E2E performance #130
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: E2E performance | |
| on: | |
| workflow_dispatch: | |
| inputs: | |
| install_ipex: | |
| description: Install Intel PyTorch Extension | |
| type: boolean | |
| default: true | |
| suite: | |
| description: Test suite | |
| type: choice | |
| options: | |
| - all | |
| - huggingface | |
| - timm_models | |
| - torchbench | |
| default: all | |
| mode: | |
| description: Inference/Training | |
| type: choice | |
| options: | |
| - all | |
| - inference | |
| - training | |
| default: all | |
| dtype: | |
| description: Data type | |
| type: choice | |
| options: | |
| - all | |
| - amp_bf16 | |
| - amp_fp16 | |
| - bfloat16 | |
| - float16 | |
| - float32 | |
| default: all | |
| models: | |
| description: Run all models or a subset | |
| type: choice | |
| options: | |
| - all | |
| - subset | |
| default: subset | |
| only_one_model: | |
| description: Run only this one model | |
| type: string | |
| default: "" | |
| runner_label: | |
| description: Runner label, keep empty for default | |
| type: string | |
| default: "" | |
| TORCH_COMPILE_DEBUG: | |
| description: TORCH_COMPILE_DEBUG | |
| type: string | |
| default: "" | |
| permissions: read-all | |
| env: | |
| TRITON_DISABLE_LINE_INFO: 1 | |
| USE_AOT_DEVLIST: pvc | |
| PYTHON_VERSION: "3.10" | |
| BENCHMARK_REPO: weishi-deng/benchmark.git | |
| BENCHMARK_BRANCH: main | |
| INSTALL_IPEX: ${{ inputs.install_ipex }} | |
| TORCH_XPU_OPS_COMMIT: 39522db63ce045f52c9d61a286018c266cd00479 | |
| jobs: | |
| setup: | |
| name: Setup | |
| runs-on: | |
| - glados | |
| - spr | |
| - cpu | |
| outputs: | |
| suite: ${{ steps.set-matrix.outputs.suite }} | |
| mode: ${{ steps.set-matrix.outputs.mode }} | |
| dtype: ${{ steps.set-matrix.outputs.dtype }} | |
| models: ${{ steps.set-matrix.outputs.models }} | |
| timeout-minutes: 10 | |
| steps: | |
| - name: Set matrix | |
| id: set-matrix | |
| run: | | |
| if [[ -z "${{ inputs.suite }}" || "${{ inputs.suite }}" == "all" ]]; then | |
| suite='["huggingface", "timm_models", "torchbench"]' | |
| else | |
| suite='["${{ inputs.suite }}"]' | |
| fi | |
| if [[ -z "${{ inputs.mode }}" || "${{ inputs.mode }}" == "all" ]]; then | |
| mode='["inference", "training"]' | |
| else | |
| mode='["${{ inputs.mode }}"]' | |
| fi | |
| if [[ -z "${{ inputs.dtype }}" || "${{ inputs.dtype }}" == "all" ]]; then | |
| dtype='["amp_bf16", "amp_fp16", "bfloat16", "float16", "float32"]' | |
| else | |
| dtype='["${{ inputs.dtype }}"]' | |
| fi | |
| if [[ -z "${{ inputs.models }}" ]]; then | |
| models="subset" | |
| else | |
| models="${{ inputs.models }}" | |
| fi | |
| echo "suite=$suite" >> $GITHUB_OUTPUT | |
| echo "mode=$mode" >> $GITHUB_OUTPUT | |
| echo "dtype=$dtype" >> $GITHUB_OUTPUT | |
| echo "models=$models" >> $GITHUB_OUTPUT | |
| print_inputs: | |
| name: Print inputs | |
| needs: setup | |
| runs-on: Linux | |
| steps: | |
| - name: Print inputs | |
| run: | | |
| echo "${{ toJSON(github.event.inputs) }}" | |
| - name: Print setup outputs | |
| run: | | |
| echo "${{ toJSON(needs.setup.outputs) }}" | |
| build: | |
| name: Test | |
| needs: setup | |
| runs-on: | |
| - ${{ inputs.runner_label || 'max1550' }} | |
| strategy: | |
| matrix: | |
| suite: ${{ fromJson(needs.setup.outputs.suite) }} | |
| mode: ${{ fromJson(needs.setup.outputs.mode) }} | |
| dtype: ${{ fromJson(needs.setup.outputs.dtype) }} | |
| fail-fast: false | |
| timeout-minutes: 720 | |
| defaults: | |
| run: | |
| shell: bash -noprofile --norc -eo pipefail -c "source /home/runner/intel/oneapi/setvars.sh > /dev/null; source {0}" | |
| steps: | |
| - name: Checkout repository | |
| uses: actions/checkout@v4 | |
| - name: Get benchmark commit id | |
