self-push.yml

name: Self-hosted runner (push)

on:
  push:
    branches:
      - main
      - ci_*
      - ci-*
    paths:
      - "src/**"
      - "tests/**"
      - ".github/**"
      - "templates/**"
      - "utils/**"
  repository_dispatch:

env:
  HF_HOME: /mnt/cache
  TRANSFORMERS_IS_CI: yes
  OMP_NUM_THREADS: 8
  MKL_NUM_THREADS: 8
  PYTEST_TIMEOUT: 60
  TF_FORCE_GPU_ALLOW_GROWTH: true
  RUN_PT_TF_CROSS_TESTS: 1

jobs:
  setup:
    name: Setup
    runs-on: ubuntu-latest
    outputs:
      matrix: ${{ steps.set-matrix.outputs.matrix }}
      test_map: ${{ steps.set-matrix.outputs.test_map }}
    steps:
      - name: Checkout transformers
        uses: actions/checkout@v2
        with:
          fetch-depth: 2

      - name: Cleanup
        run: |
          rm -rf tests/__pycache__
          rm -rf tests/models/__pycache__
          rm -rf reports

      - name: Fetch the tests to run
        # TODO: add `git-python` in the docker images
        run: |
          pip install --upgrade git-python
          python utils/tests_fetcher.py --diff_with_last_commit | tee test_preparation.txt

      - name: Report fetched tests
        uses: actions/upload-artifact@v2
        with:
          name: test_fetched
          path: test_preparation.txt

      - id: set-matrix
        name: Organize tests into models
        # The `keys` is used as GitHub actions matrix for jobs, i.e. `models/bert`, `tokenization`, `pipeline`, etc.
        # The `test_map` is used to get the actual identified test files under each key.
        # If no test to run (so no `test_map.json` file), create a dummy map (empty matrix will fail)
        run: |
          if [ -f test_map.json ]; then
              keys=$(python3 -c 'import json; fp = open("test_map.json"); test_map = json.load(fp); fp.close(); d = list(test_map.keys()); print(d)')
              test_map=$(python3 -c 'import json; fp = open("test_map.json"); test_map = json.load(fp); fp.close(); print(test_map)')
          else
              keys=$(python3 -c 'keys = ["dummy"]; print(keys)')
              test_map=$(python3 -c 'test_map = {"dummy": []}; print(test_map)')
          fi
          echo $keys
          echo $test_map
          echo "::set-output name=matrix::$keys"
          echo "::set-output name=test_map::$test_map"

  run_tests_single_gpu:
    name: Model tests
    needs: setup
    # `dummy` means there is no test to run
    if: contains(fromJson(needs.setup.outputs.matrix), 'dummy') != true
    strategy:
      fail-fast: false
      matrix:
        folders: ${{ fromJson(needs.setup.outputs.matrix) }}
        machine_type: [single-gpu]
    runs-on: [self-hosted, docker-gpu, '${{ matrix.machine_type }}']
    container:
      image: huggingface/transformers-all-latest-gpu
      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
    steps:
      - name: NVIDIA-SMI
        run: |
          nvidia-smi

      - name: Are GPUs recognized by our DL frameworks
        working-directory: /transformers
        run: |
          utils/print_env_pt.py
          TF_CPP_MIN_LOG_LEVEL=3 python3 -c "import tensorflow as tf; print('TF GPUs available:', bool(tf.config.list_physical_devices('GPU')))"
          TF_CPP_MIN_LOG_LEVEL=3 python3 -c "import tensorflow as tf; print('Number of TF GPUs available:', len(tf.config.list_physical_devices('GPU')))"

      - name: Echo folder ${{ matrix.folders }}
        shell: bash
        # For folders like `models/bert`, set an env. var. (`matrix_folders`) to `models_bert`, which will be used to
        # set the artifact folder names (because the character `/` is not allowed).
        run: |
          echo "${{ matrix.folders }}"
          echo "${{ fromJson(needs.setup.outputs.test_map)[matrix.folders] }}"
          matrix_folders=${{ matrix.folders }}
          matrix_folders=${matrix_folders/'models/'/'models_'}
          echo "$matrix_folders"
          echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV

