TorchBench V3 nightly (A100) #74
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| name: TorchBench V3 nightly (A100) | |
| on: | |
| workflow_dispatch: | |
| schedule: | |
| - cron: '00 18 * * *' # run at 6:00 PM UTC, K8s containers will roll out at 12PM EST | |
| jobs: | |
| run-benchmark: | |
| environment: docker-s3-upload | |
| env: | |
| BASE_CONDA_ENV: "torchbench" | |
| CONDA_ENV: "torchbench-v3-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: ${{ github.repository_owner == 'pytorch' }} | |
| runs-on: [a100-runner] | |
| steps: | |
| - name: Checkout TorchBench v3.0 branch | |
| uses: actions/checkout@v3 | |
| with: | |
| ref: v3.0 | |
| path: benchmark | |
| - name: Tune Nvidia GPU | |
| run: | | |
| sudo nvidia-smi -pm 1 | |
| sudo nvidia-smi -ac 1215,1410 | |
| nvidia-smi | |
| - name: Clone and setup conda env | |
| run: | | |
| CONDA_ENV=${BASE_CONDA_ENV} . "${SETUP_SCRIPT}" | |
| conda create --name "${CONDA_ENV}" --clone "${BASE_CONDA_ENV}" | |
| - name: Install TorchBench | |
| run: | | |
| set -x | |
| . "${SETUP_SCRIPT}" | |
| pushd benchmark | |
| python install.py | |
| - name: Run the torch-nightly userbenchmark | |
| run: | | |
| . "${SETUP_SCRIPT}" | |
| # remove old results | |
| if [ -d benchmark-output ]; then rm -Rf benchmark-output; fi | |
| pushd benchmark | |
| if [ -d .userbenchmark ]; then rm -Rf .userbenchmark; fi | |
| python run_benchmark.py torch-nightly -c v3-cuda-tests.yaml | |
| cp -r ./.userbenchmark/torch-nightly ../benchmark-output | |
| - name: Detect potential regressions | |
| continue-on-error: true | |
| run: | | |
| . "${SETUP_SCRIPT}" | |
| pushd benchmark | |
| RESULTS=($(find ${PWD}/../benchmark-output -name "metrics-*.json" -maxdepth 2 | sort -r)) | |
| # TODO: the following assumes only one metrics-*.json is found. It will keep | |
| # overwriting gh-issue.md if multiple are found. | |
| for r in ${RESULTS[@]}; do | |
| python regression_detector.py --platform "${PLATFORM_NAME}" --treatment "${r}" --owner @xuzhao9 \ | |
| --gh-issue-path gh-issue.md --errors-path errors.txt | |
| done | |
| rm -r ../benchmark-output || true | |
| cp -r ./.userbenchmark/torch-nightly ../benchmark-output | |
| - name: Copy artifact and upload to scribe and Amazon S3 | |
| run: | | |
| . "${SETUP_SCRIPT}" | |
| pushd benchmark | |
| LATEST_RESULT=$(find ../benchmark-output/ -name "metrics-*.json" | sort -r | head -1) | |
| echo "Benchmark result file: ${LATEST_RESULT}" | |
| # Upload the result json to Scribe | |
| python ./scripts/userbenchmark/upload_scribe.py --userbenchmark_json "${LATEST_RESULT}" --userbenchmark_platform "${PLATFORM_NAME}" | |
| # Upload the result json to Amazon S3 | |
| python ./scripts/userbenchmark/upload_s3.py --upload-file "${LATEST_RESULT}" --userbenchmark_platform "${PLATFORM_NAME}" | |
| - name: Copy regression results to Amazon S3 and kick off bisection | |
| if: env.TORCHBENCH_REGRESSION_DETECTED | |
| run: | | |
| . "${SETUP_SCRIPT}" | |
| pushd benchmark | |
| LATEST_REGRESSION_RESULT=$(find ../benchmark-output/ -name "regression-*.yaml" | sort -r | head -1) | |
| # Upload the regression json to Amazon S3 | |
| python ./scripts/userbenchmark/upload_s3.py --upload-file "${LATEST_REGRESSION_RESULT}" --userbenchmark_platform "${PLATFORM_NAME}" | |
| # Get the workflow ID from | |
| # https://api.github.com/repos/pytorch/benchmark/actions/workflows | |
| # And dispatch the bisection workflow | |
| curl -u xuzhao9:${{ secrets.TORCHBENCH_ACCESS_TOKEN }} \ | |
| -X POST \ | |
| -H "Accept: application/vnd.github.v3+json" \ | |
| https://api.github.com/repos/pytorch/benchmark/actions/workflows/57994037/dispatches \ | |
| -d '{"ref": "main", "inputs": {"regression_date": "${{ env.TORCHBENCH_REGRESSION_DETECTED }}" } }' | |
| - name: Upload result to GH Actions Artifact | |
| uses: actions/upload-artifact@v3 | |
| with: | |
| name: TorchBench V3 result | |
| path: benchmark-output/ | |
| - name: Clean up Conda env | |
| if: always() | |
| run: | | |
| . "${SETUP_SCRIPT}" | |
| conda deactivate && conda deactivate | |
| conda remove -n "${CONDA_ENV}" --all |