[CI] support nightly ci for per pr by labels (#6483)

### What this PR does / why we need it?

This PR refactors the nightly CI workflows (A2 and A3) to support
running tests against a specific PR's code, in addition to the existing
scheduled/dispatch runs using pre-built images.

#### Motivation:
Previously, nightly tests could only be triggered by schedule or
workflow_dispatch, always using the pre-built nightly image. This change
allows developers to trigger nightly tests against their own PR's source
code, enabling early validation without waiting for a nightly build.

#### Changes
Trigger logic (parse-trigger job)

A new parse-trigger job is introduced in both
schedule_nightly_test_a2.yaml and schedule_nightly_test_a3.yaml to
centralize trigger evaluation:

`schedule / workflow_dispatch`: runs all tests with the pre-built image
(existing behavior preserved)
`pull_request (labeled + synchronize)`: runs only when:The PR has the
nightly-test label, and /nightly [test-names] comment exists (latest one
wins)

1. /nightly or /nightly all — runs all tests
2. /nightly test1 test2 — runs only named tests (comma-wrapped for exact
matching)

#### How to trigger
1. Add the nightly-test label to your PR
2. Comment /nightly (all tests) or /nightly test1 test2 (specific tests)
4. Re-triggering: add another /nightly comment and push a new commit
(synchronize event)

### Does this PR introduce _any_ user-facing change?
None

### How was this patch tested?

- vLLM version: v0.14.1
- vLLM main:
dc917cceb8

---------

Signed-off-by: hfadzxy <starmoon_zhang@163.com>

.github/workflows/_e2e_nightly_multi_node.yaml

@@ -45,6 +45,12 @@ on:
         default: main
         type: string
         description: used for pr level tests
+      is_pr_test:
+        required: true
+        type: boolean
+      is_run:
+        required: true  
+        type: boolean
     secrets:
       KUBECONFIG_B64:
         required: true
@@ -68,6 +74,7 @@ jobs:
     name: ${{ inputs.config_file_path }}
     # This is the runner with no NPU for k8s controller
     runs-on: ${{ inputs.runner }}
+    if: ${{ inputs.is_run }}
     container:
       image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-cpu
       env:
@@ -78,8 +85,19 @@ jobs:
         - name: Decode kubeconfig from secrets
           run: |
             # Decode and save kubeconfig
-            echo "${{ secrets.KUBECONFIG_B64 }}" | base64 -d > "$KUBECONFIG"
+            if [ "${{ inputs.is_pr_test }}" = "true" ]; then
-
+              echo "PR test mode"
+              if [ "${{ inputs.soc_version }}" = "a3" ]; then
+                echo "Using A3 cached kubeconfig"
+                cp /root/.cache/.kube/kubeconfig.yaml "$KUBECONFIG"
+              else
+                echo "Using A2 cached kubeconfig"
+                cp /root/.cache/.kube/hk_001_kb.yaml "$KUBECONFIG"
+              fi
+            else
+              echo "Decoding kubeconfig from secrets"
+              echo "${{ secrets.KUBECONFIG_B64 }}" | base64 -d > "$KUBECONFIG"
+            fi
         - name: Checkout code
           uses: actions/checkout@v6
 
@@ -133,9 +151,13 @@ jobs:
             image="${{ inputs.image }}"
             config_file_path="${{ inputs.config_file_path }}"
             fail_tag=FAIL_TAG_"${{ inputs.config_file_path }}"
-            echo "FAIL_TAG=${fail_tag}" >> "$GITHUB_ENV"
+            is_pr_test="${{ inputs.is_pr_test }}"
+            vllm_version="${{ inputs.vllm_version }}"
+            vllm_ascend_ref="${{ inputs.vllm_ascend_ref }}"
+            vllm_ascend_remote_url="${{ inputs.vllm_ascend_remote_url }}"
+            echo "FAIL_TAG=${fail_tag}" >> $GITHUB_ENV
 
