TorchBench V2 nightly #76
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 V2 nightly | |
| on: | |
| workflow_dispatch: | |
| schedule: | |
| - cron: '0 14 * * *' # run at 2 PM UTC | |
| jobs: | |
| run-benchmark: | |
| environment: docker-s3-upload | |
| env: | |
| TORCHBENCH_VER: "v2" | |
| CONFIG_VER: "v2" | |
| CONDA_ENV_NAME: "torchbench-v2-nightly-ci" | |
| OUTPUT_DIR: ".torchbench/v2-nightly-ci" | |
| BISECTION_ROOT: ".torchbench/v2-bisection-ci" | |
| SCRIBE_GRAPHQL_ACCESS_TOKEN: ${{ secrets.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 | |
| AWS_DEFAULT_REGION: us-east-1 | |
| BUILD_ENVIRONMENT: benchmark-nightly | |
| SETUP_SCRIPT: "/data/nvme/bin/setup_instance.sh" | |
| if: ${{ github.repository_owner == 'pytorch' }} | |
| runs-on: [self-hosted, bm-runner] | |
| steps: | |
| - name: Checkout | |
| uses: actions/checkout@v3 | |
| with: | |
| ref: v2.0 | |
| - name: Create conda env | |
| run: | | |
| python3 ./utils/python_utils.py --create-conda-env ${CONDA_ENV_NAME} | |
| - name: Install PyTorch nightly | |
| run: | | |
| . activate "${CONDA_ENV_NAME}" | |
| . "${SETUP_SCRIPT}" | |
| python utils/cuda_utils.py --install-torch-deps | |
| python utils/cuda_utils.py --install-torch-nightly | |
| - name: Install Torchbench models | |
| run: | | |
| . activate "${CONDA_ENV_NAME}" | |
| . "${SETUP_SCRIPT}" | |
| python install.py | |
| - name: Run benchmark | |
| run: | | |
| . activate "${CONDA_ENV_NAME}" | |
| . "${SETUP_SCRIPT}" | |
| WORKFLOW_HOME="${HOME}/${{ env.OUTPUT_DIR }}/gh${GITHUB_RUN_ID}" | |
| bash ./.github/scripts/run.sh "${WORKFLOW_HOME}" | |
| - name: Generate the bisection config | |
| run: | | |
| set -x | |
| . activate "${CONDA_ENV_NAME}" | |
| . "${SETUP_SCRIPT}" | |
| WORKFLOW_HOME="${HOME}/${{ env.OUTPUT_DIR }}/gh${GITHUB_RUN_ID}" | |
| mkdir -p benchmark-output/ | |
| # Update the self-hosted pytorch version | |
| pushd "${HOME}/pytorch" | |
| git fetch origin | |
| popd | |
| pip install gitpython pyyaml dataclasses argparse | |
| # Compare the result from yesterday and report any perf signals | |
| python ./.github/scripts/generate-abtest-config.py \ | |
| --pytorch-dir "${HOME}/pytorch" \ | |
| --github-issue "${WORKFLOW_HOME}/gh-issue.md" \ | |
| --benchmark-dir "${WORKFLOW_HOME}" \ | |
| --out "${WORKFLOW_HOME}/bisection.yaml" | |
| # Include in the GitHub artifact | |
| if [ -f "${WORKFLOW_HOME}/gh-issue.md" ]; then | |
| cp "${WORKFLOW_HOME}/bisection.yaml" ./benchmark-output/ | |
| cp "${WORKFLOW_HOME}/gh-issue.md" ./benchmark-output/ | |
| # Setup the bisection environment | |
| BISECTION_HOME="${HOME}/${{ env.BISECTION_ROOT }}/bisection-gh${GITHUB_RUN_ID}" | |
| mkdir -p "${BISECTION_HOME}" | |
| mv ./benchmark-output/gh-issue.md "${BISECTION_HOME}/gh-issue.md" | |
| cp ./benchmark-output/bisection.yaml "${BISECTION_HOME}/config.yaml" | |
| fi | |
| - name: Dispatch the bisection workflow | |
| if: env.TORCHBENCH_PERF_SIGNAL | |
| run: | | |
| # Get the workflow ID from | |
| # https://api.github.com/repos/pytorch/benchmark/actions/workflows | |
| curl -u xuzhao9:${{ secrets.TORCHBENCH_ACCESS_TOKEN }} \ | |
| -X POST \ | |
| -H "Accept: application/vnd.github.v3+json" \ | |
| https://api.github.com/repos/pytorch/benchmark/actions/workflows/16176850/dispatches \ | |
| -d '{"ref": "main", "inputs": {"issue_name": "bisection-gh'"${GITHUB_RUN_ID}"'" } }' | |
| - name: Copy artifact and upload to scribe | |
| run: | | |
| . activate "${CONDA_ENV_NAME}" | |
| TODAY=$(date "+%Y%m%d%H%M%S") | |
| LATEST_RESULT=$(find ${HOME}/${{ env.OUTPUT_DIR }}/gh${GITHUB_RUN_ID} -name "*.json" | sort -r | head -1) | |
| echo "Benchmark result file: ${LATEST_RESULT}" | |
| mkdir -p benchmark-output/ | |
| cp "${LATEST_RESULT}" ./benchmark-output/benchmark-result-${CONFIG_VER}-${TODAY}.json | |
| # Load environment variables | |
| CONFIG_DIR=torchbenchmark/score/configs/${CONFIG_VER} | |
| CONFIG_ENV=${CONFIG_DIR}/config-${CONFIG_VER}.env | |
| # Load environment variables | |
| set -a; source "${CONFIG_ENV}"; set +a | |
| SCORE_FILE="./benchmark-result-${CONFIG_VER}-score-${TODAY}.json" | |
| # Generate score file | |
| python compute_score.py --score_version "${CONFIG_VER}" --benchmark_data_file "${LATEST_RESULT}" --output-json "${SCORE_FILE}" | |
| # Upload result to Scribe | |
| python scripts/upload_scribe_${CONFIG_VER}.py --pytest_bench_json "${LATEST_RESULT}" --torchbench_score_file "${SCORE_FILE}" | |
| - name: Upload artifact | |
| uses: actions/upload-artifact@v3 | |
| with: | |
| name: Benchmark result | |
| path: benchmark-output/ | |
| - name: Destroy conda env | |
| run: | | |
| conda env remove --name "${CONDA_ENV_NAME}" |