[Misc] main2main upgrade vllm to v0.23.0 (vllm-project#10534)

SOMEONEUNSEEN · web-flow · commit edcae83da32a · 2026-06-24T14:16:23.000+08:00
### What this PR does / why we need it? This PR upgrades the default vLLM version for the vllm-ascend project from v0.22.1 to v0.23.0, and re-enables release-tag CI testing. **1. vLLM Version Update** - Updated `.github/vllm-release-tag.commit` to `v0.23.0` - Updated default `VLLM_TAG` to `v0.23.0` in all Dockerfiles (`Dockerfile`, `Dockerfile.310p`, `Dockerfile.a3`, `Dockerfile.a5`, and their openEuler variants) - Updated version references in documentation **2. CI Workflow Optimization** - Re-enabled release-tag CI testing: - `pr_e2e_command.yml`: Restored release-tag test matrix - `pr_test.yaml`: Restored release-tag test matrix - `schedule_update_estimated_times.yaml`: Restored release-tag scheduled job and timing updates - `schedule_vllm_e2e_test.yaml`: Re-enabled weekly upstream E2E test schedule **3. v2 Model Runner Compatibility Handling** Due to API differences between vLLM v0.23.0 and main branch, v2 model runner patches are only supported on main branch: - Added `_V2_MODEL_RUNNER_SUPPORTED` flag in `vllm_ascend/patch/worker/__init__.py` - Added fallback logic in `vllm_ascend/worker/worker.py` to automatically fall back to v1 when v2 model runner is set on v0.23.0 - Updated related tests to skip v2 model runner tests on v0.23.0 **4. Documentation Updates** - Updated branch status in `README.md` and `README.zh.md` to indicate main branch supports CI coverage for both vLLM main and v0.23.0 - Updated example branch name in `docs/source/community/slash-commands.md` ### Does this PR introduce _any_ user-facing change? Yes. Users must upgrade their upstream vLLM installation to v0.23.0. This release of vllm-ascend is **not** compatible with vLLM v0.22.1 or earlier. The `VLLM_USE_V2_MODEL_RUNNER` flag remains unsupported on the v0.23.0 release tag (only supported on vLLM main). ### How was this patch tested? full test https://github.com/vllm-project/vllm-ascend/actions/runs/27616069133/job/81653220088?pr=10534 https://github.com/vllm-project/vllm-ascend/actions/runs/27665474683/job/81818968014?pr=10534 https://github.com/vllm-project/vllm-ascend/actions/runs/28016862824/job/82923233743?pr=10534 https://github.com/vllm-project/vllm-ascend/actions/runs/28073790792/job/83114305076?pr=10534 - vLLM version: v0.22.1 - vLLM main: vllm-project/vllm@967c5c3 --------- Signed-off-by: liyishi <1252651434@qq.com>
diff --git a/.github/vllm-release-tag.commit b/.github/vllm-release-tag.commit
@@ -1 +1 @@
-v0.22.1
+v0.23.0
diff --git a/.github/workflows/pr_e2e_command.yml b/.github/workflows/pr_e2e_command.yml
@@ -181,10 +181,7 @@ jobs:
       matrix:
         vllm_version:
           - ${{ needs.e2e-test.outputs.main_commit }}
-          # Dropped vLLM v0.22.1 (release-tag) compatibility from main. Uncomment
-          # the line below to re-enable release-tag e2e when maintaining a new
-          # release tag (update .github/vllm-release-tag.commit accordingly).
-          # - ${{ needs.e2e-test.outputs.release_tag }}
+          - ${{ needs.e2e-test.outputs.release_tag }}
     needs: [e2e-test]
     if: ${{ needs.e2e-test.outputs.has_tests == 'true' }}
     uses: ./.github/workflows/_selected_tests.yaml
diff --git a/.github/workflows/pr_test.yaml b/.github/workflows/pr_test.yaml
@@ -207,10 +207,7 @@ jobs:
       matrix:
         vllm_version:
           - ${{ needs.lint-and-select-tests.outputs.main_commit }}
-          # Dropped vLLM v0.22.1 (release-tag) compatibility from main. Uncomment
-          # the line below to re-enable release-tag CI when maintaining a new
-          # release tag (update .github/vllm-release-tag.commit accordingly).
-          # - ${{ needs.lint-and-select-tests.outputs.release_tag }}
+          - ${{ needs.lint-and-select-tests.outputs.release_tag }}
     uses: ./.github/workflows/_selected_tests.yaml
     with:
       vllm: ${{ matrix.vllm_version }}
diff --git a/.github/workflows/schedule_update_estimated_times.yaml b/.github/workflows/schedule_update_estimated_times.yaml
@@ -109,22 +109,18 @@ jobs:
       upload_timing: true
       continue_on_error: true
 
