worker-vllm/.github/workflows/serverless-model-tests.yml

name: Serverless Model Tests

on:
  pull_request:
    branches:
      - "**"
  push:
    branches:
      - "main"

permissions:
  contents: read

# Only one run per ref at a time — these tests spin up real H100 endpoints.
concurrency:
  group: serverless-model-tests-${{ github.ref }}
  cancel-in-progress: true

jobs:
  # On PRs we only smoke-test one (small, cheap) model to keep review loops fast.
  test-pr:
    if: github.event_name == 'pull_request'
    runs-on: [blacksmith-8vcpu-ubuntu-2204, linux]
    timeout-minutes: 60
    steps:
      - name: Checkout
        uses: actions/checkout@v4

      - name: Set up QEMU
        uses: docker/setup-qemu-action@v3

      - name: Set up Docker Buildx
        uses: docker/setup-buildx-action@v3

      - name: Login to Docker Hub
        uses: docker/login-action@v3
        with:
          username: ${{ secrets.DOCKERHUB_USERNAME }}
          password: ${{ secrets.DOCKERHUB_TOKEN }}

      - name: Set up Python
        uses: actions/setup-python@v5
        with:
          python-version: "3.11"

      - name: Install script dependencies
        run: pip install pyyaml

      - name: Run serverless e2e test
        id: e2e_test
        run: python scripts/serverless_e2e_test.py --config configs/qwen/qwen3_8b.yaml --build
        env:
          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY_2 }}
          DOCKERHUB_REPO: ${{ vars.DOCKERHUB_REPO || 'runpod' }}
          DOCKERHUB_IMG: ${{ vars.DOCKERHUB_IMG || 'worker-v1-vllm' }}
          HUGGINGFACE_ACCESS_TOKEN: ${{ secrets.HUGGINGFACE_ACCESS_TOKEN_2 }}

      - name: Cleanup safety net
        if: always()
        run: |
          if [ -n "${{ steps.e2e_test.outputs.endpoint_id }}" ]; then
            curl -sf -X DELETE "https://rest.runpod.io/v1/endpoints/${{ steps.e2e_test.outputs.endpoint_id }}" \
              -H "Authorization: Bearer ${{ secrets.RUNPOD_API_KEY_2 }}" || true
          fi
          if [ -n "${{ steps.e2e_test.outputs.template_id }}" ]; then
            curl -sf -X DELETE "https://rest.runpod.io/v1/templates/${{ steps.e2e_test.outputs.template_id }}" \
              -H "Authorization: Bearer ${{ secrets.RUNPOD_API_KEY_2 }}" || true
          fi

  # On push to main we test every tuned model config in parallel.
  discover:
    if: github.event_name == 'push'
    runs-on: ubuntu-latest
    outputs:
      configs: ${{ steps.list.outputs.configs }}
    steps:
      - name: Checkout
        uses: actions/checkout@v4

      - name: List model configs
        id: list
        run: |
          CONFIGS=$(find configs -name '*.yaml' | sort | jq -R -s -c 'split("\n") | map(select(length > 0))')
          echo "configs=$CONFIGS" >> "$GITHUB_OUTPUT"

  # The image is the same regardless of which model config is under test (configs
  # only supply endpoint env vars), so build/push it once and let every matrix job
  # in test-main reuse that same tag instead of rebuilding per config.
  build:
    if: github.event_name == 'push'
    runs-on: [blacksmith-8vcpu-ubuntu-2204, linux]
    outputs:
      image: ${{ steps.build.outputs.image }}
    steps:
      - name: Checkout
        uses: actions/checkout@v4

      - name: Set up QEMU
        uses: docker/setup-qemu-action@v3

      - name: Set up Docker Buildx
        uses: docker/setup-buildx-action@v3

      - name: Login to Docker Hub
        uses: docker/login-action@v3
        with:
          username: ${{ secrets.DOCKERHUB_USERNAME }}
          password: ${{ secrets.DOCKERHUB_TOKEN }}

      - name: Build and push image
        id: build
        env:
          DOCKERHUB_REPO: ${{ vars.DOCKERHUB_REPO || 'runpod' }}
          DOCKERHUB_IMG: ${{ vars.DOCKERHUB_IMG || 'worker-v1-vllm' }}
          RELEASE_VERSION: test-${{ github.sha }}
          HUGGINGFACE_ACCESS_TOKEN: ${{ secrets.HUGGINGFACE_ACCESS_TOKEN }}
        run: |
          docker buildx bake --push
          echo "image=${DOCKERHUB_REPO}/${DOCKERHUB_IMG}:${RELEASE_VERSION}" >> "$GITHUB_OUTPUT"

  test-main:
    needs: [discover, build]
    if: github.event_name == 'push'
    runs-on: [blacksmith-8vcpu-ubuntu-2204, linux]
    timeout-minutes: 60
    strategy:
      fail-fast: false
      matrix:
        config: ${{ fromJSON(needs.discover.outputs.configs) }}
    steps:
      - name: Checkout
        uses: actions/checkout@v4

      - name: Set up Python
        uses: actions/setup-python@v5
        with:
          python-version: "3.11"

      - name: Install script dependencies
        run: pip install pyyaml

      - name: Run serverless e2e test
        id: e2e_test
        run: python scripts/serverless_e2e_test.py --config "${{ matrix.config }}" --image "${{ needs.build.outputs.image }}"
        env:
          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY_2 }}
          HUGGINGFACE_ACCESS_TOKEN: ${{ secrets.HUGGINGFACE_ACCESS_TOKEN_2 }}

      - name: Cleanup safety net
        if: always()
        run: |
          if [ -n "${{ steps.e2e_test.outputs.endpoint_id }}" ]; then
            curl -sf -X DELETE "https://rest.runpod.io/v1/endpoints/${{ steps.e2e_test.outputs.endpoint_id }}" \
              -H "Authorization: Bearer ${{ secrets.RUNPOD_API_KEY_2 }}" || true
          fi
          if [ -n "${{ steps.e2e_test.outputs.template_id }}" ]; then
            curl -sf -X DELETE "https://rest.runpod.io/v1/templates/${{ steps.e2e_test.outputs.template_id }}" \
              -H "Authorization: Bearer ${{ secrets.RUNPOD_API_KEY_2 }}" || true
          fi