| uses: ./.github/actions/get-commit-id | |
| with: | |
| repository: ${{ env.BENCHMARK_REPO }} | |
| branch: ${{ env.BENCHMARK_BRANCH }} | |
| variable: BENCHMARK_COMMIT_ID | |
| - name: Load pip cache | |
| id: pip-cache | |
| uses: ./.github/actions/load | |
| with: | |
| path: $HOME/.cache/pip | |
| # pip cache per commit id just to minimize network traffic | |
| key: pip-$PYTHON_VERSION-$GITHUB_SHA | |
| - name: Install Python | |
| uses: actions/setup-python@v5 | |
| with: | |
| python-version: ${{ env.PYTHON_VERSION }} | |
| - name: Install Python build dependencies | |
| run: | | |
| pip install wheel | |
| - name: Setup PyTorch | |
| if: ${{ env.INSTALL_IPEX == 'true' }} | |
| uses: ./.github/actions/setup-pytorch | |
| - name: Setup PyTorch without IPEX | |
| if: ${{ env.INSTALL_IPEX == 'false' }} | |
| uses: ./.github/actions/setup-pytorch | |
| with: | |
| repository: pytorch/pytorch | |
| torch_xpu_ops_commit: ${{ env.TORCH_XPU_OPS_COMMIT }} | |
| - name: Install latest nightly wheels including IPEX | |
| if: ${{ env.INSTALL_IPEX == 'true' }} | |
| uses: ./.github/actions/install-wheels | |
| with: | |
| gh_token: ${{ secrets.GITHUB_TOKEN }} | |
| python_version: ${{ env.PYTHON_VERSION }} | |
| - name: Install latest nightly wheels excluding IPEX | |
| if: ${{ env.INSTALL_IPEX == 'false' }} | |
| uses: ./.github/actions/install-wheels | |
| with: | |
| gh_token: ${{ secrets.GITHUB_TOKEN }} | |
| python_version: ${{ env.PYTHON_VERSION }} | |
| wheels_pattern: "!(intel_extension_for_pytorch-*|torch-*).whl" | |
| - name: Setup fake IPEX | |
| if: ${{ env.INSTALL_IPEX == 'false' }} | |
| uses: ./.github/actions/setup-fake-ipex | |
| - name: Identify pinned versions | |
| run: | | |
| cd pytorch | |
| echo "TRANSFORMERS_VERSION=$(<.ci/docker/ci_commit_pins/huggingface.txt)" >> "${GITHUB_ENV}" | |
| echo "TIMM_COMMIT_ID=$(<.ci/docker/ci_commit_pins/timm.txt)" >> "${GITHUB_ENV}" | |
| # TIMM depends on torch and torchvision, so, in general, it needs to be installed before | |
| # installing custom torch and torchvision, but instead we install its dependencies except | |
| # torch and torchvision. | |
| - name: Install TIMM Models | |
| if: ${{ matrix.suite == 'timm_models' }} | |
| run: | | |
| # install timm without dependencies | |
| pip install --no-deps git+https://github.com/huggingface/pytorch-image-models@$TIMM_COMMIT_ID | |
| # install timm dependencies without torch and torchvision | |
| pip install $(curl -sSL https://raw.githubusercontent.com/huggingface/pytorch-image-models/$TIMM_COMMIT_ID/requirements.txt | grep -vE torch) | |
| - name: Generate Triton cache key | |
| id: triton-key | |
| run: | | |
| COMPOSITE_KEY=$(echo $PYTHON_VERSION $GITHUB_SHA | sha256sum - | cut -d\ -f1) | |
| echo "key=triton-$COMPOSITE_KEY" >> $GITHUB_OUTPUT | |
| - name: Load Triton wheels from a cache | |
| id: triton-cache | |
| uses: ./.github/actions/load | |
| with: | |
| path: python/dist | |
| key: ${{ steps.triton-key.outputs.key }} | |
| - name: Build Triton wheels | |
| if: ${{ steps.triton-cache.outputs.status == 'miss' }} | |
| uses: ./.github/actions/setup-triton | |
| with: | |
| command: DEBUG=1 python setup.py bdist_wheel | |
| - name: Install Triton | |
| run: | | |
| pip install python/dist/*.whl | |
| - name: Save Triton wheels to a cache | |
| if: ${{ steps.triton-cache.outputs.status == 'miss' }} | |
| uses: ./.github/actions/save | |
| with: | |
| path: ${{ steps.triton-cache.outputs.path }} | |
| dest: ${{ steps.triton-cache.outputs.dest }} | |
| - name: Install python test dependencies | |
| run: | | |
| pip install pyyaml pandas scipy numpy psutil pyre_extensions torchrec | |
| - name: Install transformers package | |
| if: ${{ matrix.suite == 'huggingface' }} | |
