tfhe-rs/.github/workflows/gpu_full_h100_tests.yml

# Compile and test tfhe-cuda-backend on an H100 VM on hyperstack
name: gpu_full_h100_tests

env:
  CARGO_TERM_COLOR: always
  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
  RUSTFLAGS: "-C target-cpu=native"
  RUST_BACKTRACE: "full"
  RUST_MIN_STACK: "8388608"
  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}

on:
  workflow_dispatch:


permissions: {}

# zizmor: ignore[concurrency-limits] concurrency is managed after instance setup to ensure safe provisioning

jobs:
  setup-instance:
    name: gpu_full_h100_tests/setup-instance
    runs-on: ubuntu-latest
    outputs:
      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
      # otherwise we'll try to run the next job on a non-existing on-demand instance.
      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label }}
      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
    steps:
      - name: Start remote instance
        id: start-remote-instance
        continue-on-error: true
        uses: zama-ai/slab-github-runner@973c1d22702de8d0acd2b34e83404c96ed92c264 # v1.4.2
        with:
          mode: start
          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
          slab-url: ${{ secrets.SLAB_BASE_URL }}
          job-secret: ${{ secrets.JOB_SECRET }}
          backend: hyperstack
          profile: single-h100

      # This will allow to fallback on permanent instances running on Hyperstack.
      - name: Use permanent remote instance
        id: use-permanent-instance
        if: env.SECRETS_AVAILABLE == 'true' && steps.start-remote-instance.outcome == 'failure'
        run: |
          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"

  cuda-tests-linux:
    name: gpu_full_h100_tests/cuda-tests-linux
    needs: [ setup-instance ]
    concurrency:
      group: ${{ github.workflow_ref }}
      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
    strategy:
      fail-fast: false
      # explicit include-based build matrix, of known valid options
      matrix:
        include:
          - os: ubuntu-22.04
            cuda: "12.8"
            gcc: 11 
    steps:
      - name: Checkout tfhe-rs
        uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8
        with:
          persist-credentials: 'false'
          token: ${{ secrets.REPO_CHECKOUT_TOKEN }}

      - name: Setup Hyperstack dependencies
        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
        uses: ./.github/actions/gpu_setup
        with:
          cuda-version: ${{ matrix.cuda }}
          gcc-version: ${{ matrix.gcc }}

      - name: Install latest stable
        uses: dtolnay/rust-toolchain@e97e2d8cc328f1b50210efc529dca0028893a2d9 # zizmor: ignore[stale-action-refs] this action doesn't create releases
        with:
          toolchain: stable
      - name: Enable nvidia multi-process service
        run: |
          nvidia-cuda-mps-control -d
      - name: Run core crypto, integer and internal CUDA backend tests
        run: |
          make test_gpu

      - name: Run user docs tests
        run: |
          make test_user_doc_gpu

      - name: Test C API
        run: |
          make test_c_api_gpu

      - name: Run High Level API Tests
        run: |
          make test_high_level_api_gpu

  slack-notify:
    name: gpu_full_h100_tests/slack-notify
    needs: [ setup-instance, cuda-tests-linux ]
    runs-on: ubuntu-latest
    if: ${{ failure() }}
    continue-on-error: true
    steps:
      - name: Send message
        uses: rtCamp/action-slack-notify@e31e87e03dd19038e411e38ae27cbad084a90661
        env:
          SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
          SLACK_MESSAGE: "Full H100 tests finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.ACTION_RUN_URL }})"

  teardown-instance:
    name: gpu_full_h100_tests/teardown-instance
    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
    needs: [ setup-instance, cuda-tests-linux ]
    runs-on: ubuntu-latest
    steps:
      - name: Stop instance
        id: stop-instance
        uses: zama-ai/slab-github-runner@973c1d22702de8d0acd2b34e83404c96ed92c264 # v1.4.2
        with:
          mode: stop
          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
          slab-url: ${{ secrets.SLAB_BASE_URL }}
          job-secret: ${{ secrets.JOB_SECRET }}
          label: ${{ needs.setup-instance.outputs.runner-name }}

      - name: Slack Notification
        if: ${{ failure() }}
        uses: rtCamp/action-slack-notify@e31e87e03dd19038e411e38ae27cbad084a90661
        env:
          SLACK_COLOR: ${{ job.status }}
          SLACK_MESSAGE: "Instance teardown (cuda-h100-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"