perf-models.yaml

name: "Model perf regressions and output report"

on:
  workflow_dispatch:
  schedule:
    - cron: "0 2,7,10,14,17,20,23 * * *"
  workflow_call:

jobs:
  build-docker-artifact:
    uses: ./.github/workflows/build-docker-artifact.yaml
    secrets: inherit
  build-artifact:
    needs: build-docker-artifact
    uses: ./.github/workflows/build-artifact.yaml
    secrets: inherit
  models-perf:
    needs: build-artifact
    strategy:
      # Do not fail-fast because we need to ensure all tests go to completion
      # so we try not to get hanging machines
      fail-fast: false
      matrix:
        test-info: [
          {name: "GS", arch: grayskull, runs-on: ["perf-grayskull", "self-reset"], machine-type: "bare_metal"},
          {name: "N300 WH B0", arch: wormhole_b0, runs-on: ["perf-wormhole_b0", "self-reset"], machine-type: "bare_metal"},
        ]
        model-type: [llm_javelin, cnn_javelin, other]
    name: "${{ matrix.model-type }} ${{ matrix.test-info.name }}"
    env:
      TT_METAL_ENV: ${{ vars.TT_METAL_ENV }}
      ARCH_NAME: ${{ matrix.test-info.arch }}
      LOGURU_LEVEL: INFO
      TTNN_CONFIG_OVERRIDES: '{"enable_fast_runtime_mode": true}'
    environment: dev
    runs-on: ${{ matrix.test-info.runs-on }}
    steps:
      - uses: tenstorrent-metal/metal-workflows/.github/actions/checkout-with-submodule-lfs@v2.0.0
      - name: Enable Performance mode
        run: |
          sudo cpupower frequency-set -g performance
      - name: Ensure weka mount is active
        run: |
          sudo systemctl restart mnt-MLPerf.mount
          sudo /etc/rc.local
          ls -al /mnt/MLPerf/bit_error_tests
      - name: Set up dynamic env vars for build
        run: |
          echo "TT_METAL_HOME=$(pwd)" >> $GITHUB_ENV
          echo "PYTHONPATH=$(pwd)" >> $GITHUB_ENV
      - uses: actions/download-artifact@v4
        with:
          name: TTMetal_build_${{ matrix.test-info.arch }}
      - name: Extract files
        run: tar -xvf ttm_${{ matrix.test-info.arch }}.tar
      - uses: ./.github/actions/install-python-deps
      - uses: ./.github/actions/load-docker-image
      - name: Run performance regressions
        id: performance_tests
        timeout-minutes: 30
        run: |
          ./scripts/docker/run_docker_cmd.sh -e TTNN_CONFIG_OVERRIDES=${TTNN_CONFIG_OVERRIDES} -- \
          /bin/bash -c "source python_env/bin/activate \  
            && ./tests/scripts/run_tests.sh --tt-arch $ARCH_NAME --pipeline-type ${{ matrix.model-type }}_models_performance_${{ matrix.test-info.machine-type }}"
      - uses: ./.github/actions/slack-report
        if: ${{ steps.performance_tests.outcome != 'success' }}
        with:
          slack_webhook_url: ${{ secrets.SLACK_WEBHOOK_URL }}
      - name: Check perf report exists
        id: check-perf-report
        if: ${{ !cancelled() }}
        run: |
          ls -hal
          export PERF_REPORT_FILENAME=Models_Perf_$(date +%Y_%m_%d).csv
          ls -hal $PERF_REPORT_FILENAME
          echo "perf_report_filename=$PERF_REPORT_FILENAME" >> "$GITHUB_OUTPUT"
      - name: Upload perf report
        if: ${{ !cancelled() && steps.check-perf-report.conclusion == 'success' }}
        uses: actions/upload-artifact@v4
        with:
          name: perf-report-csv-${{ matrix.model-type }}-${{ matrix.test-info.arch }}-${{ matrix.test-info.machine-type }}
          path: "${{ steps.check-perf-report.outputs.perf_report_filename }}"
      - name: Disable Performance mode
        if: always()
        run: |
          sudo cpupower frequency-set -g ondemand