| uses: ./.github/actions/install-dependency | |
| with: | |
| package: transformers | |
| repository: huggingface/transformers | |
| ref: ${{ env.TRANSFORMERS_VERSION }} | |
| try-tag-prefix: v | |
| - name: Clone pytorch benchmark | |
| if: ${{ matrix.suite == 'torchbench' }} | |
| uses: actions/checkout@v4 | |
| with: | |
| repository: ${{ env.BENCHMARK_REPO }} | |
| ref: ${{ env.BENCHMARK_BRANCH }} | |
| submodules: recursive | |
| path: benchmark | |
| - name: Install pytorch benchmark | |
| if: ${{ matrix.suite == 'torchbench' }} | |
| run: | | |
| cd benchmark | |
| python install.py | |
| pip install -e . | |
| - name: Run e2e performance benchmarks | |
| env: | |
| HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }} | |
| run: | | |
| export WORKSPACE=$GITHUB_WORKSPACE | |
| if [[ "${{ inputs.TORCH_COMPILE_DEBUG }}" == "1" ]] ; then | |
| export TORCH_COMPILE_DEBUG="1" | |
| # torch will save debug logs to $TORCH_COMPILE_DEBUG_DIR/torch_compile_debug | |
| export TORCH_COMPILE_DEBUG_DIR=$GITHUB_WORKSPACE | |
| fi | |
| cd pytorch | |
| # if "only_one_model" is set, then test this model | |
| # if "models" == "subset", then test the models from .github/models/performance/{suite}.txt | |
| # otherwise test all models | |
| if [[ "${{ inputs.only_one_model }}" ]]; then | |
| $GITHUB_WORKSPACE/scripts/inductor_xpu_test.sh ${{ matrix.suite }} ${{ matrix.dtype }} ${{ matrix.mode }} performance xpu 0 static 1 0 ${{ inputs.only_one_model }} | |
| elif [[ "${{ needs.setup.outputs.models }}" == "subset" ]]; then | |
| while read model; do | |
| $GITHUB_WORKSPACE/scripts/inductor_xpu_test.sh ${{ matrix.suite }} ${{ matrix.dtype }} ${{ matrix.mode }} performance xpu 0 static 1 0 $model | |
| done < $GITHUB_WORKSPACE/.github/models/performance/${{ matrix.suite }}.txt | |
| else | |
| $GITHUB_WORKSPACE/scripts/inductor_xpu_test.sh ${{ matrix.suite }} ${{ matrix.dtype }} ${{ matrix.mode }} performance xpu 0 static 1 0 | |
| fi | |
| - name: Report environment details | |
| run: | | |
| mkdir -p inductor_log | |
| TIMESTAMP=$(date '+%Y%m%d%H%M%S') | |
| echo "TIMESTAMP=$TIMESTAMP" >> "${GITHUB_ENV}" | |
| source ./scripts/capture-hw-details.sh --quiet | |
| cat <<EOF | tee inductor_log/.env | |
| TIMESTAMP=$TIMESTAMP | |
| JOB_NAME=${{ join(matrix.*, '-') }} | |
| GITHUB_RUN_ID=$GITHUB_RUN_ID | |
| GITHUB_RUN_NUMBER=$GITHUB_RUN_NUMBER | |
| GITHUB_RUN_ATTEMPT=$GITHUB_RUN_ATTEMPT | |
| PYTHON_VERSION=$PYTHON_VERSION | |
| PYTORCH_REPO=$PYTORCH_REPO | |
| PYTORCH_COMMIT_ID=$PYTORCH_COMMIT_ID | |
| IPEX_REPO=$IPEX_REPO | |
| IPEX_COMMIT_ID=$IPEX_COMMIT_ID | |
| LLVM_REPO=llvm/llvm-project | |
| LLVM_COMMIT_ID=$LLVM_COMMIT_ID | |
| BENCHMARK_REPO=$BENCHMARK_REPO | |
| BENCHMARK_COMMIT_ID=$BENCHMARK_COMMIT_ID | |
| TRITON_REPO=$GITHUB_REPOSITORY | |
| TRITON_COMMIT_ID=$GITHUB_SHA | |
| TORCHVISION_COMMIT_ID=$TORCHVISION_COMMIT_ID | |
| TORCHTEXT_COMMIT_ID=$TORCHTEXT_COMMIT_ID | |
| TORCHAUDIO_COMMIT_ID=$TORCHAUDIO_COMMIT_ID | |
| TRANSFORMERS_VERSION=$TRANSFORMERS_VERSION | |
| TIMM_COMMIT_ID=$TIMM_COMMIT_ID | |
| LIBIGC1_VERSION=$LIBIGC1_VERSION | |
| LEVEL_ZERO_VERSION=$LEVEL_ZERO_VERSION | |
| GPU_DEVICE=$GPU_DEVICE | |
| AGAMA_VERSION=$AGAMA_VERSION | |
| EOF | |
| - name: Copy reports | |
| run: | | |
| if [[ -d torch_compile_debug ]]; then | |
| cp -rT torch_compile_debug inductor_log | |
| fi | |
| - name: Upload test logs | |
| uses: actions/upload-artifact@v4 | |
| with: | |
| name: logs-${{ join(matrix.*, '-') }} | |
| path: inductor_log | |
| - name: Save pip cache | |
| if: ${{ steps.pip-cache.outputs.status == 'miss' }} | |
| uses: ./.github/actions/save | |
| with: | |
| path: ${{ steps.pip-cache.outputs.path }} | |
| dest: ${{ steps.pip-cache.outputs.dest }} |