tfhe-rs/.github/workflows/gpu_fast_h100_tests.yml

# Compile and test tfhe-cuda-backend on an H100 VM on hyperstack
name: gpu_fast_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 }}
  SLACKIFY_MARKDOWN: true
  IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
  PULL_REQUEST_MD_LINK: ""
  CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
  # Secrets will be available only to zama-ai organization members
  SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
  EXTERNAL_CONTRIBUTION_RUNNER: "gpu_ubuntu-22.04"

on:
  # Allows you to run this workflow manually from the Actions tab as an alternative.
  workflow_dispatch:
  pull_request:
    types: [ labeled ]

permissions:
  contents: read

jobs:
  should-run:
    name: gpu_fast_h100_tests/should-run
    runs-on: ubuntu-latest
    permissions:
      pull-requests: read
    outputs:
      gpu_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.gpu_any_changed }}
    steps:
      - name: Checkout tfhe-rs
        uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8
        with:
          fetch-depth: 0
          persist-credentials: 'false'
          token: ${{ env.CHECKOUT_TOKEN }}

      - name: Check for file changes
        id: changed-files
        uses: tj-actions/changed-files@ed68ef82c095e0d48ec87eccea555d944a631a4c # v46.0.5
        with:
          files_yaml: |
            gpu:
              - tfhe/Cargo.toml
              - tfhe/build.rs
              - backends/tfhe-cuda-backend/**
              - tfhe/src/core_crypto/gpu/**
              - tfhe/src/integer/gpu/**
              - tfhe/src/integer/server_key/radix_parallel/tests_unsigned/**
              - tfhe/src/integer/server_key/radix_parallel/tests_signed/**
              - tfhe/src/integer/server_key/radix_parallel/tests_cases_unsigned.rs
              - tfhe/src/shortint/parameters/**
              - tfhe/src/high_level_api/**
              - tfhe/src/c_api/**
              - 'tfhe/docs/**/**.md'
              - '.github/workflows/gpu_fast_h100_tests.yml'
              - scripts/integer-tests.sh
              - ci/slab.toml

  setup-instance:
    name: gpu_fast_h100_tests/setup-instance
    needs: should-run
    if: github.event_name != 'pull_request' ||
      (github.event.action != 'labeled' && needs.should-run.outputs.gpu_test == 'true') ||
      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true')
    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 || steps.start-github-instance.outputs.runner_group }}
      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
    steps:
      - name: Start remote instance
        id: start-remote-instance
        if: env.SECRETS_AVAILABLE == 'true'
        continue-on-error: true
        uses: zama-ai/slab-github-runner@79939325c3c429837c10d6041e4fd8589d328bac
        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"

      # This instance will be spawned especially for pull-request from forked repository
      - name: Start GitHub instance
        id: start-github-instance
        if: env.SECRETS_AVAILABLE == 'false'
        run: |
          echo "runner_group=${EXTERNAL_CONTRIBUTION_RUNNER}" >> "$GITHUB_OUTPUT"

  cuda-tests-linux:
    name: gpu_fast_h100_tests/cuda-tests-linux
    needs: [ should-run, setup-instance ]
    if: github.event_name != 'pull_request' ||
      (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
    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@08c6903cd8c0fde910a37f88322edcfb5dd907a8
        with:
          persist-credentials: 'false'
          token: ${{ env.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 }}
          github-instance: ${{ env.SECRETS_AVAILABLE == 'false' }}

      - 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 and internal CUDA backend tests
        run: |
          BIG_TESTS_INSTANCE=TRUE make test_core_crypto_gpu
          BIG_TESTS_INSTANCE=TRUE make test_integer_compression_gpu
          BIG_TESTS_INSTANCE=TRUE make test_cuda_backend

      - name: Run user docs tests
        run: |
          BIG_TESTS_INSTANCE=TRUE make test_user_doc_gpu

      - name: Test C API
        run: |
          BIG_TESTS_INSTANCE=TRUE make test_c_api_gpu

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

  slack-notify:
    name: gpu_fast_h100_tests/slack-notify
    needs: [ setup-instance, cuda-tests-linux ]
    runs-on: ubuntu-latest
    if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' && failure() }}
    continue-on-error: true
    steps:
      - name: Set pull-request URL
        if: env.SECRETS_AVAILABLE == 'true' && github.event_name == 'pull_request'
        run: |
          echo "PULL_REQUEST_MD_LINK=[pull-request](${PR_BASE_URL}${PR_NUMBER}), "  >> "${GITHUB_ENV}"
        env:
          PR_BASE_URL: ${{ vars.PR_BASE_URL }}
          PR_NUMBER: ${{ github.event.pull_request.number }}

      - name: Send message
        if: env.SECRETS_AVAILABLE == 'true'
        uses: rtCamp/action-slack-notify@e31e87e03dd19038e411e38ae27cbad084a90661
        env:
          SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
          SLACK_MESSAGE: "Fast H100 tests finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"

  teardown-instance:
    name: gpu_fast_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 remote instance
        id: stop-instance
        if: env.SECRETS_AVAILABLE == 'true'
        uses: zama-ai/slab-github-runner@79939325c3c429837c10d6041e4fd8589d328bac
        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() }}
        continue-on-error: true
        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 }})"