-            required_params=("size" "replicas" "image" "config_file_path")
+            required_params=("size" "replicas" "image" "config_file_path" "is_pr_test" "vllm_version" "vllm_ascend_ref" "vllm_ascend_remote_url")
             for param in "${required_params[@]}"; do
               if [ -z "${!param}" ]; then
                 echo "Error: Parameter '$param' is required but empty"
@@ -158,6 +180,10 @@ jobs:
               -D config_file_path="$config_file_path" \
               -D npu_per_node="$npu_per_node" \
               -D fail_tag="$fail_tag" \
+              -D is_pr_test="$is_pr_test" \
+              -D vllm_version="$vllm_version" \
+              -D vllm_ascend_ref="$vllm_ascend_ref" \
+              -D vllm_ascend_remote_url="$vllm_ascend_remote_url" \
               --outfile lws.yaml
 
             kubectl apply -f ./lws.yaml

.github/workflows/_e2e_nightly_single_node.yaml

@@ -36,6 +36,16 @@ on:
       name:
         required: false
         type: string
+      vllm_version:
+        required: false
+        type: string
+        default: "v0.16.0"
+      is_pr_test:
+        required: true
+        type: boolean
+      is_run:
+        required: true
+        type: boolean
 
 # 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.
@@ -54,6 +64,7 @@ jobs:
   e2e-nightly:
     name: ${{ inputs.name || inputs.config_file_path || inputs.tests }}
     runs-on: ${{ inputs.runner }}
+    if: ${{ inputs.is_run }}
     timeout-minutes: 600
     container:
       image: ${{ inputs.image }}
@@ -65,6 +76,60 @@ jobs:
         run: |
           npu-smi info
           cat /usr/local/Ascend/ascend-toolkit/latest/"$(uname -i)"-linux/ascend_toolkit_install.info
+      
+      - name: uninstall vlm vllm-ascend and remove code (if pr test)
+        if: ${{ inputs.is_pr_test }}
+        run: |
+          pip uninstall -y vllm vllm-ascend || true
+          rm -rf /vllm-workspace/vllm /vllm-workspace/vllm-ascend
+
+      - name: Checkout vllm-project/vllm repo
+        if: ${{ inputs.is_pr_test }}
+        uses: actions/checkout@v6
+        with:
+          repository: vllm-project/vllm
+          ref: ${{ inputs.vllm_version }}
+          path: ./temp-vllm 
+          fetch-depth: 1
+
+      - name: Checkout vllm-project/vllm-ascend repo
+        if: ${{ inputs.is_pr_test }}
+        uses: actions/checkout@v6
+        with:
+          path: ./temp-vllm-ascend
+          fetch-depth: 1
+
+      - name: Move code to /vllm-workspace
+        if: ${{ inputs.is_pr_test }}
+        run: |
+          mv ./temp-vllm /vllm-workspace/vllm
+          mv ./temp-vllm-ascend /vllm-workspace/vllm-ascend
+          ls -R /vllm-workspace
+
+      - name: Install vllm-project/vllm from source
+        if: ${{ inputs.is_pr_test }}
+        working-directory: /vllm-workspace/vllm
+        run: |
+          VLLM_TARGET_DEVICE=empty pip install -e .
+      
+      - name: Install vllm-project/vllm-ascend
+        if: ${{ inputs.is_pr_test }}
+        working-directory: /vllm-workspace/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 .
+
+      - name: Install aisbench
+        if: ${{ inputs.is_pr_test }}
+        shell: bash -l {0} 
+        run: |
+          git clone -b v3.0-20250930-master --depth 1 https://gitee.com/aisbench/benchmark.git /vllm-workspace/vllm-ascend/benchmark
+          cd /vllm-workspace/vllm-ascend/benchmark
+          pip install pytest asyncio pytest-asyncio
+          pip install -e . -r requirements/api.txt -r requirements/extra.txt
+          python3 -m pip cache purge
 
       - name: Show vLLM and vLLM-Ascend version
         working-directory: /vllm-workspace

.github/workflows/_e2e_nightly_single_node_models.yaml

@@ -40,6 +40,9 @@ on:
         required: false
         type: boolean
         default: false
+      is_run:
+        required: true
+        type: boolean
 
 # 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.
@@ -58,6 +61,7 @@ jobs:
   e2e-nightly:
     name: ${{inputs.model_list}} accuracy test
     runs-on: ${{ inputs.runner }}
+    if: ${{ inputs.is_run }}
     container:
       image: "${{ inputs.image }}"
       env:

