CI/CD для GPU‑нагрузок: контейнеры и пайплайны

				
					name: ci
on:
  push: { branches: ["main"] }
  pull_request:
jobs:
  lint-test:
    runs-on: ubuntu-22.04
    steps:
      - uses: actions/checkout@v4
      - run: pipx install ruff && ruff check .
      - run: pipx install pytest && pytest -q

  build-image:
    needs: lint-test
    runs-on: ubuntu-22.04
    steps:
      - uses: actions/checkout@v4
      - uses: docker/setup-buildx-action@v3
      - uses: docker/login-action@v3
        with: { registry: ${{ secrets.REGISTRY }}, username: ${{ secrets.USER }}, password: ${{ secrets.PASSWORD }} }
      - name: Build & push
        run: |
          docker buildx build --push \
            --build-arg CUDA_VER=12.4.1 \
            --cache-to=type=registry,ref=${{ secrets.REGISTRY }}/proj/cache,mode=max \
            --cache-from=type=registry,ref=${{ secrets.REGISTRY }}/proj/cache \
            -t ${{ secrets.REGISTRY }}/proj/app:${{ github.sha }} .

  gpu-smoke:
    needs: build-image
    runs-on: self-hosted-gpu              # раннер с GPU
    steps:
      - name: Pull image
        run: docker pull ${{ secrets.REGISTRY }}/proj/app:${{ github.sha }}
      - name: Smoke (≤120s)
        timeout-minutes: 5
        run: |
          docker run --rm --gpus 1 \
            -e MODELS_DIR=/nvme/models \
            -v /nvme/models:/nvme/models \
            ${{ secrets.REGISTRY }}/proj/app:${{ github.sha }} \
            python -c "import torch; assert torch.cuda.is_available(); print(torch.cuda.get_device_name(0))"

  perf-gates:
    needs: gpu-smoke
    runs-on: self-hosted-gpu
    steps:
      - name: Perf test
        run: |
          docker run --rm --gpus 1 ${{ secrets.REGISTRY }}/proj/app:${{ github.sha }} \
            python perf_gate.py --p95-ms 1200 --vram-headroom-gb 2.0