torchao.yml

name: Torchao nightly workflow (A100)
on:
  workflow_dispatch:
  # TODO: REMOVE THIS BEFORE THE PR IS MERGED!
  pull_request:

jobs:
  run-benchmark:
    environment: docker-s3-upload
    env:
      BASE_CONDA_ENV: "torchbench"
      CONDA_ENV:  "torchao-nightly"
      PLATFORM_NAME: "gcp_a100"
      SETUP_SCRIPT: "/workspace/setup_instance.sh"
      TORCHBENCH_USERBENCHMARK_SCRIBE_GRAPHQL_ACCESS_TOKEN: ${{ secrets.TORCHBENCH_USERBENCHMARK_SCRIBE_GRAPHQL_ACCESS_TOKEN }}
      AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
      AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
      IS_GHA: 1
      BUILD_ENVIRONMENT: benchmark-nightly
    if: ${{ github.repository_owner == 'pytorch' }}
    runs-on: [a100-runner]
    steps:
      - name: Checkout TorchBench
        uses: actions/checkout@v3
        with:
          path: benchmark
      - name: Tune Nvidia GPU
        run: |
          sudo nvidia-smi -pm 1
          sudo nvidia-smi -ac 1215,1410
          nvidia-smi
      - name: Clone and setup conda env
        run: |
          CONDA_ENV=${BASE_CONDA_ENV} . "${SETUP_SCRIPT}"
          conda create --name "${CONDA_ENV}" --clone "${BASE_CONDA_ENV}"
      - name: Run the torchao userbenchmark
        run: |
          . "${SETUP_SCRIPT}"
          # remove old results if exists
          if [ -d benchmark-output ]; then rm -Rf benchmark-output; fi
          pushd benchmark
          if [ -d .userbenchmark ]; then rm -Rf .userbenchmark; fi
          # Install torchao
          python install.py torchao
          python run_benchmark.py torchao --ci
          cp -r ./.userbenchmark/torchao ../benchmark-output
      - name: Upload result to GH Actions Artifact
        uses: actions/upload-artifact@v3
        with:
          name: Torchao nightly result
          path: benchmark-output/
      - name: Clean up Conda env
        if: always()
        run: |
          . "${SETUP_SCRIPT}"
          conda deactivate && conda deactivate
          conda remove -n "${CONDA_ENV}" --all