.github/workflows/schedule_nightly_test_a2.yaml

@@ -24,10 +24,15 @@ on:
       # Run test at 24:00 Beijing time (UTC+8)
       - cron: "0 16 * * *"
   workflow_dispatch:
-  pull_request: 
+  pull_request:
     branches:
       - 'main'
-    types: [ labeled ]
+    types: [labeled, synchronize]
+
+permissions:
+  contents: read
+  pull-requests: read
+  issues: read
 
 # 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.
@@ -42,9 +47,84 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
+  parse-trigger:
+    name: Parse trigger and determine test scope
+    runs-on: linux-aarch64-a2b3-0
+    outputs:
+      should_run: ${{ steps.parse.outputs.should_run }}
+      test_filter: ${{ steps.parse.outputs.test_filter }}
+      is_pr_event: ${{ steps.parse.outputs.is_pr_event }}
+    steps:
+      - name: Parse trigger
+        id: parse
+        uses: actions/github-script@v7
+        with:
+          script: |
+            const eventName = context.eventName;
+
+            function parseNightlyComment(body) {
+              if (!body) return null;
+              const match = body.trim().match(/^\/nightly(?:\s+(.+))?$/m);
+              if (!match) return null;
+              const args = (match[1] || '').trim();
+              if (!args || args === 'all') return 'all';
+              // Wrap with commas for exact-name matching: ",name1,name2,"
+              return ',' + args.split(/\s+/).join(',') + ',';
+            }
+
+            // schedule / workflow_dispatch: run all tests with pre-built image
+            if (eventName === 'schedule' || eventName === 'workflow_dispatch') {
+              core.setOutput('should_run', 'true');
+              core.setOutput('test_filter', 'all');
+              core.setOutput('is_pr_event', 'false');
+              return;
+            }
+
+            // pull_request (labeled / synchronize)
+            if (eventName === 'pull_request') {
+              const labels = context.payload.pull_request.labels.map(l => l.name);
+              if (!labels.includes('nightly-test')) {
+                core.setOutput('should_run', 'false');
+                core.setOutput('test_filter', '');
+                core.setOutput('is_pr_event', 'true');
+                return;
+              }
+              // Search comments for latest /nightly command
+              const prNumber = context.payload.pull_request.number;
+              const comments = await github.paginate(github.rest.issues.listComments, {
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                issue_number: prNumber,
+                per_page: 100,
+              });
+              let testFilter = null;
+              for (let i = comments.length - 1; i >= 0; i--) {
+                const result = parseNightlyComment(comments[i].body);
+                if (result !== null) { testFilter = result; break; }
+              }
+              // No /nightly comment found: do not run any tests
+              if (testFilter === null) {
+                core.info('nightly-test label present but no /nightly comment found; skipping.');
+                core.setOutput('should_run', 'false');
+                core.setOutput('test_filter', '');
+                core.setOutput('is_pr_event', 'true');
+                return;
+              }
+              core.setOutput('should_run', 'true');
+              core.setOutput('test_filter', testFilter);
+              core.setOutput('is_pr_event', 'true');
+              return;
+            }
+
+            // Fallback
+            core.setOutput('should_run', 'false');
+            core.setOutput('test_filter', '');
+            core.setOutput('is_pr_event', 'false');
+
   single-node-tests:
     name: single-node
-    if: always() && (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch')
+    needs: [parse-trigger]
+    if: always() && needs.parse-trigger.outputs.should_run == 'true'
     strategy:
       fail-fast: false
       matrix:
@@ -61,10 +141,25 @@ jobs:
       tests: ${{ matrix.test_config.tests }}
       name: ${{ matrix.test_config.name }}
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a2'
+      is_pr_test: >-
+        ${{
+          needs.parse-trigger.outputs.is_pr_event == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
+      is_run: >-
+        ${{
+          needs.parse-trigger.outputs.should_run == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
 
