           VLLM_TARGET_DEVICE=empty pip install -e .
 
       - name: Install vllm-project/vllm-ascend
-        env:
-          PIP_EXTRA_INDEX_URL: https://mirrors.huaweicloud.com/ascend/repos/pypi
         run: |
           pip install -r requirements-dev.txt
           pip install -v -e .

Dockerfile

@@ -46,8 +46,7 @@ RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm
 
 # Install vllm-ascend
 # Append `libascend_hal.so` path (devlib) to LD_LIBRARY_PATH
-RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi && \
-    source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
+RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
     source /usr/local/Ascend/nnal/atb/set_env.sh && \
     export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/`uname -i`-linux/devlib && \
     python3 -m pip install -v -e /vllm-workspace/vllm-ascend/ --extra-index https://download.pytorch.org/whl/cpu/ && \

Dockerfile.openEuler

@@ -43,8 +43,7 @@ RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/ -
     python3 -m pip cache purge
 
 # Install vllm-ascend
-RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi && \
-    source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
+RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
     source /usr/local/Ascend/nnal/atb/set_env.sh && \
     export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/`uname -i`-linux/devlib && \
     python3 -m pip install -v -e /vllm-workspace/vllm-ascend/ --extra-index https://download.pytorch.org/whl/cpu/ && \

README.md

@@ -38,7 +38,7 @@ By using vLLM Ascend plugin, popular open-source models, including Transformer-l
 - Software:
   * Python >= 3.9, < 3.12
   * CANN >= 8.1.RC1
-  * PyTorch >= 2.5.1, torch-npu >= 2.5.1.post1.dev20250528
+  * PyTorch >= 2.5.1, torch-npu >= 2.5.1
   * vLLM (the same version as vllm-ascend)
 
 ## Getting Started

README.zh.md

@@ -39,7 +39,7 @@ vLLM 昇腾插件 (`vllm-ascend`) 是一个由社区维护的让vLLM在Ascend NP
 - 软件：
   * Python >= 3.9, < 3.12
   * CANN >= 8.1.RC1
-  * PyTorch >= 2.5.1, torch-npu >= 2.5.1.post1.dev20250528
+  * PyTorch >= 2.5.1, torch-npu >= 2.5.1
   * vLLM (与vllm-ascend版本一致)
 
 ## 开始使用

docs/source/installation.md

@@ -9,11 +9,11 @@ This document describes how to install vllm-ascend manually.
 - A hardware with Ascend NPU. It's usually the Atlas 800 A2 series.
 - Software:
 
-    | Software      | Supported version                | Note                                      |
-    |---------------|----------------------------------|-------------------------------------------|
-    | CANN          | >= 8.1.RC1                       | Required for vllm-ascend and torch-npu    |
-    | torch-npu     | >= 2.5.1.post1.dev20250528       | Required for vllm-ascend                  |
-    | torch         | >= 2.5.1                         | Required for torch-npu and vllm           |
+    | Software  | Supported version | Note                                   |
+    |-----------|-------------------|----------------------------------------| 
+    | CANN      | >= 8.1.RC1        | Required for vllm-ascend and torch-npu |
+    | torch-npu | >= 2.5.1          | Required for vllm-ascend               |
+    | torch     | >= 2.5.1          | Required for torch-npu and vllm        |
 
