e2e-performance.yml

name: E2E performance
run-name: ${{ inputs.run_name }}

on:
  workflow_dispatch:
    inputs:
      pytorch_ref:
        description: PyTorch ref, keep empty for default
        type: string
        default: ""
      suite:
        description: Test suite
        type: choice
        options:
          - all
          - huggingface
          - timm_models
          - torchbench
        default: all
      mode:
        description: Inference, inference-no-freezing, or training
        type: choice
        options:
          - all
          - inference
          - inference-no-freezing
          - 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: ""
      run_name:
        description: Custom run name
        type: string
        default: "E2E performance"

permissions: read-all

jobs:
  setup:
    name: Setup
    runs-on: Linux
    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", "inference-no-freezing", "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: |
          cat <<EOF
          ${{ toJSON(github.event.inputs) }}
          EOF

      - name: Print setup outputs
        run: |
          cat <<EOF
          ${{ toJSON(needs.setup.outputs) }}
          EOF

  run_tests:
    name: Run test matrix
    needs: setup
    strategy:
      matrix:
        suite: ${{ fromJson(needs.setup.outputs.suite) }}
        mode: ${{ fromJson(needs.setup.outputs.mode) }}
        dtype: ${{ fromJson(needs.setup.outputs.dtype) }}
      fail-fast: false
    uses: ./.github/workflows/e2e-reusable.yml
    with:
      pytorch_ref: ${{ inputs.pytorch_ref }}
      suite: ${{ matrix.suite }}
      mode: ${{ matrix.mode }}
      test_mode: performance
      dtype: ${{ matrix.dtype }}
      models: ${{ inputs.models }}
      only_one_model: ${{ inputs.only_one_model }}
      runner_label: ${{ inputs.runner_label }}
      TORCH_COMPILE_DEBUG: ${{ inputs.TORCH_COMPILE_DEBUG }}