   single-node-yaml-tests:
     name: single-node
-    if: always() && (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch')
+    needs: [parse-trigger]
+    if: always() && needs.parse-trigger.outputs.should_run == 'true'
     strategy:
       fail-fast: false
       matrix:
@@ -84,11 +179,26 @@ jobs:
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a2'
       config_file_path: ${{ matrix.test_config.config_file_path }}
       name: ${{ matrix.test_config.name }}
+      is_pr_test: >-
+        ${{
+          needs.parse-trigger.outputs.is_pr_event == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
+      is_run: >-
+        ${{
+          needs.parse-trigger.outputs.should_run == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
+
 
   multi-node-tests:
     name: multi-node
-    if: always() && (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch')
+    needs: [parse-trigger, single-node-tests, single-node-yaml-tests]
-    needs: [single-node-tests, single-node-yaml-tests]
+    if: always() && needs.parse-trigger.outputs.should_run == 'true'
     strategy:
       fail-fast: false
       max-parallel: 1
@@ -108,40 +218,55 @@ jobs:
       replicas: 1
       size: ${{ matrix.test_config.size }}
       config_file_path: ${{ matrix.test_config.config_file_path }}
+      vllm_ascend_ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.ref_name }}
+      is_pr_test: >-
+        ${{
+          needs.parse-trigger.outputs.is_pr_event == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
+      is_run: >-
+        ${{
+          needs.parse-trigger.outputs.should_run == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
     secrets:
       KUBECONFIG_B64: ${{ secrets.KUBECONFIG_HK_001_INTERNAL_B64 }}
 
   single-node-accuracy-tests:
-    if: >-
+    needs: [parse-trigger]
-      ${{
+    if: always() && needs.parse-trigger.outputs.should_run == 'true'
-        github.event_name == 'schedule' ||
-        github.event_name == 'workflow_dispatch' ||
-        contains(github.event.pull_request.labels.*.name, 'accuracy-test')
-      }}
     strategy:
       fail-fast: false
       matrix:
         test_config:
-          - os: linux-aarch64-a2b3-1
+          - name: accuracy-group-1
+            os: linux-aarch64-a2b3-1
             model_list:
               - Qwen3-8B
               - Qwen2-Audio-7B-Instruct
               - Qwen3-8B-W8A8
               - Qwen3-VL-8B-Instruct
               - Qwen2.5-Omni-7B
-          - os: linux-aarch64-a2b3-1
+          - name: accuracy-group-2
+            os: linux-aarch64-a2b3-1
             model_list:
               - ERNIE-4.5-21B-A3B-PT
               - InternVL3_5-8B-hf
               - Molmo-7B-D-0924
               - Llama-3.2-3B-Instruct
               - llava-onevision-qwen2-0.5b-ov-hf
-          - os: linux-aarch64-a2b3-2
+          - name: accuracy-group-3
+            os: linux-aarch64-a2b3-2
             model_list:
               - Qwen3-30B-A3B
               - Qwen3-VL-30B-A3B-Instruct
               - Qwen3-30B-A3B-W8A8
-          - os: linux-aarch64-a2b3-4
+          - name: accuracy-group-4
+            os: linux-aarch64-a2b3-4
             model_list:
               - Qwen3-Next-80B-A3B-Instruct
               - Qwen3-Omni-30B-A3B-Instruct
@@ -151,10 +276,18 @@ jobs:
       runner: ${{ matrix.test_config.os }}
       model_list: ${{ toJson(matrix.test_config.model_list) }}
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.1-910b-ubuntu22.04-py3.11'
+      is_run: >-
+        ${{
+          needs.parse-trigger.outputs.should_run == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
       upload: false
 
   doc-test:
     name: doc-test
+    needs: [parse-trigger]
     if: always() && (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch')
     strategy:
       # Each version should be tested

.github/workflows/schedule_nightly_test_a3.yaml

@@ -25,10 +25,15 @@ on:
       # Run test at 24:00 Beijing time (UTC+8)
       - cron: "0 16 * * *"
   workflow_dispatch:
-  pull_request: 
+  pull_request:
     branches:
       - 'main'
-    types: [ labeled ]
+    types: [ labeled, synchronize ]
+
+permissions:
+  contents: read
+  pull-requests: read
+  issues: read
 
