cifar_benchmark.yaml

name: CIFAR-10 benchmark CML
on:
  workflow_dispatch:
    inputs:
      git-ref:
        description: Repo reference (branch, tag or SHA)
        default: "main"
        required: true
        type: string
      benchmark:
        description: Benchmark to run (cifar-10-8b or cifar-10-16b)
        default: "cifar-10-16b"
        type: choice
        options:
          - "cifar-10-8b"
          - "cifar-10-16b"
      instance_type:
        description: Instance type on which to launch benchmarks
        default: "hpc7a.96xlarge"
        type: choice
        options:
          - "m6i.metal"
          - "u-6tb1.112xlarge"
          - "hpc7a.96xlarge" 
      num_samples:
        description: Number of samples to use
        default: "3"
        type: string
        required: true
      p_error:
        description: P-error to use
        default: "0.01"
        type: string
        required: true

  # FIXME: Add recurrent launching
  # https://github.com/zama-ai/concrete-ml-internal/issues/1851

# Global environnement variables
env:
  # Github action url (used by slack notification)
  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
  AGENT_TOOLSDIRECTORY: /opt/hostedtoolcache
  RUNNER_TOOL_CACHE: /opt/hostedtoolcache
  SLAB_PROFILE: big-cpu


# Jobs
jobs:
  setup-ec2:
    name: Setup EC2 instance
    runs-on: ubuntu-latest
    outputs:
      runner-name: ${{ steps.start-instance.outputs.label }}
      instance-id: ${{ steps.start-instance.outputs.ec2-instance-id }}
    steps:
      - name: Start instance
        id: start-instance
        uses: zama-ai/slab-github-runner@867256e1784c920d89ce3411ffa031420c7ced11
        with:
          mode: start
          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
          slab-url: ${{ secrets.SLAB_BASE_URL }}
          job-secret: ${{ secrets.JOB_SECRET }}
          profile: ${{ env.SLAB_PROFILE }}

  run-cifar-10:
    needs: [setup-ec2]
    name: Run benchmark
    runs-on: ${{ needs.setup-ec2.outputs.runner-name }}
    env:
      PIP_INDEX_URL: ${{ secrets.PIP_INDEX_URL }}
      PIP_EXTRA_INDEX_URL: ${{ secrets.PIP_EXTRA_INDEX_URL }}
    steps:
      - name: Add masks
        run: |
          echo "::add-mask::${{ secrets.INTERNAL_PYPI_URL_FOR_MASK }}"
          echo "::add-mask::${{ secrets.INTERNAL_REPO_URL_FOR_MASK }}"
          echo "::add-mask::${{ secrets.INTERNAL_PYPI_URL }}"
          echo "::add-mask::${{ secrets.INTERNAL_REPO_URL }}"

      - name: Checkout code
        uses: actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11
        with:
          lfs: true
          ref: ${{ github.event.inputs.git-ref }}

      - name: Set up Python
        uses: actions/setup-python@0a5c61591373683505ea898e09a3ea4f39ef2b9c
        with:
          python-version: "3.8"

      - name: Install dependencies
        id: install-deps
        run: |
          apt update
          apt install --no-install-recommends -y gnome-keyring
          apt install -y graphviz* graphviz-dev libgraphviz-dev pkg-config python3-dev
          apt-mark hold docker.io
          ./script/make_utils/setup_os_deps.sh
          make setup_env

      # CIFAR-10-8b benchmark
      - name: Benchmark - CIFAR-10-8b
        if: github.event.inputs.benchmark == 'cifar-10-8b'
        run: |
          source .venv/bin/activate
          NUM_SAMPLES=${{ github.event.inputs.num_samples }} python3 ./use_case_examples/cifar/cifar_brevitas_with_model_splitting/infer_fhe.py
          python3 ./benchmarks/convert_cifar.py --model-name "8-bit-split-v0"

      # CIFAR-10-16b benchmark
      - name: Benchmark - CIFAR-10-16b
        if: github.event.inputs.benchmark == 'cifar-10-16b'
        run: |
          source .venv/bin/activate
          NUM_SAMPLES=${{ github.event.inputs.num_samples }} P_ERROR=${{ github.event.inputs.p_error }} python3 ./use_case_examples/cifar/cifar_brevitas_training/evaluate_one_example_fhe.py
          python3 ./benchmarks/convert_cifar.py --model-name "16-bits-trained-v0"

      - name: Archive raw predictions
        uses: actions/upload-artifact@v4.3.0
        with:
          name: predictions.csv
          path: inference_results.csv

      - name: Archive metrics
        uses: actions/upload-artifact@v4.3.0
        with:
          name: metrics.json
          path: to_upload.json

      - name: Archive MLIR
        uses: actions/upload-artifact@v4.3.0
        with:
          name: mlir.txt
          path: cifar10.mlir

      - name: Archive Graph
        uses: actions/upload-artifact@v4.3.0
        with:
          name: graph.txt
          path: cifar10.graph

      - name: Archive client
        uses: actions/upload-artifact@v4.3.0
        with:
          name: client.zip
          path: client_server/client.zip

      - name: Archive server
        uses: actions/upload-artifact@v4.3.0
        with:
          name: server.zip
          path: client_server/server.zip

      # We need to keep this as the last step to avoid not uploading the artifacts
      # if the step crashes
      - name: Upload results
        id: upload-results
        run: |
          # Log the json
          cat to_upload.json | jq
          
          # We need to sleep to avoid log issues
          sleep 1.

          # Upload the json to the benchmark database
          curl --fail-with-body \
          -H "Authorization: Bearer ${{ secrets.NEW_ML_PROGRESS_TRACKER_TOKEN }}" \
          -H "Content-Type: application/json; charset=UTF-8" \
          -d @to_upload.json \
          -X POST "${{ secrets.NEW_ML_PROGRESS_TRACKER_URL }}experiment"

  teardown-ec2:
    name: Teardown EC2 instance (fast-tests)
    if: ${{ always() }}
    needs: [ setup-ec2, run-cifar-10 ]
    runs-on: ubuntu-latest
    steps:
      - name: Stop instance
        id: stop-instance
        uses: zama-ai/slab-github-runner@ab65ad70bb9f9e9251e4915ea5612bcad23cd9b1
        with:
          mode: stop
          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
          slab-url: ${{ secrets.SLAB_BASE_URL }}
          job-secret: ${{ secrets.JOB_SECRET }}
          profile: ${{ env.SLAB_PROFILE }}
          label: ${{ needs.setup-ec2.outputs.runner-name }}

  slack-notification:
    runs-on: ubuntu-20.04
    needs: [run-cifar-10]
    steps:
      - name: Slack Notification
        if: ${{ always() }}
        continue-on-error: true
        uses: rtCamp/action-slack-notify@b24d75fe0e728a4bf9fc42ee217caa686d141ee8
        env:
          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
          SLACK_COLOR: ${{ needs.run-cifar-10.result }}
          SLACK_MESSAGE: "Benchmark action: ${{ github.event.inputs.benchmark }} (${{ env.ACTION_RUN_URL }}) ended with result: ${{ needs.run-cifar-10.result }}"
          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}