      - name: Update clone
        working-directory: /transformers
        run: git fetch && git checkout ${{ github.sha }}

      - name: Run all non-slow selected tests on GPU
        working-directory: /transformers
        run: |
          python3 -m pytest -n 2 --dist=loadfile -v --make-reports=${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }} ${{ fromJson(needs.setup.outputs.test_map)[matrix.folders] }}

      - name: Failure short reports
        if: ${{ failure() }}
        continue-on-error: true
        run: cat /transformers/reports/${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }}/failures_short.txt

      - name: Test suite reports artifacts
        if: ${{ always() }}
        uses: actions/upload-artifact@v2
        with:
          name: ${{ matrix.machine_type }}_run_all_tests_gpu_${{ env.matrix_folders }}_test_reports
          path: /transformers/reports/${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }}

  run_tests_multi_gpu:
    name: Model tests
    needs: setup
    # `dummy` means there is no test to run
    if: contains(fromJson(needs.setup.outputs.matrix), 'dummy') != true
    strategy:
      fail-fast: false
      matrix:
        folders: ${{ fromJson(needs.setup.outputs.matrix) }}
        machine_type: [multi-gpu]
    runs-on: [self-hosted, docker-gpu, '${{ matrix.machine_type }}']
    container:
      image: huggingface/transformers-all-latest-gpu
      options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
    steps:
      - name: NVIDIA-SMI
        run: |
          nvidia-smi

      - name: Are GPUs recognized by our DL frameworks
        working-directory: /transformers
        run: |
          utils/print_env_pt.py
          TF_CPP_MIN_LOG_LEVEL=3 python3 -c "import tensorflow as tf; print('TF GPUs available:', bool(tf.config.list_physical_devices('GPU')))"
          TF_CPP_MIN_LOG_LEVEL=3 python3 -c "import tensorflow as tf; print('Number of TF GPUs available:', len(tf.config.list_physical_devices('GPU')))"

      - name: Echo folder ${{ matrix.folders }}
        shell: bash
        # For folders like `models/bert`, set an env. var. (`matrix_folders`) to `models_bert`, which will be used to
        # set the artifact folder names (because the character `/` is not allowed).
        run: |
          echo "${{ matrix.folders }}"
          echo "${{ fromJson(needs.setup.outputs.test_map)[matrix.folders] }}"
          matrix_folders=${{ matrix.folders }}
          matrix_folders=${matrix_folders/'models/'/'models_'}
          echo "$matrix_folders"
          echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV

      - name: Update clone
        working-directory: /transformers
        run: git fetch && git checkout ${{ github.sha }}

      - name: Run all non-slow selected tests on GPU
        env:
          MKL_SERVICE_FORCE_INTEL: 1
        working-directory: /transformers
        run: |
          python3 -m pytest -n 2 --dist=loadfile -v --make-reports=${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }} ${{ fromJson(needs.setup.outputs.test_map)[matrix.folders] }}

      - name: Failure short reports
        if: ${{ failure() }}
        continue-on-error: true
        run: cat /transformers/reports/${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }}/failures_short.txt

      - name: Test suite reports artifacts
        if: ${{ always() }}
        uses: actions/upload-artifact@v2
        with:
          name: ${{ matrix.machine_type }}_run_all_tests_gpu_${{ env.matrix_folders }}_test_reports
          path: /transformers/reports/${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }}

  run_tests_torch_cuda_extensions_single_gpu:
    name: Torch CUDA extension tests
    needs: setup
    if: contains(fromJson(needs.setup.outputs.matrix), 'deepspeed') || contains(fromJson(needs.setup.outputs.matrix), 'extended')
    strategy:
      fail-fast: false
      matrix:
        machine_type: [single-gpu]
    runs-on: [self-hosted, docker-gpu, '${{ matrix.machine_type }}']
    container:
      image: nvcr.io/nvidia/pytorch:21.03-py3
      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
    steps:
      - name: Checkout transformers
        uses: actions/checkout@v2
        with:
          fetch-depth: 2

      - name: NVIDIA-SMI
        run: |
          nvidia-smi

      - name: Install dependencies
        run: |
          apt -y update && apt install -y libaio-dev
          pip install --upgrade pip
          pip install .[deepspeed-testing]

