CP kv cache management and input process

.github/actionlint.yaml

@@ -7,7 +7,4 @@ self-hosted-runner:
     - linux-aarch64-a2-4
     - linux-aarch64-a2-8
     - linux-arm64-npu-static-8
-    - linux-aarch64-310p-1
-    - linux-aarch64-310p-2
-    - linux-aarch64-310p-4
     - ubuntu-24.04-arm

.github/workflows/accuracy_test.yaml

@@ -29,15 +29,35 @@ on:
     types: [ labeled ]
   workflow_dispatch:
     inputs:
-      vllm-ascend-version:
-        description: 'vllm-ascend:'
+      vllm-version:
+        description: 'vllm version:'
         required: true
         type: choice
+        # Please also update this when bump matched version
         # Current supported vLLM versions
         options:
-          - latest
           - main
-        default: main
+          - v0.10.0
+          - v0.9.1
+          - v0.7.3
+      vllm-ascend-version:
+        description: 'vllm-ascend version:'
+        required: true
+        type: choice
+        options:
+          - main
+          - v0.9.1-dev
+          - v0.7.3-dev
+      models:
+        description: 'model:'
+        required: true
+        type: choice
+        options:
+          - all
+          - Qwen/Qwen2.5-VL-7B-Instruct
+          - Qwen/Qwen3-8B-Base
+          - Qwen/Qwen3-30B-A3B
+        default: 'all'
 
 # 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.
@@ -56,27 +76,58 @@ jobs:
     # test will be triggered when tag '*-accuracy-test' & 'ready-for-test' or workflow_dispatch job
     if:  >-
       ${{
-      contains(github.event.pull_request.labels.*.name, 'accuracy-test') &&
+      (contains(github.event.pull_request.labels.*.name, 'accuracy-test') ||
+      contains(github.event.pull_request.labels.*.name, 'vl-accuracy-test') ||
+      contains(github.event.pull_request.labels.*.name, 'moe-accuracy-test') ||
+      contains(github.event.pull_request.labels.*.name, 'dense-accuracy-test')) &&
       contains(github.event.pull_request.labels.*.name, 'ready-for-test') ||
       github.event_name == 'workflow_dispatch' || github.event_name == 'schedule'
       }}
-    runs-on: ${{ matrix.runner }}
+    runs-on: >-
+      ${{
+          (matrix.model_name == 'Qwen/Qwen3-30B-A3B' && 'linux-aarch64-a2-2') ||
+          'linux-aarch64-a2-1'
+      }}
     strategy:
       matrix:
-        include:
-          - model_name: Qwen3-8B-Base
-            runner: linux-aarch64-a2-1
-          - model_name: Qwen2.5-VL-7B-Instruct
-            runner: linux-aarch64-a2-1
-          - model_name: Qwen3-30B-A3B
-            runner: linux-aarch64-a2-2
-      fail-fast: false
+        # the accuracy test will run:
+        # 1. workflow_dispatch with models input
+        #   - all: Qwen/Qwen3-30B-A3B, Qwen/Qwen2.5-VL-7B-Instruct, Qwen/Qwen3-8B-Base
+        #   - specified but not all: Qwen/Qwen3-30B-A3B, Qwen/Qwen2.5-VL-7B-Instruct, Qwen/Qwen3-8B-Base
+        # 2. PR labeled with "*-accuracy-test"
+        #   - accuracy-test: Qwen/Qwen3-8B-Base, Qwen/Qwen2.5-VL-7B-Instruct, Qwen/Qwen3-30B-A3B
+        #   - dense-accuracy-test: Qwen/Qwen3-8B-Base
+        #   - vl-accuracy-test: Qwen/Qwen2.5-VL-7B-Instruct
+        #   - moe-accuracy-test: Qwen/Qwen3-30B-A3B
+        model_name: ${{ fromJSON(
+          (github.event_name == 'schedule' &&
+            '["Qwen/Qwen3-30B-A3B","Qwen/Qwen2.5-VL-7B-Instruct","Qwen/Qwen3-8B-Base"]') ||
+          (github.event.inputs.models == 'all' &&
+            '["Qwen/Qwen3-30B-A3B","Qwen/Qwen2.5-VL-7B-Instruct","Qwen/Qwen3-8B-Base"]') ||
+          (github.event.inputs.models == 'Qwen/Qwen3-30B-A3B' &&
+            '["Qwen/Qwen3-30B-A3B"]') ||
+          (github.event.inputs.models == 'Qwen/Qwen2.5-VL-7B-Instruct' &&
+            '["Qwen/Qwen2.5-VL-7B-Instruct"]') ||
+          (github.event.inputs.models == 'Qwen/Qwen3-8B-Base' &&
+            '["Qwen/Qwen3-8B-Base"]') ||
+          contains(github.event.pull_request.labels.*.name, 'accuracy-test') &&
+            '["Qwen/Qwen3-8B-Base","Qwen/Qwen2.5-VL-7B-Instruct", "Qwen/Qwen3-30B-A3B"]' ||
+          contains(github.event.pull_request.labels.*.name, 'dense-accuracy-test') &&
+            '["Qwen/Qwen3-8B-Base"]' ||
+          contains(github.event.pull_request.labels.*.name, 'vl-accuracy-test') &&
+            '["Qwen/Qwen2.5-VL-7B-Instruct"]' ||
+          contains(github.event.pull_request.labels.*.name, 'moe-accuracy-test') &&
+            '["Qwen/Qwen3-30B-A3B"]'
+         ) }}
 
