linux_cuda_wheel.yaml

name: Build and test Linux CUDA wheels

on:
  pull_request:
  push:
    branches:
      - nightly
      - main
      - release/*
    tags:
        - v[0-9]+.[0-9]+.[0-9]+-rc[0-9]+
  workflow_dispatch:

concurrency:
  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
  cancel-in-progress: true

permissions:
  id-token: write
  contents: write

defaults:
  run:
    shell: bash -l -eo pipefail {0}

jobs:
  generate-matrix:
    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
    with:
      package-type: wheel
      os: linux
      test-infra-repository: pytorch/test-infra
      test-infra-ref: main
      with-cpu: disable
      with-xpu: disable
      with-rocm: disable
      with-cuda: enable
      build-python-only: "disable"
  build:
    needs: generate-matrix
    strategy:
      fail-fast: false
    name: Build and Upload wheel
    uses: pytorch/test-infra/.github/workflows/build_wheels_linux.yml@main
    with:
      repository: pytorch/torchcodec
      ref: ""
      test-infra-repository: pytorch/test-infra
      test-infra-ref: main
      build-matrix: ${{ needs.generate-matrix.outputs.matrix }}
      post-script: packaging/post_build_script.sh
      smoke-test-script: packaging/fake_smoke_test.py
      package-name: torchcodec
      trigger-event: ${{ github.event_name }}
      build-platform: "python-build-package"
      build-command: "BUILD_AGAINST_ALL_FFMPEG_FROM_S3=1 ENABLE_CUDA=1 python -m build --wheel -vvv --no-isolation"

  install-and-test:
    runs-on: linux.4xlarge.nvidia.gpu
    strategy:
      fail-fast: false
      matrix:
          # 3.9 corresponds to the minimum python version for which we build
          # the wheel unless the label cliflow/binaries/all is present in the
          # PR.
          # For the actual release we should add that label and change this to
          # include more python versions.
        python-version: ['3.9']
        cuda-version: ['11.8', '12.1', '12.4']
        ffmpeg-version-for-tests: ['5', '6', '7']
    container:
      image: "pytorch/manylinux-builder:cuda${{ matrix.cuda-version }}"
      options: "--gpus all -e NVIDIA_DRIVER_CAPABILITIES=video,compute,utility"
    needs: build
    steps:
      - name: Setup env vars
        run: |
          cuda_version_without_periods=$(echo "${{ matrix.cuda-version }}" | sed 's/\.//g')
          echo cuda_version_without_periods=${cuda_version_without_periods} >> $GITHUB_ENV
      - uses: actions/download-artifact@v3
        with:
          name: pytorch_torchcodec__3.9_cu${{ env.cuda_version_without_periods }}_x86_64
          path: pytorch/torchcodec/dist/
      - name: Setup miniconda using test-infra
        uses: pytorch/test-infra/.github/actions/setup-miniconda@main
        with:
          python-version: ${{ matrix.python-version }}
          #
          # For some reason nvidia::libnpp=12.4 doesn't install but nvidia/label/cuda-12.4.0::libnpp does.
          # So we use the latter convention for libnpp.
          # We install conda packages at the start because otherwise conda may have conflicts with dependencies.
          default-packages: "nvidia/label/cuda-${{ matrix.cuda-version }}.0::libnpp nvidia::cuda-nvrtc=${{ matrix.cuda-version }} nvidia::cuda-toolkit=${{ matrix.cuda-version }} nvidia::cuda-cudart=${{ matrix.cuda-version }} nvidia::cuda-driver-dev=${{ matrix.cuda-version }} conda-forge::ffmpeg=${{ matrix.ffmpeg-version-for-tests }}"
      - name: Check env
        run: |
          ${CONDA_RUN} env
          ${CONDA_RUN} conda info
          ${CONDA_RUN} nvidia-smi
          ${CONDA_RUN} conda list
      - name: Assert ffmpeg exists
        run: |
          ${CONDA_RUN} ffmpeg -buildconf
      - name: Update pip
        run: ${CONDA_RUN} python -m pip install --upgrade pip
      - name: Install PyTorch
        run: |
          ${CONDA_RUN} python -m pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu${{ env.cuda_version_without_periods }}
          ${CONDA_RUN} python -c 'import torch; print(f"{torch.__version__}"); print(f"{torch.__file__}"); print(f"{torch.cuda.is_available()=}")'
      - name: Install torchcodec from the wheel
        run: |
          wheel_path=`find pytorch/torchcodec/dist -type f -name "*.whl"`
          echo Installing $wheel_path
          ${CONDA_RUN} python -m pip install $wheel_path -vvv

      - name: Check out repo
        uses: actions/checkout@v3

      - name: Install test dependencies
        run: |
          # Ideally we would find a way to get those dependencies from pyproject.toml
          ${CONDA_RUN} python -m pip install numpy pytest pillow

      - name: Delete the src/ folder just for fun
        run: |
          # The only reason we checked-out the repo is to get access to the
          # tests. We don't care about the rest. Out of precaution, we delete
          # the src/ folder to be extra sure that we're running the code from
          # the installed wheel rather than from the source.
          # This is just to be extra cautious and very overkill because a)
          # there's no way the `torchcodec` package from src/ can be found from
          # the PythonPath: the main point of `src/` is precisely to protect
          # against that and b) if we ever were to execute code from
          # `src/torchcodec`, it would fail loudly because the built .so files
          # aren't present there.
          rm -r src/
          ls
      - name: Smoke test
        run: |
          ${CONDA_RUN} python test/decoders/manual_smoke_test.py
      - name: Run Python tests
        run: |
          ${CONDA_RUN} FAIL_WITHOUT_CUDA=1 pytest test -vvv
      - name: Run Python benchmark
        run: |
          ${CONDA_RUN} time python benchmarks/decoders/gpu_benchmark.py --devices=cuda:0,cpu --resize_devices=none