      - name: Are GPUs recognized by our DL frameworks
        run: |
          utils/print_env_pt.py

      - name: Run all non-slow selected tests on GPU
        # TODO: Here we pass all tests in the 2 folders for simplicity. It's better to pass only the identified tests.
        run: |
          python -m pytest -n 1 --dist=loadfile -v --make-reports=${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu tests/deepspeed tests/extended

      - name: Failure short reports
        if: ${{ failure() }}
        continue-on-error: true
        run: cat reports/${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu/failures_short.txt

      - name: Test suite reports artifacts
        if: ${{ always() }}
        uses: actions/upload-artifact@v2
        with:
          name: ${{ matrix.machine_type }}_run_tests_torch_cuda_extensions_gpu_test_reports
          path: reports/${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu

  run_tests_torch_cuda_extensions_multi_gpu:
    name: Torch CUDA extension tests
    needs: setup
    if: contains(fromJson(needs.setup.outputs.matrix), 'deepspeed') || contains(fromJson(needs.setup.outputs.matrix), 'extended')
    strategy:
      fail-fast: false
      matrix:
        machine_type: [multi-gpu]
    runs-on: [self-hosted, docker-gpu, '${{ matrix.machine_type }}']
    container:
      image: nvcr.io/nvidia/pytorch:21.03-py3
      options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
    steps:
      - name: Checkout transformers
        uses: actions/checkout@v2
        with:
          fetch-depth: 2

      - name: NVIDIA-SMI
        run: |
          nvidia-smi

      - name: Install dependencies
        run: |
          apt -y update && apt install -y libaio-dev
          pip install --upgrade pip
          rm -rf ~/.cache/torch_extensions/ # shared between conflicting builds
          pip install .[testing,deepspeed,fairscale]

      - name: Are GPUs recognized by our DL frameworks
        run: |
          utils/print_env_pt.py

      - name: Run all non-slow selected tests on GPU
        # TODO: Here we pass all tests in the 2 folders for simplicity. It's better to pass only the identified tests.
        run: |
          python -m pytest -n 1 --dist=loadfile -v --make-reports=${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu tests/deepspeed tests/extended

      - name: Failure short reports
        if: ${{ failure() }}
        continue-on-error: true
        run: cat reports/${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu/failures_short.txt

      - name: Test suite reports artifacts
        if: ${{ always() }}
        uses: actions/upload-artifact@v2
        with:
          name: ${{ matrix.machine_type }}_run_tests_torch_cuda_extensions_gpu_test_reports
          path: reports/${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu

  send_results:
    name: Send results to webhook
    runs-on: ubuntu-latest
    if: always()
    needs: [
        setup,
        run_tests_single_gpu,
        run_tests_multi_gpu,
        run_tests_torch_cuda_extensions_single_gpu,
        run_tests_torch_cuda_extensions_multi_gpu
    ]
    steps:
      - uses: actions/checkout@v2
      - uses: actions/download-artifact@v2
      - name: Send message to Slack
        env:
          CI_SLACK_BOT_TOKEN: ${{ secrets.CI_SLACK_BOT_TOKEN }}
          CI_SLACK_CHANNEL_ID: ${{ secrets.CI_SLACK_CHANNEL_ID }}
          CI_SLACK_CHANNEL_ID_DAILY: ${{ secrets.CI_SLACK_CHANNEL_ID_DAILY }}
          CI_SLACK_CHANNEL_DUMMY_TESTS: ${{ secrets.CI_SLACK_CHANNEL_DUMMY_TESTS }}
          CI_SLACK_REPORT_CHANNEL_ID: ${{ secrets.CI_SLACK_CHANNEL_ID }}
          CI_EVENT: push
          CI_TITLE: ${{ github.event.head_commit.message }}
          CI_COMMIT_URL: ${{ github.event.head_commit.url }}
        # We pass `needs.setup.outputs.matrix` as the argument. A processing in `notification_service.py` to change
        # `models/bert` to `models_bert` is required, as the artifact names use `_` instead of `/`.
        run: |
          pip install slack_sdk
          python utils/notification_service.py "${{ needs.setup.outputs.matrix }}"