+      fail-fast: false
     name: ${{ matrix.model_name }} accuracy
     container:
       image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.2.rc1-910b-ubuntu22.04-py3.11
       env:
+        DATASET_SOURCE: ModelScope
         VLLM_USE_MODELSCOPE: True
+        USE_MODELSCOPE_HUB: 1
         # 1. If version specified (work_dispatch), do specified branch accuracy test
         # 2. If no version (labeled PR), do accuracy test by default ref:
         # The branch, tag or SHA to checkout. When checking out the repository that
@@ -88,10 +139,10 @@ jobs:
       - name: Checkout repository
         uses: actions/checkout@v4
 
-      - name: Set model name as output
-        id: set_output
+      - name: Check npu and CANN info
         run: |
-          echo "model_name=${{ matrix.model_name }}" >> $GITHUB_OUTPUT
+          npu-smi info
+          cat /usr/local/Ascend/ascend-toolkit/latest/"$(uname -i)"-linux/ascend_toolkit_install.info
 
       - name: Config mirrors
         run: |
@@ -110,19 +161,19 @@ jobs:
         uses: actions/checkout@v4
         with:
           repository: vllm-project/vllm
-          ref: v0.10.0
           path: ./vllm-empty
+          # Please also update this when bump matched version
+          ref: ${{ github.event.inputs.vllm-version || 'v0.10.0' }}
 
       - name: Install vllm-project/vllm from source
         working-directory: ./vllm-empty
-        run: |
-          VLLM_TARGET_DEVICE=empty pip install -e .
+        run: VLLM_TARGET_DEVICE=empty pip install -e .
 
       - name: Resolve vllm-ascend version
         run: |
           VERSION_INPUT="${{ github.event.inputs.vllm-ascend-version }}"
 
-          if [[ "$VERSION_INPUT" == "latest" ]]; then
+          if [[ "$VERSION_INPUT" == "main" ]]; then
             TAGS=$(git ls-remote --tags --sort=-v:refname https://github.com/vllm-project/vllm-ascend "v*" | cut -f2 | sed 's|refs/tags/||')
             LATEST_TAG=$(echo "$TAGS" | head -n1)
             if [[ -z "$LATEST_TAG" ]]; then
@@ -148,8 +199,8 @@ jobs:
           PIP_EXTRA_INDEX_URL: https://mirrors.huaweicloud.com/ascend/repos/pypi
         run: |
           pip install -r requirements-dev.txt
-          pip install -v -e .
-
+          pip install -v -e . 
+            
       - name: Get vLLM commit hash and URL
         working-directory: ./vllm-empty
         run: |
@@ -162,6 +213,15 @@ jobs:
           VLLM_ASCEND_COMMIT=$(git rev-parse --short=7 HEAD)
           echo "VLLM_ASCEND_COMMIT=$VLLM_ASCEND_COMMIT" >> $GITHUB_ENV
 
+      - name: Print resolved hashes
+        run: |
+          echo "vLLM       : ${{ env.VLLM_COMMIT }}"
+          echo "vLLM-Ascend: ${{ env.VLLM_ASCEND_COMMIT }}"
+
+      - name: Install lm-eval, ray, and datasets
+        run: |
+            pip install lm-eval==0.4.8
+
       - name: Collect version info
         run: |
           for dir in /usr/local/Ascend/ascend-toolkit/*; do
@@ -182,27 +242,37 @@ jobs:
             pip show torch_npu | grep "Version:" | awk '{print "GHA_TORCH_NPU_VERSION="$2}'
             pip show vllm | grep "Version:" | awk '{print "GHA_VLLM_VERSION="$2}' | sed 's/+.*//'
           } >> "$GITHUB_ENV"
+
+      - name: Print versions
+        run: |
+          echo "CANN: ${{ env.GHA_CANN_VERSION }}"
+          echo "Torch NPU: ${{ env.GHA_TORCH_NPU_VERSION }}"
+          echo "Torch: ${{ env.GHA_TORCH_VERSION }}"
+          echo "vLLM: ${{ env.GHA_VLLM_VERSION }}"
+          echo "vLLM Ascend: ${{ env.GHA_VLLM_ASCEND_VERSION }}"
 
