Bump torch version to 2.7.1

.github/workflows/release_whl.yml

@@ -90,7 +90,8 @@ jobs:
           --exclude libc10.so \
           --exclude libc_sec.so \
           --exclude "libascend*.so" \
-          --exclude "libtorch*.so"
+          --exclude "libtorch*.so" \
+          --exclude "liberror_manager.so"
         done
         rm -f dist/*.whl
         mv dist/repaired/*.whl dist/

README.md

@@ -42,7 +42,7 @@ By using vLLM Ascend plugin, popular open-source models, including Transformer-l
 - Software:
   * Python >= 3.9, < 3.12
   * CANN >= 8.2.rc1
-  * PyTorch >= 2.5.1, torch-npu >= 2.5.1.post1.dev20250619
+  * PyTorch >= 2.7.1, torch-npu >= 2.7.1.dev20250724
   * vLLM (the same version as vllm-ascend)
 
 ## Getting Started

README.zh.md

@@ -42,7 +42,7 @@ 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 >= 8.2.rc1
-  * PyTorch >= 2.5.1, torch-npu >= 2.5.1.post1.dev20250619
+  * PyTorch >= 2.5.1, torch-npu >= 2.7.1.dev20250724
   * vLLM (与vllm-ascend版本一致)
 
 ## 开始使用

docs/source/installation.md

@@ -12,8 +12,8 @@ This document describes how to install vllm-ascend manually.
     | Software      | Supported version                | Note                                      |
     |---------------|----------------------------------|-------------------------------------------|
     | CANN          | >= 8.2.RC1                       | Required for vllm-ascend and torch-npu    |
-    | torch-npu     | >= 2.5.1.post1.dev20250619       | Required for vllm-ascend, No need to install manually, it will be auto installed in below steps |
-    | torch         | >= 2.5.1                         | Required for torch-npu and vllm           |
+    | torch-npu     | >= 2.7.1.dev20250724             | Required for vllm-ascend, No need to install manually, it will be auto installed in below steps |
+    | torch         | >= 2.7.1                         | 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.

pyproject.toml

@@ -12,9 +12,9 @@ requires = [
     "scipy",
     "setuptools>=64",
     "setuptools-scm>=8",
-    "torch-npu==2.5.1.post1.dev20250619",
-    "torch>=2.5.1",
-    "torchvision<0.21.0",
+    "torch-npu==2.7.1.dev20250724",
+    "torch>=2.7.1",
+    "torchvision",
     "wheel",
     "msgpack",
     "quart",

requirements.txt

@@ -10,8 +10,8 @@ pyyaml
 scipy
 setuptools>=64
 setuptools-scm>=8
-torch>=2.5.1
-torchvision<0.21.0
+torch>=2.7.1
+torchvision
 wheel
 # Remove after https://github.com/vllm-project/vllm-ascend/issues/2034
 transformers<4.54.0
@@ -26,4 +26,4 @@ numba
 # Install torch_npu
 --pre
 --extra-index-url https://mirrors.huaweicloud.com/ascend/repos/pypi
-torch-npu==2.5.1.post1.dev20250619
+torch-npu==2.7.1.dev20250724

tests/e2e/singlecard/ops/test_fused_moe.py

@@ -19,9 +19,6 @@
 
 Run `pytest tests/ops/test_fused_moe.py`.
 """
-# fused moe ops test will hit the infer_schema error, we need add the patch
-# here to make the test pass.
-import vllm_ascend.patch.worker.patch_common.patch_utils  # type: ignore[import]  # isort: skip  # noqa
 
 from unittest.mock import MagicMock, patch
 

tests/ut/test_platform.py

@@ -543,13 +543,9 @@ def test_get_piecewise_backend_cls_returns_correct_value(self):
 