-  # Dropped vLLM v0.22.1 (release-tag) compatibility from main. Uncomment this
-  # job to re-enable release-tag timing collection when maintaining a new
-  # release tag (update .github/vllm-release-tag.commit accordingly).
-  # run-tests-release:
-  #   needs: select-tests
-  #   if: ${{ needs.select-tests.outputs.has_tests == 'true' }}
-  #   uses: ./.github/workflows/_selected_tests.yaml
-  #   with:
-  #     vllm: ${{ needs.select-tests.outputs.release_tag }}
-  #     test_groups: ${{ needs.select-tests.outputs.test_groups }}
-  #     upload_timing: true
-  #     continue_on_error: true
+  run-tests-release:
+    needs: select-tests
+    if: ${{ needs.select-tests.outputs.has_tests == 'true' }}
+    uses: ./.github/workflows/_selected_tests.yaml
+    with:
+      vllm: ${{ needs.select-tests.outputs.release_tag }}
+      test_groups: ${{ needs.select-tests.outputs.test_groups }}
+      upload_timing: true
+      continue_on_error: true
 
   update-estimated-times:
-    needs: [run-tests-main]
-    # needs: [run-tests-main, run-tests-release]
+    needs: [run-tests-main, run-tests-release]
     if: always()
     runs-on: ubuntu-latest
     steps:
diff --git a/.github/workflows/schedule_vllm_e2e_test.yaml b/.github/workflows/schedule_vllm_e2e_test.yaml
@@ -18,14 +18,8 @@
 name: E2E-upstream
 
 on:
-  # Dropped vLLM v0.22.1 (release-tag) compatibility from main. This workflow
-  # runs vllm-ascend main against the v0.22.1 release tag and would now fail, so
-  # the weekly schedule is disabled (left manually runnable via workflow_dispatch).
-  # Re-enable the cron below and update .github/vllm-release-tag.commit when
-  # maintaining a new release tag.
-  workflow_dispatch:
-  # schedule:
-  #   - cron: '0 8 * * 0' # every Sunday at 8:00
+  schedule:
+    - cron: '0 8 * * 0' # every Sunday at 8:00
 # Bash shells do not use ~/.profile or ~/.bashrc so these shells need to be explicitly
 # declared as "shell: bash -el {0}" on steps that need to be properly activated.
 # It's used to activate ascend-toolkit environment variables.
diff --git a/Dockerfile b/Dockerfile
@@ -47,7 +47,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL} && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.22.1
+ARG VLLM_TAG=v0.23.0
 ARG VLLM_COMMIT=""
 RUN if [ -n "$VLLM_COMMIT" ]; then \
       git init /vllm-workspace/vllm && \
diff --git a/Dockerfile.310p b/Dockerfile.310p
@@ -33,7 +33,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL} && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.22.1
+ARG VLLM_TAG=v0.23.0
 ARG VLLM_COMMIT=""
 RUN if [ -n "$VLLM_COMMIT" ]; then \
       git init /vllm-workspace/vllm && \
diff --git a/Dockerfile.310p.openEuler b/Dockerfile.310p.openEuler
@@ -32,7 +32,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL} && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.22.1
+ARG VLLM_TAG=v0.23.0
 ARG VLLM_COMMIT=""
 RUN if [ -n "$VLLM_COMMIT" ]; then \
       git init /vllm-workspace/vllm && \
diff --git a/Dockerfile.a3 b/Dockerfile.a3
@@ -49,7 +49,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL} && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.22.1
+ARG VLLM_TAG=v0.23.0
 ARG VLLM_COMMIT=""
 RUN if [ -n "$VLLM_COMMIT" ]; then \
       git init /vllm-workspace/vllm && \
diff --git a/Dockerfile.a3.openEuler b/Dockerfile.a3.openEuler
@@ -47,7 +47,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL} && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.22.1
+ARG VLLM_TAG=v0.23.0
 ARG VLLM_COMMIT=""
 RUN if [ -n "$VLLM_COMMIT" ]; then \
       git init /vllm-workspace/vllm && \
diff --git a/Dockerfile.a5 b/Dockerfile.a5
@@ -49,7 +49,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL} && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.22.1
+ARG VLLM_TAG=v0.23.0
 RUN git clone --depth 1 -b $VLLM_TAG $VLLM_REPO /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/Dockerfile.a5.openEuler b/Dockerfile.a5.openEuler
@@ -47,7 +47,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL} && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.22.1
+ARG VLLM_TAG=v0.23.0
 RUN git clone --depth 1 -b $VLLM_TAG $VLLM_REPO /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/Dockerfile.openEuler b/Dockerfile.openEuler
@@ -47,7 +47,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL} && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.22.1
+ARG VLLM_TAG=v0.23.0
 ARG VLLM_COMMIT=""
 RUN if [ -n "$VLLM_COMMIT" ]; then \
       git init /vllm-workspace/vllm && \
diff --git a/README.md b/README.md
@@ -86,7 +86,7 @@ Below are the maintained branches:
 