-      - name: Run accuracy test
+      - name: Run Accuracy Test
         id: report
+        working-directory: ./benchmarks
         env:
-          VLLM_WORKER_MULTIPROC_METHOD: spawn
-          VLLM_USE_MODELSCOPE: True
-          VLLM_VERSION: ${{ env.GHA_VLLM_VERSION }}
-          VLLM_COMMIT: ${{ env.VLLM_COMMIT }}
-          VLLM_ASCEND_VERSION: ${{ env.GHA_VLLM_ASCEND_VERSION || github.ref }}
-          VLLM_ASCEND_COMMIT: ${{ env.VLLM_ASCEND_COMMIT }}
-          CANN_VERSION: ${{ env.GHA_CANN_VERSION }}
-          TORCH_VERSION: ${{ env.GHA_TORCH_VERSION }}
-          TORCH_NPU_VERSION: ${{ env.GHA_TORCH_NPU_VERSION }}
+          PYTORCH_NPU_ALLOC_CONF: max_split_size_mb:256
         run: |
           model_base_name=$(basename ${{ matrix.model_name }})
           markdown_name="${model_base_name}"
+          echo "markdown_name=$markdown_name"
           echo "markdown_name=$markdown_name" >> $GITHUB_OUTPUT
-          mkdir -p ./benchmarks/accuracy
-          pytest -sv ./tests/e2e/singlecard/models/test_lm_eval_correctness.py \
-          --config ./tests/e2e/singlecard/models/configs/${{ matrix.model_name }}.yaml \
-          --report_output ./benchmarks/accuracy/${model_base_name}.md 
+          mkdir -p ./accuracy
+
+          python ./scripts/run_accuracy.py \
+            --model "${{ matrix.model_name }}" \
+            --output "./accuracy/${markdown_name}.md" \
+            --vllm_ascend_version "${{ env.GHA_VLLM_ASCEND_VERSION || github.ref }}" \
+            --cann_version "${{ env.GHA_CANN_VERSION }}" \
+            --torch_npu_version "${{ env.GHA_TORCH_NPU_VERSION }}" \
+            --torch_version "${{ env.GHA_TORCH_VERSION }}" \
+            --vllm_version "${{ env.GHA_VLLM_VERSION }}" \
+            --vllm_commit "${{ env.VLLM_COMMIT }}" \
+            --vllm_ascend_commit "${{ env.VLLM_ASCEND_COMMIT }}" \
 
       - name: Generate step summary
         if: ${{ always() }}
@@ -214,7 +284,19 @@ jobs:
           SAFE_VLLM_ASCEND_VERSION="${GHA_VLLM_ASCEND_VERSION//\//-}"
           echo "SAFE_VLLM_ASCEND_VERSION=$SAFE_VLLM_ASCEND_VERSION" >> "$GITHUB_ENV"
 
+      - name: Check report first line for failure
+        id: check_report
+        run: |
+          REPORT_PATH="./benchmarks/accuracy/${{ steps.report.outputs.markdown_name }}.md"
+          echo "Scanning $REPORT_PATH for ❌ …"
+          if grep -q '❌' "$REPORT_PATH"; then
+            echo "contains_fail=true" >> $GITHUB_OUTPUT
+          else
+            echo "contains_fail=false" >> $GITHUB_OUTPUT
+          fi
+
       - name: Upload Report 
+        if: ${{ github.event_name == 'workflow_dispatch' && steps.check_report.outputs.contains_fail == 'false' }}
         uses: actions/upload-artifact@v4
         with:
           name: "report-${{ env.SAFE_VLLM_ASCEND_VERSION }}-${{ steps.report.outputs.markdown_name }}"
@@ -223,24 +305,20 @@ jobs:
           retention-days: 90
           overwrite: true
 
