install-test.yml

name: Test Package Installation

on:
  pull_request:
    branches: [main]
    paths:
      - 'pyproject.toml'
      - 'pyproject.vllm.toml'
      - 'uv.lock'
      - 'uv.vllm.lock'
      - 'areal/**'
      - '.github/workflows/install-test.yml'
  push:
    branches: [main]
    paths:
      - 'pyproject.toml'
      - 'pyproject.vllm.toml'
      - 'uv.lock'
      - 'uv.vllm.lock'
      - 'areal/**'
      - '.github/workflows/install-test.yml'
  workflow_dispatch:

concurrency:
  group: install-test-${{ github.ref }}
  cancel-in-progress: true

jobs:
  install-test:
    strategy:
      fail-fast: false
      matrix:
        os: [ubuntu-latest, macos-latest]
        python-version: ['3.12']
    runs-on: ${{ matrix.os }}
    name: Install test (${{ matrix.os }}, Python ${{ matrix.python-version }})

    steps:
      - name: Checkout repository
        uses: actions/checkout@v6

      - name: Install uv
        uses: astral-sh/setup-uv@v7
        with:
          enable-cache: true
          cache-dependency-glob: |
            uv.lock
            uv.vllm.lock

      - name: Set up Python
        run: uv python install ${{ matrix.python-version }}

      - name: Install package (basic)
        run: uv sync

      - name: Verify package import
        run: |
          uv run python -c "import areal; print(f'areal version: {areal.__version__}')"

      - name: Verify core modules are importable
        run: |
          uv run python -c "
          from areal import (
              TrainController,
              RolloutController,
              WorkflowExecutor,
              StalenessManager,
              workflow_context,
              current_platform,
          )
          print('All core modules imported successfully')
          "

      - name: Build wheel
        run: uv build --wheel

      - name: Verify wheel artifact
        run: |
          ls -la dist/
          python -m zipfile -l dist/*.whl | head -20

  install-test-cuda-extras:
    # Test CUDA extras installation on Linux only (CUDA packages have Linux-only wheels)
    # Note: flash-attn is excluded because it requires CUDA compilation infrastructure
    # which is not available on standard GitHub runners
    # sglang and vllm are conflicting extras, so we test each separately
    runs-on: ubuntu-latest
    strategy:
      fail-fast: false
      matrix:
        variant: [sglang, vllm]
    name: Install test (Linux, CUDA extras, ${{ matrix.variant }})

    steps:
      - name: Checkout repository
        uses: actions/checkout@v6

      - name: Install uv
        uses: astral-sh/setup-uv@v7
        with:
          enable-cache: true
          cache-dependency-glob: |
            uv.lock
            uv.vllm.lock

      - name: Set up Python
        run: uv python install 3.12

      - name: Install package with CUDA extras (excluding flash-attn)
        # flash-attn requires CUDA toolkit for compilation, skip it in CI
        # Test individual extras that have pre-built wheels
        run: |
          if [ "${{ matrix.variant }}" = "vllm" ]; then
            cp pyproject.vllm.toml pyproject.toml
            cp uv.vllm.lock uv.lock
          fi
          uv sync --extra ${{ matrix.variant }} --extra megatron --extra tms

      - name: Verify package import with CUDA extras
        run: |
          uv run python -c "import areal; print(f'areal version: {areal.__version__}')"

      - name: Verify CUDA-dependent packages are installed
        env:
          VARIANT: ${{ matrix.variant }}
        run: |
          uv run python -c "
          import os, importlib.util
          variant = os.environ['VARIANT']
          packages = [variant, 'megatron.core']
          for pkg in packages:
              spec = importlib.util.find_spec(pkg)
              status = 'installed' if spec else 'not found'
              print(f'{pkg}: {status}')
          "

  install-test-docker:
    # Test editable installation in Docker runtime image
    # This mirrors the installation method documented for Docker users
    # Only runs on manual dispatch due to large image pull time
    # and `no space on device left` errors on busy runners
    if: github.event_name == 'workflow_dispatch'
    runs-on: ubuntu-latest
    strategy:
      fail-fast: false
      matrix:
        variant: [sglang, vllm]
    name: Install test (Docker ${{ matrix.variant }} image)
    container:
      image: ghcr.io/inclusionai/areal-runtime:dev-${{ matrix.variant }}

    steps:
      - name: Checkout repository
        uses: actions/checkout@v6

      - name: Install package (editable, no deps)
        # This is the documented Docker installation method
        # Dependencies are pre-installed in the runtime image
        run: uv pip install -e . --no-deps

      - name: Verify package import
        run: |
          python -c "import areal; print(f'areal version: {areal.__version__}')"

      - name: Verify core modules are importable
        run: |
          python -c "
          from areal import (
              TrainController,
              RolloutController,
              WorkflowExecutor,
              StalenessManager,
              workflow_context,
              current_platform,
          )
          print('All core modules imported successfully')
          "

      - name: Run Docker installation validation
        run: |
          python areal/tools/validate_docker_installation.py