 | Branch           | Status       | Note                                 |
 |------------------|--------------|--------------------------------------|
-| main             | Maintained   | CI commitment for vLLM main branch |
+| main             | Maintained   | CI commitment for vLLM main branch and vLLM v0.23.0 tag |
 | v0.7.1-dev       | Unmaintained | Outdated, no longer maintained. |
 | v0.7.3-dev       | Unmaintained | Only bug fixes are allowed, and no new release tags anymore. |
 | v0.9.1-dev       | Unmaintained | Only bug fixes are allowed, and no new release tags anymore. |
diff --git a/README.zh.md b/README.zh.md
@@ -80,7 +80,7 @@ vllm-ascend有主干分支和开发分支。
 
 | 分支              | 状态         | 备注                  |
 |------------------|--------------|----------------------|
-| main             | Maintained   | 基于vLLM main分支CI看护   |
+| main             | Maintained   | 基于vLLM main分支和vLLM最新版本（v0.23.0）CI看护   |
 | v0.7.1-dev       | Unmaintained | 不再维护 |
 | v0.7.3-dev       | Unmaintained | 只允许Bug修复，不会再发布新版本 |
 | v0.9.1-dev       | Unmaintained | 只允许Bug修复，不会再发布新版本 |
diff --git a/docs/source/community/slash-commands.md b/docs/source/community/slash-commands.md
@@ -65,7 +65,7 @@ Use `--branch <name>` to specify a target branch. Without `--branch`, all argume
 /nightly qwen3-vl-32b-instruct-w8a8
 
 # Run on a specific release branch
-/nightly qwen3-vl-32b-instruct-w8a8 --branch releases/v0.22.1
+/nightly qwen3-vl-32b-instruct-w8a8 --branch releases/v0.23.0
 
 # Run all tests on a specific branch
 /nightly all --branch my-feature-branch
diff --git a/tests/e2e/pull_request/one_card/model_runner_v2/test_basic.py b/tests/e2e/pull_request/one_card/model_runner_v2/test_basic.py
@@ -22,12 +22,18 @@
 from vllm import SamplingParams
 
 from tests.e2e.conftest import VllmRunner
+from vllm_ascend.utils import vllm_version_is
 
 MODELS = ["Qwen/Qwen3-0.6B", "vllm-ascend/DeepSeek-V2-Lite-W8A8"]
 
 MAIN_MODELS = ["LLM-Research/Meta-Llama-3.1-8B-Instruct"]
 EGALE_MODELS = ["vllm-ascend/EAGLE-LLaMA3.1-Instruct-8B"]
 
+pytestmark = pytest.mark.skipif(
+    vllm_version_is("0.23.0"),
+    reason="v2 model runner patches not supported on v0.23.0",
+)
+
 
 @pytest.mark.skipif(True, reason="Fix me, it's broken after CANN and trition-ascend are upgraded.")
 @pytest.mark.parametrize("model", MODELS)
diff --git a/tests/e2e/pull_request/one_card/test_guided_decoding.py b/tests/e2e/pull_request/one_card/test_guided_decoding.py
@@ -28,6 +28,7 @@
 from vllm.sampling_params import SamplingParams, StructuredOutputsParams
 
 from tests.e2e.conftest import VllmRunner
+from vllm_ascend.utils import vllm_version_is
 
 MODEL_NAME = "Qwen/Qwen3-0.6B"
 
@@ -37,6 +38,9 @@
 @pytest.fixture(params=[False, True], ids=["v1", "v2"])
 def model_runner_env(request):
     use_v2_model_runner = request.param
+    if use_v2_model_runner and vllm_version_is("0.23.0"):
+        pytest.skip("No need to support v2 model runner for vLLM tag version.")
+
     with patch.dict(os.environ, {"VLLM_USE_V2_MODEL_RUNNER": "1" if use_v2_model_runner else "0"}):
         yield
 
diff --git a/vllm_ascend/distributed/utils.py b/vllm_ascend/distributed/utils.py
@@ -9,12 +9,12 @@
 
 
 def get_decode_context_model_parallel_world_size() -> int:
-    """Return DCP world size (v0.22.1 helper removed on vLLM main)."""
+    """Return DCP world size (v0.21.0 helper removed on vLLM main)."""
     return get_dcp_group().world_size
 
 
 def get_decode_context_model_parallel_rank() -> int:
-    """Return DCP rank within group (v0.22.1 helper removed on vLLM main)."""
+    """Return DCP rank within group (v0.21.0 helper removed on vLLM main)."""
     return get_dcp_group().rank_in_group
 