-    outputs:
-      model_name: ${{ steps.set_output.outputs.model_name }}
-
   create_pr:
     runs-on: ubuntu-latest
     needs: accuracy_tests
-    if: ${{ github.event_name == 'workflow_dispatch' && github.event.inputs.vllm-ascend-version == 'latest' }}
+    if: ${{ github.event_name == 'workflow_dispatch' }}
     env:
       UPSTREAM_REPO: vllm-project/vllm-ascend
-
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
         with:
           repository: vllm-ascend-ci/vllm-ascend
           token: ${{ secrets.PAT_TOKEN }}
           ref: main
-      
+
       - name: Add upstream remote
         run: |
           git remote add upstream https://github.com/${{ env.UPSTREAM_REPO }}.git
@@ -272,7 +350,7 @@ jobs:
           find ./docs/source/developer_guide/evaluation/accuracy_report -maxdepth 1 -type f -name '*.md' ! -name 'index.md' -delete
           find ./docs/source/developer_guide/evaluation/accuracy_report -mindepth 2 -type f -name '*.md' -exec mv -f {} ./docs/source/developer_guide/evaluation/accuracy_report \;
           find ./docs/source/developer_guide/evaluation/accuracy_report -mindepth 1 -type d -empty -delete
-      
+
       - name: Update accuracy_report/index.md
         run: |
           REPORT_DIR="./docs/source/developer_guide/evaluation/accuracy_report"
@@ -312,10 +390,16 @@ jobs:
               head: `vllm-ascend-ci:${{ env.BRANCH_NAME }}`,
               base: '${{ github.event.inputs.vllm-ascend-version }}',
               title: `[Doc] Update accuracy reports for ${{ github.event.inputs.vllm-ascend-version }}`,
-              body: `The accuracy results running on NPU Altlas A2 have changed, updating reports for: All models (Qwen/Qwen3-30B-A3B, Qwen2.5-VL-7B-Instruct, Qwen3-8B-Base)
+              body: `The accuracy results running on NPU Altlas A2 have changed, updating reports for:
+            ${{ 
+              github.event.inputs.models == 'all' 
+                && 'All models (Qwen/Qwen3-30B-A3B, Qwen2.5-VL-7B-Instruct, Qwen3-8B-Base)' 
+                || github.event.inputs.models 
+            }}
+
+            - [Workflow run][1]
 
-              - [Workflow run][1]
-
-              [1]: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}`
+            [1]: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}`
             });
             core.info(`Created PR #${pr.data.number}`);
+

.github/workflows/release_whl.yml

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

.github/workflows/vllm_ascend_test.yaml

@@ -196,24 +196,20 @@ jobs:
           pytest -sv tests/e2e/singlecard/test_guided_decoding.py
           pytest -sv tests/e2e/singlecard/test_camem.py
           pytest -sv tests/e2e/singlecard/test_embedding.py
-
-          # ------------------------------------ v1 spec decode test ------------------------------------ #
-          pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py
-          # TODO: revert me when test_v1_spec_decode.py::test_ngram_correctness is fixed
-          pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py
-
-          # All other tests, ignore: 310p test, accuracy test.
           pytest -sv tests/e2e/singlecard/ \
           --ignore=tests/e2e/singlecard/test_offline_inference.py \
           --ignore=tests/e2e/singlecard/test_ilama_lora.py \
           --ignore=tests/e2e/singlecard/test_guided_decoding.py \
           --ignore=tests/e2e/singlecard/test_camem.py \
           --ignore=tests/e2e/singlecard/test_embedding.py \
           --ignore=tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py \
-          --ignore=tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py \
-          --ignore=tests/e2e/singlecard/test_offline_inference_310p.py \
-          --ignore=tests/e2e/singlecard/models/test_lm_eval_correctness.py
-  e2e-2-cards:
+          --ignore=tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py
+          # ------------------------------------ v1 spec decode test ------------------------------------ #
+          VLLM_USE_MODELSCOPE=True pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py
+          # TODO: revert me when test_v1_spec_decode.py::test_ngram_correctness is fixed
+          VLLM_USE_MODELSCOPE=True pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py
+
+  e2e-4-cards:
     needs: [e2e]
     if: ${{ needs.e2e.result == 'success' }}
     strategy:
@@ -281,11 +277,7 @@ jobs:
           pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_QwQ
           pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeek_dbo
           pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeekV3_dbo
-          pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_alltoallv
-          pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen3_W4A8DYNAMIC
-          pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeek_W4A8DYNAMIC
           pytest -sv tests/e2e/multicard/test_data_parallel.py
           pytest -sv tests/e2e/multicard/ --ignore=tests/e2e/multicard/test_ilama_lora_tp2.py \
             --ignore=tests/e2e/multicard/test_offline_inference_distributed.py \
-            --ignore=tests/e2e/multicard/test_data_parallel.py \
-            --ignore=tests/e2e/multicard/test_offline_inference_310p.py
+            --ignore=tests/e2e/multicard/test_data_parallel.py