     @patch("torch.distributed.is_hccl_available", return_value=True)
     @patch("torch_npu._C._distributed_c10d.ProcessGroupHCCL")
-    @patch("torch_npu._C._distributed_c10d.ProcessGroupHCCL.Options")
     @patch("torch.distributed.ProcessGroup")
-    def test_successful_initialization(self, mock_pg, mock_options_cls,
-                                       mock_pg_hccl, _):
+    def test_successful_initialization(self, mock_pg, mock_pg_hccl, _):
         mock_prefix = MagicMock(spec=PrefixStore)
-        mock_options = MagicMock(spec=ProcessGroup.Options)
-        mock_options_cls.return_value = mock_options
         mock_backend = MagicMock()
         mock_pg_hccl.return_value = mock_backend
         group_rank = 0
@@ -574,8 +570,7 @@ def test_successful_initialization(self, mock_pg, mock_options_cls,
             timeout=timedelta(seconds=30),
         )
 
-        mock_pg.assert_called_once_with(mock_prefix, group_rank, group_size,
-                                        unittest.mock.ANY)
+        mock_pg.assert_called_once_with(mock_prefix, group_rank, group_size)
         mock_pg_hccl.assert_called_once_with(mock_prefix, group_rank,
                                              group_size, unittest.mock.ANY)
         mock_backend._set_sequence_number_for_group.assert_called_once()

vllm_ascend/__init__.py

@@ -23,9 +23,5 @@ def register():
 
 
 def register_model():
-    # fix pytorch schema check error, remove this line after pytorch
-    # is upgraded to 2.7.0
-    import vllm_ascend.patch.worker.patch_common.patch_utils  # noqa: F401
-
     from .models import register_model
     register_model()

vllm_ascend/patch/__init__.py

@@ -75,20 +75,6 @@
 #    Future Plan:
 #       Remove this patch when vllm merged them.
 #
-# ** File: worker/patch_common/patch_utils.py **
-# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-#   1. `vllm.utils.direct_register_custom_op`
-#    Why:
-#       pytorch 2.7.o is not compatible with pytorch 2.5.1. While vllm is based on pytorch 2.7.0, but vllm ascend
-#       is based on pytorch 2.5.1, so we need to use this patch to make vllm compatible with pytorch 2.5.1.
-#    How：
-#       patch __annotations__ check to make it compatible with pytorch 2.5.1.
-#    Related PR (if no, explain why):
-#       This is the problem in vllm-ascend
-#    Future Plan:
-#       Remove this patch once pytorch 2.7.0 is supported for vllm ascend.
-#
-# ** File: worker/patch_0_10_0/patch_sampler_gather_logprobs.py **
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.v1.sample.sampler.Sampler.gather_logprobs`
 #    Why:

vllm_ascend/patch/worker/patch_common/__init__.py

@@ -15,9 +15,6 @@
 # limitations under the License.
 #
 
-# patch_utils should be the first import, because it will be used by other
-# patch files.
-import vllm_ascend.patch.worker.patch_common.patch_utils  # noqa isort:skip
 import vllm_ascend.patch.worker.patch_common.patch_distributed  # noqa
 import vllm_ascend.patch.worker.patch_common.patch_linear  # noqa
 import vllm_ascend.patch.worker.patch_common.patch_minicpm  # noqa

vllm_ascend/platform.py

@@ -259,16 +259,10 @@ def stateless_init_device_torch_dist_pg(
 
         assert is_hccl_available()
 
-        # TODO(Yizhou): The reason we need to set options while vllm does not
-        # seems to be related to the version of PyTorch. In the latest version,
-        # there is no need to set options. While in the older version, 2.5.1
-        # specifically, we need to set options.
-        options = ProcessGroup.Options(backend=backend)
         pg: ProcessGroup = ProcessGroup(
             prefix_store,
             group_rank,
             group_size,
-            options,
         )
 
         backend_options = ProcessGroupHCCL.Options()

vllm_ascend/quantization/quant_config.py

@@ -15,10 +15,6 @@
 # limitations under the License.
 # This file is a part of the vllm-ascend project.
 #
-# By using quantization case, this file is called before worker patch achieve,
-# we need to import patch_utils here first to make sure the patch is applied.
-import vllm_ascend.patch.worker.patch_common.patch_utils  # type: ignore[import]  # isort: skip  # noqa
-
 from types import MappingProxyType
 from typing import Any, Callable, Dict, List, Mapping, Optional