TorchBench Nightly #32
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 Nightly | |
| on: | |
| workflow_dispatch: | |
| inputs: | |
| test_pytorch-triton: | |
| default: false | |
| description: 'Run TorchBench with pytorch-triton before triton-nightly' | |
| required: false | |
| type: boolean | |
| workflow_run: | |
| workflows: ["Wheels"] | |
| types: [completed] | |
| env: | |
| TRITON_USE_ASSERT_ENABLED_LLVM: "TRUE" | |
| jobs: | |
| TorchBench-Nvidia: | |
| strategy: | |
| matrix: | |
| runner: [[self-hosted, A100], [self-hosted, H100]] | |
| runs-on: ${{ matrix.runner }} | |
| steps: | |
| - name: Clear cache | |
| run: | | |
| rm -rf ${HOME}/.triton | |
| rm -rf ${HOME}/torchbench | |
| - name: Checkout TorchBench | |
| uses: actions/checkout@v4 | |
| with: | |
| repository: 'pytorch/benchmark' | |
| ref: '7617d3f53f8ae8db938cda80173926968c7f0d4b' | |
| path: 'torchbench' | |
| submodules: 'recursive' | |
| - name: Create Python venv | |
| run: | | |
| python3 -m venv ${HOME}/torchbench/venv | |
| source ${HOME}/torchbench/venv/bin/activate | |
| echo PATH=${PATH} >> $GITHUB_ENV | |
| # Torch nightly version below should match TorchBench commit above | |
| - name: Install TorchBench and models | |
| run: | | |
| sudo apt-get install --yes libpango-1.0-0 | |
| python3 -m pip install --upgrade pip | |
| python3 -m pip install lightning-utilities pyre_extensions | |
| python3 -m pip install --upgrade torchaudio==2.2.0.dev20231017 torch==2.2.0.dev20231017 torchrec==0.5.0.dev20231017 torchvision==0.17.0.dev20231017 --index-url https://download.pytorch.org/whl/nightly/cu121 | |
| cd torchbench | |
| python3 install.py -v | |
| - name: Tune Nvidia GPU (A100) | |
| if: ${{ contains(matrix.runner, 'A100') }} | |
| run: | | |
| sudo nvidia-smi -pm 1 | |
| sudo nvidia-smi --lock-gpu-clocks=1350,1350 | |
| nvidia-smi | |
| - name: Tune Nvidia GPU (H100) | |
| if: ${{ contains(matrix.runner, 'H100') }} | |
| run: | | |
| echo "ENABLE_TMA=1" >> "${GITHUB_ENV}" | |
| nvidia-smi | |
| - name: Run TorchBench with pytorch-triton | |
| if: ${{ inputs.test_pytorch-triton }} | |
| run: | | |
| cd torchbench | |
| python3 run_benchmark.py dynamo --performance --inference --inductor --devices=cuda --output-directory ./ --output dynamo_pytorch-triton.csv --timing --stats 2>&1 | tee dynamo_log_pytorch-triton.log | |
| - name: Install the latest Triton wheel | |
| run: | | |
| python3 -m pip uninstall --yes pytorch-triton | |
| python3 -m pip install --upgrade triton-nightly --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/ | |
| - name: Run TorchBench with triton-nightly | |
| run: | | |
| cd torchbench | |
| python3 run_benchmark.py dynamo --performance --inference --inductor --devices=cuda --output-directory ./ --output dynamo_triton-nightly.csv --timing --stats 2>&1 | tee dynamo_log_triton-nightly.log | |
| - name: Create artifacts archive | |
| run: | | |
| cd torchbench | |
| tar -czf artifacts.tar.gz *.log *.csv | |
| - name: Upload artifacts archive | |
| uses: actions/upload-artifact@v2 | |
| with: | |
| name: TorchBench Dynamo Logs | |
| path: ${HOME}/torchbench/torchbench_dynamo_logs.tar.gz | |
| - name: Cleanup and reset GPU clocks | |
| if: always() | |
| run: | | |
| sudo nvidia-smi -i 0 -rgc |