 # 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.
@@ -42,9 +47,84 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
+  parse-trigger:
+    name: Parse trigger and determine test scope
+    runs-on: linux-aarch64-a2b3-0
+    outputs:
+      should_run: ${{ steps.parse.outputs.should_run }}
+      test_filter: ${{ steps.parse.outputs.test_filter }}
+      is_pr_event: ${{ steps.parse.outputs.is_pr_event }}
+    steps:
+      - name: Parse trigger
+        id: parse
+        uses: actions/github-script@v7
+        with:
+          script: |
+            const eventName = context.eventName;
+
+            function parseNightlyComment(body) {
+              if (!body) return null;
+              const match = body.trim().match(/^\/nightly(?:\s+(.+))?$/m);
+              if (!match) return null;
+              const args = (match[1] || '').trim();
+              if (!args || args === 'all') return 'all';
+              // Wrap with commas for exact-name matching: ",name1,name2,"
+              return ',' + args.split(/\s+/).join(',') + ',';
+            }
+
+            // schedule / workflow_dispatch: run all tests with pre-built image
+            if (eventName === 'schedule' || eventName === 'workflow_dispatch') {
+              core.setOutput('should_run', 'true');
+              core.setOutput('test_filter', 'all');
+              core.setOutput('is_pr_event', 'false');
+              return;
+            }
+
+            // pull_request (labeled / synchronize)
+            if (eventName === 'pull_request') {
+              const labels = context.payload.pull_request.labels.map(l => l.name);
+              if (!labels.includes('nightly-test')) {
+                core.setOutput('should_run', 'false');
+                core.setOutput('test_filter', '');
+                core.setOutput('is_pr_event', 'true');
+                return;
+              }
+              // Search comments for latest /nightly command
+              const prNumber = context.payload.pull_request.number;
+              const comments = await github.paginate(github.rest.issues.listComments, {
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                issue_number: prNumber,
+                per_page: 100,
+              });
+              let testFilter = null;
+              for (let i = comments.length - 1; i >= 0; i--) {
+                const result = parseNightlyComment(comments[i].body);
+                if (result !== null) { testFilter = result; break; }
+              }
+              // No /nightly comment found: do not run any tests
+              if (testFilter === null) {
+                core.info('nightly-test label present but no /nightly comment found; skipping.');
+                core.setOutput('should_run', 'false');
+                core.setOutput('test_filter', '');
+                core.setOutput('is_pr_event', 'true');
+                return;
+              }
+              core.setOutput('should_run', 'true');
+              core.setOutput('test_filter', testFilter);
+              core.setOutput('is_pr_event', 'true');
+              return;
+            }
+
+            // Fallback
+            core.setOutput('should_run', 'false');
+            core.setOutput('test_filter', '');
+            core.setOutput('is_pr_event', 'false');
+
   multi-node-tests:
     name: multi-node
-    if: always() && (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch')
+    needs: [parse-trigger]
+    if: always() && needs.parse-trigger.outputs.should_run == 'true'
     strategy:
       fail-fast: false
       max-parallel: 1
@@ -106,13 +186,28 @@ jobs:
       replicas: 1
       size: ${{ matrix.test_config.size }}
       config_file_path: ${{ matrix.test_config.config_file_path }}
+      vllm_ascend_ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.ref_name }}
+      is_pr_test: >-
+        ${{
+          needs.parse-trigger.outputs.is_pr_event == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
+      is_run: >-
+        ${{
+          needs.parse-trigger.outputs.should_run == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
     secrets:
       KUBECONFIG_B64: ${{ secrets.KUBECONFIG_B64 }}
-  
+
   single-node-tests:
     name: single-node
-    if: always() && (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch')
+    needs: [parse-trigger, multi-node-tests]
-    needs: [multi-node-tests]
+    if: always() && needs.parse-trigger.outputs.should_run == 'true'
     strategy:
       fail-fast: false
       matrix:
@@ -126,11 +221,25 @@ jobs:
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a3'
       tests: ${{ matrix.test_config.tests }}
       name: ${{ matrix.test_config.name }}
+      is_pr_test: >-
+        ${{
+          needs.parse-trigger.outputs.is_pr_event == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
+      is_run: >-
+        ${{
+          needs.parse-trigger.outputs.should_run == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
 