 
diff --git a/vllm_ascend/patch/__init__.py b/vllm_ascend/patch/__init__.py
@@ -1007,7 +1007,6 @@
 # ** 29. File: platform/patch_mamba_manager.py**
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.v1.core.single_type_kv_cache_manager.MambaManager`
-#      `vllm.v1.core.single_type_kv_cache_manager.spec_manager_map[MambaSpec]`
 #    Why:
 #       Upstream hybrid prefix cache lookup does not support PCP/DCP.
 #    How:
diff --git a/vllm_ascend/patch/worker/__init__.py b/vllm_ascend/patch/worker/__init__.py
@@ -17,11 +17,21 @@
 
 from vllm.triton_utils import HAS_TRITON
 
-from vllm_ascend.utils import is_310p
+from vllm_ascend.utils import is_310p, vllm_version_is
+
+# The v2 model runner is intentionally NOT made compatible with the v0.23.0
+# release. vLLM v0.23.0 and the verified main commit are diverged, and the v2
+# worker patches target main-only APIs; rather than maintain a separate v0.23.0
+# compatibility path we keep v2 main-only. With v0.23.0 installed this flag is
+# False, so none of the patch_v2.* / routed-experts-capture patches below are
+# imported and the v2 worker stays dormant (the release uses the v1 runner).
+_V2_MODEL_RUNNER_SUPPORTED = not vllm_version_is("0.23.0")
 
 if HAS_TRITON:
     import vllm_ascend.patch.worker.patch_triton
-    import vllm_ascend.patch.worker.patch_v2.patch_triton  # noqa
+
+    if _V2_MODEL_RUNNER_SUPPORTED:
+        import vllm_ascend.patch.worker.patch_v2.patch_triton  # noqa
 
 
 import vllm_ascend.patch.worker.patch_weight_utils  # noqa
@@ -53,9 +63,13 @@
 import vllm_ascend.patch.worker.patch_deepseek_mtp  # noqa
 import vllm_ascend.patch.worker.patch_gqa_c8  # noqa
 
-import vllm_ascend.patch.worker.patch_v2.patch_uva  # noqa
-import vllm_ascend.patch.worker.patch_v2.patch_input_batch  # noqa
-import vllm_ascend.patch.worker.patch_v2.patch_model_state  # noqa
-import vllm_ascend.patch.worker.patch_v2.patch_block_table  # noqa
-import vllm_ascend.patch.worker.patch_v2.patch_attn_utils  # noqa
-import vllm_ascend.patch.worker.patch_routed_experts_capture  # noqa
+if _V2_MODEL_RUNNER_SUPPORTED:
+    import vllm_ascend.patch.worker.patch_v2.patch_uva  # noqa
+    import vllm_ascend.patch.worker.patch_v2.patch_input_batch  # noqa
+    import vllm_ascend.patch.worker.patch_v2.patch_model_state  # noqa
+    import vllm_ascend.patch.worker.patch_v2.patch_block_table  # noqa
+    import vllm_ascend.patch.worker.patch_v2.patch_attn_utils  # noqa
+
+# only patch routed experts capture in main2main.
+if _V2_MODEL_RUNNER_SUPPORTED:
+    import vllm_ascend.patch.worker.patch_routed_experts_capture  # noqa
diff --git a/vllm_ascend/spec_decode/llm_base_proposer.py b/vllm_ascend/spec_decode/llm_base_proposer.py
@@ -60,9 +60,8 @@
 def patch_tensor_parallel_group(tp_group):
     """Temporarily swap the global TP group for draft-model spec decode.
 
-    Backports vllm 0.21's ``patch_tensor_parallel_group`` which was removed
-    on vLLM main. Used so the draft model can run with a TP degree that
-    differs from the target model.
+    vllm-ascend local implementation for swapping the global TP group so the
+    draft model can run with a TP degree that differs from the target model.
     """
     old_tp_group = _ps.get_tp_group()
     _ps._TP_STATE_PATCHED = True
diff --git a/vllm_ascend/worker/worker.py b/vllm_ascend/worker/worker.py
@@ -70,6 +70,7 @@
     get_ascend_device_type,
     register_ascend_customop,
     setup_ascend_local_comm_res,
+    vllm_version_is,
 )
 from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
 
@@ -159,6 +160,9 @@ def __init__(
             WEIGHT_LOADER_V2_SUPPORTED.remove("UnquantizedLinearMethod")
 
         self.use_v2_model_runner = envs_vllm.VLLM_USE_V2_MODEL_RUNNER
+        if self.use_v2_model_runner and vllm_version_is("0.23.0"):
+            logger.warning("VLLM_USE_V2_MODEL_RUNNER is not supported on vllm 0.23.0; falling back to v1 model runner.")
+            self.use_v2_model_runner = False
         self._pp_send_work: list[Handle] = []
 
         ascend_compilation_config = get_ascend_config().ascend_compilation_config