 You have 2 way to install:
 - **Using pip**: first prepare env manually or via CANN image, then install `vllm-ascend` using pip.
@@ -156,7 +156,6 @@ cd ..
 # Install vLLM Ascend
 git clone  --depth 1 --branch |vllm_ascend_version| https://github.com/vllm-project/vllm-ascend.git
 cd vllm-ascend
-export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi
 pip install -v -e .
 cd ..
 ```

pyproject.toml

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

requirements.txt

@@ -10,6 +10,7 @@ pyyaml
 scipy
 setuptools>=64
 setuptools-scm>=8
+torch-npu==2.5.1
 torch>=2.5.1
 torchvision<0.21.0
 wheel
@@ -20,8 +21,3 @@ quart
 
 # Required for N-gram speculative decoding
 numba
-
-# Install torch_npu
---pre
---extra-index-url https://mirrors.huaweicloud.com/ascend/repos/pypi
-torch-npu==2.5.1.post1.dev20250528

setup.py

@@ -152,7 +152,7 @@ def configure(self, ext: CMakeExtension) -> None:
             # if pybind11 is installed via pip
             pybind11_cmake_path = (subprocess.check_output(
                 [python_executable, "-m", "pybind11",
-                 "--cmakedir"]).decode().strip())
+                 "--cmake"]).decode().strip())
         except subprocess.CalledProcessError as e:
             # else specify pybind11 path installed from source code on CI container
             raise RuntimeError(f"CMake configuration failed: {e}")

tests/multicard/test_data_parallel.py

@@ -47,7 +47,16 @@
 #    Related PR (if no, explain why):
 #    Future Plan:
 #       Remove those patch when vllm merged them
-#   2. `vllm.config.ParallelConfig.get_next_dp_init_port`
+#   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`
 #    Why:
 #       vllm doesn't support get port from environment.
 #    How：
@@ -56,7 +65,7 @@
 #       Need a PR to vllm to support get port from environment.
 #    Future Plan:
 #       Remove those patch when vllm merged them
-#   3. `vllm.config.ParallelConfig.ParallelConfig.stateless_init_dp_group`
+#   4. `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

vllm_ascend/patch/__init__.py

@@ -22,9 +22,10 @@
 import vllm.distributed
 import vllm.envs as envs
 from torch.distributed import ProcessGroup
-from vllm.config import ParallelConfig
+from vllm.config import ParallelConfig, VllmConfig
 from vllm.distributed.utils import \
     stateless_init_torch_distributed_process_group
+from vllm.v1.engine.core import DPEngineCoreProc
 
 from vllm_ascend.utils import NullHandle, is_310p
 
@@ -81,7 +82,22 @@ def stateless_init_dp_group(self) -> "ProcessGroup":
     return dp_group
 
 
+def _init_data_parallel(self, vllm_config: VllmConfig):
+    # Configure NPUs and stateless process group for data parallel.
+    dp_rank = vllm_config.parallel_config.data_parallel_rank
+    dp_size = vllm_config.parallel_config.data_parallel_size
+    local_dp_rank = vllm_config.parallel_config.data_parallel_rank_local
+
+    assert dp_size > 1
+    assert 0 <= local_dp_rank <= dp_rank < dp_size
+
+    self.local_dp_rank = local_dp_rank
+    self.dp_group = vllm_config.parallel_config.stateless_init_dp_group()
+    self.current_wave = 0
+
+
 vllm.distributed.parallel_state.destroy_model_parallel = ascend_destroy_model_parallel
+DPEngineCoreProc._init_data_parallel = _init_data_parallel
 ParallelConfig.get_next_dp_init_port = parallel_config_get_dp_port
 ParallelConfig.stateless_init_dp_group = stateless_init_dp_group
 

vllm_ascend/patch/platform/patch_common/patch_distributed.py

@@ -74,6 +74,12 @@ def __init__(
                          distributed_init_method=distributed_init_method,
                          is_driver_worker=is_driver_worker)
 
+        # NOTE(Yizhou): Since we do not set ASCEND_RT_VISIBLE_DEVICES in
+        # vllm_ascend, we need to set the device id manually.
+        local_dp_rank = self.vllm_config.parallel_config.data_parallel_rank_local
+        world_size = self.vllm_config.parallel_config.world_size
+        self.local_rank_across_dp = local_dp_rank * world_size + self.local_rank
+
         # Try to import mindie_turbo to accelerate vLLM inference.
         try_register_lib(
             "mindie_turbo",
@@ -117,7 +123,7 @@ def initialize_cache(self, num_gpu_blocks: int,
 
     def init_device(self):
         if self.device_config.device.type == "npu":
-            self.device = torch.device(f"npu:{self.local_rank}")
+            self.device = torch.device(f"npu:{self.local_rank_across_dp}")
             NPUPlatform.set_device(self.device)
             NPUPlatform.empty_cache()
             self.init_npu_memory = NPUPlatform.mem_get_info()[0]