   single-node-yaml-tests:
     name: single-node
-    if: always() && (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch')
+    if: always() && needs.parse-trigger.outputs.should_run == 'true'
-    needs: [multi-node-tests]
+    needs: [parse-trigger, multi-node-tests]
     strategy:
       fail-fast: false
       matrix:
@@ -190,11 +299,25 @@ jobs:
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a3'
       config_file_path: ${{ matrix.test_config.config_file_path }}
       name: ${{ matrix.test_config.name }}
+      is_pr_test: >-
+        ${{
+          needs.parse-trigger.outputs.is_pr_event == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
+      is_run: >-
+        ${{
+          needs.parse-trigger.outputs.should_run == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
 
   custom-ops-tests:
     name: test ops
-    if: always() && (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch')
+    needs: [parse-trigger, multi-node-tests]
-    needs: multi-node-tests
+    if: always() && needs.parse-trigger.outputs.should_run == 'true'
     strategy:
       fail-fast: false
       matrix:
@@ -208,3 +331,17 @@ jobs:
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a3'
       tests: ${{ matrix.test_config.tests }}
       name: ${{ matrix.test_config.name }}
+      is_pr_test: >-
+        ${{
+          needs.parse-trigger.outputs.is_pr_event == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}
+      is_run: >-
+        ${{
+          needs.parse-trigger.outputs.should_run == 'true' && (
+            needs.parse-trigger.outputs.test_filter == 'all' ||
+            contains(needs.parse-trigger.outputs.test_filter, format(',{0},', matrix.test_config.name))
+          )
+        }}

tests/e2e/nightly/multi_node/scripts/lws-a2.yaml.jinja2

@@ -24,6 +24,14 @@ spec:
                 value: "/vllm-workspace"
               - name: FAIL_TAG
                 value: {{ fail_tag | default("FAIL_TAG") }}
+              - name: IS_PR_TEST
+                value: "{{ is_pr_test | default("false") }}"
+              - name: VLLM_VERSION
+                value: {{ vllm_version | default("latest") }}
+              - name: VLLM_ASCEND_REF
+                value: {{ vllm_ascend_ref | default("main") }}
+              - name: VLLM_ASCEND_REMOTE_URL
+                value: {{ vllm_ascend_remote_url | default("https://github.com/vllm-project/vllm-ascend.git") }}
             command:
               - sh
               - -c
@@ -76,6 +84,14 @@ spec:
                 value: "/vllm-workspace"
               - name: FAIL_TAG
                 value: {{ fail_tag | default("FAIL_TAG") }}
+              - name: IS_PR_TEST
+                value: "{{ is_pr_test | default("false") }}"
+              - name: VLLM_VERSION
+                value: {{ vllm_version | default("latest") }}
+              - name: VLLM_ASCEND_REF
+                value: {{ vllm_ascend_ref | default("main") }}
+              - name: VLLM_ASCEND_REMOTE_URL
+                value: {{ vllm_ascend_remote_url | default("https://github.com/vllm-project/vllm-ascend.git") }}
             command:
               - sh
               - -c

tests/e2e/nightly/multi_node/scripts/lws.yaml.jinja2

@@ -24,6 +24,14 @@ spec:
                 value: "/vllm-workspace"
               - name: FAIL_TAG
                 value: {{ fail_tag | default("FAIL_TAG") }}
+              - name: IS_PR_TEST
+                value: "{{ is_pr_test | default("false") }}"
+              - name: VLLM_VERSION
+                value: {{ vllm_version | default("latest") }}
+              - name: VLLM_ASCEND_REF
+                value: {{ vllm_ascend_ref | default("main") }}
+              - name: VLLM_ASCEND_REMOTE_URL
+                value: {{ vllm_ascend_remote_url | default("https://github.com/vllm-project/vllm-ascend.git") }}
             command:
               - sh
               - -c
@@ -76,6 +84,14 @@ spec:
                 value: "/vllm-workspace"
               - name: FAIL_TAG
                 value: {{ fail_tag | default("FAIL_TAG") }}
+              - name: IS_PR_TEST
+                value: "{{ is_pr_test | default("false") }}"
+              - name: VLLM_VERSION
+                value: {{ vllm_version | default("latest") }}
+              - name: VLLM_ASCEND_REF
+                value: {{ vllm_ascend_ref | default("main") }}
+              - name: VLLM_ASCEND_REMOTE_URL
+                value: {{ vllm_ascend_remote_url | default("https://github.com/vllm-project/vllm-ascend.git") }}
             command:
               - sh
               - -c

tests/e2e/nightly/multi_node/scripts/run.sh

@@ -104,8 +104,6 @@ check_npu_info() {
 
 check_and_config() {
     echo "====> Configure mirrors and git proxy"
-    # Fix me(Potabk): Currently, there have some issues with accessing GitHub via https://gh-proxy.test.osinfra.cn in certain regions.
-    # We should switch to a more stable proxy for now until the network proxy is stable enough.
     git config --global url."https://ghfast.top/https://github.com/".insteadOf "https://github.com/"
     pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
     export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi
@@ -136,6 +134,64 @@ install_extra_components() {
     echo "====> Extra components installation completed"
 }
 
+checkout_src() {
+    echo "====> Checkout source code"
+    mkdir -p "$WORKSPACE"
+    cd "$WORKSPACE"
+    pip uninstall -y vllm vllm-ascend || true
+    rm -rf "$WORKSPACE/vllm" "$WORKSPACE/vllm-ascend"
+
+    if [ ! -d "$WORKSPACE/vllm-ascend" ]; then
+        echo "Cloning vllm-ascend from $VLLM_ASCEND_REMOTE_URL"
+        git clone --depth 1 "$VLLM_ASCEND_REMOTE_URL" "$WORKSPACE/vllm-ascend"
+        cd "$WORKSPACE/vllm-ascend"
+        PR_REF=$(git ls-remote origin 'refs/pull/*/head' | grep "^${VLLM_ASCEND_REF}" | awk '{print $2}' | head -1)
+        if [ -n "$PR_REF" ]; then
+            git fetch --depth 1 origin "$PR_REF"
+            git checkout FETCH_HEAD
+        else
+            git fetch origin '+refs/pull/*/head:refs/remotes/pull/*' 2>/dev/null || true
+            git checkout "$VLLM_ASCEND_REF"
+        fi
+    fi
+
+    if [ ! -d "$WORKSPACE/vllm" ]; then
+        echo "Cloning vllm version/ref: $VLLM_VERSION"
+        git clone --depth 1 --branch "$VLLM_VERSION" https://github.com/vllm-project/vllm.git "$WORKSPACE/vllm"
+    fi
+}
+
+install_vllm() {
+    echo "====> Install vllm and vllm-ascend"
+    VLLM_TARGET_DEVICE=empty pip install -e "$WORKSPACE/vllm"
+    pip install -r "$WORKSPACE/vllm-ascend/requirements-dev.txt"
+    pip install -e "$WORKSPACE/vllm-ascend"
+}
+
+install_aisbench() {
+    echo "====> Install AISBench benchmark"
+    
+    export AIS_BENCH_URL="https://gitee.com/aisbench/benchmark.git"
+    : "${AIS_BENCH_TAG:=v3.0-20250930-master}"  
+
+    BENCH_DIR="$WORKSPACE/vllm-ascend/benchmark"
+
+    if [ -d "$BENCH_DIR" ]; then
+        echo "Removing existing benchmark directory..."
+        rm -rf "$BENCH_DIR"
+    fi
+
+    git clone -b "${AIS_BENCH_TAG}" --depth 1 \
+        "${AIS_BENCH_URL}" "${BENCH_DIR}"
+
+    cd "$BENCH_DIR"
+    pip install -e . \
+        -r requirements/api.txt \
+        -r requirements/extra.txt
+
+    python3 -m pip cache purge || echo "WARNING: pip cache purge failed, but proceeding..."
+
+}
 
 show_triton_ascend_info() {
     echo "====> Check triton ascend info"
@@ -170,6 +226,11 @@ If this is insufficient to pinpoint the error, please download and review the lo
 main() {
     check_npu_info
     check_and_config
+    if [[ "$IS_PR_TEST" == "true" ]]; then
+        checkout_src
+        install_vllm
+        install_aisbench
+    fi
     show_vllm_info
     show_triton_ascend_info
     if [[ "$CONFIG_YAML_PATH" == *"DeepSeek-V3_2-Exp-bf16.yaml" ]]; then