lightning/.github/workflows/tpu-tests.yml

name: Test PyTorch - TPU

on:
  push:
    branches: [master, "release/*"]
  pull_request_target:
    branches: [master, "release/*"]
    types: [opened, reopened, ready_for_review, labeled, synchronize]

concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref }}
  cancel-in-progress: ${{ github.event_name == 'pull_request' }}

defaults:
  run:
    shell: bash

jobs:
  test-on-tpus:
    runs-on: ubuntu-22.04
    # run only when the PR title contains 'TPU' or is a merge to master
    if: |
      (github.event_name == 'push' && github.ref == 'refs/heads/master') ||
      (startsWith(github.event_name, 'pull_request') && contains(github.event.pull_request.labels.*.name, 'run TPU'))      
    strategy:
      fail-fast: false
      matrix:
        pkg-name: ["fabric", "pytorch"]
        accelerator_type: ["v4-8"]
    timeout-minutes: 30
    env:
      XLA_VER: "2.0"
      PR_NUMBER: ${{ github.event.pull_request.number && github.event.pull_request.number || 'master' }}
      SHA: ${{ github.event.pull_request.head.sha && github.event.pull_request.head.sha || github.sha }}
      CLOUDSDK_CORE_DISABLE_PROMPTS: 1 # default to --quiet
    steps:
      - name: Set env
        run: |
          # define --zone: https://cloud.google.com/tpu/docs/regions-zones
          if [[ "${{ matrix.accelerator_type }}" == v4* ]]; then
            echo "CLOUDSDK_COMPUTE_ZONE=us-central2-b" >> $GITHUB_ENV
          else
            echo "CLOUDSDK_COMPUTE_ZONE=us-west4-a" >> $GITHUB_ENV
          fi          

      - uses: actions/checkout@v4
        with:
          ref: ${{ github.event.pull_request.head.sha }}
      - uses: actions/setup-python@v5
        with:
          python-version: "3.10"

      - uses: google-github-actions/auth@v2
        with:
          credentials_json: ${{ secrets.GKE_SA_KEY_BASE64 }}
      - uses: "google-github-actions/setup-gcloud@v2"

      - name: Time-based job cleanup
        if: always()
        run: |
          gcloud compute tpus tpu-vm list --format='value(name,createTime)' > creation_times.txt
          cat creation_times.txt

          if [ ! -s "creation_times.txt" ]; then
            echo "No existing jobs"
            exit 0
          fi

          jobs_deleted=false
          while read -r job_name created_at; do
            # Skip jobs with "keepalive" in the name
            if [[ "$job_name" == *"keepalive"* ]]; then
              echo "Skipping $job_name, has keepalive in name"
              continue
            fi

            # Convert the creation time to Unix timestamp
            created_timestamp=$(date -d "${created_at}" +%s)

            # Calculate the difference between the current time and the creation time
            current_timestamp=$(date +%s)
            age=$((current_timestamp - created_timestamp))

            # Check if the age has surpassed a timeout
            if ((age > 35 * 60)); then
              # delete the job
              gcloud compute tpus tpu-vm delete "$job_name" --async
              jobs_deleted=true
            else
              echo "Skipping $job_name, alive for $age seconds"
            fi
          done < creation_times.txt

          if [ "$jobs_deleted" = true ]; then
            sleep 5
            # diagnostics
            gcloud compute tpus tpu-vm list
          fi          

      - name: Update script
        run: |
          import os
          fname = f'tests/tests_${{ matrix.pkg-name }}/run_tpu_tests.sh'
          with open(fname) as fopen:
              data = fopen.read()
          data = data.replace('{PYTORCH_VERSION}', os.environ["XLA_VER"])
          print(data)
          with open(fname, "w") as fopen:
              fopen.write(data)          
        shell: python

      - name: Create node
        id: tpu-create
        # TPU capacity is very limited so this workflow's success is optional. continue normally if creation fails
        continue-on-error: true
        env:
          JOB_NAME: ${{ env.PR_NUMBER }}-${{ matrix.pkg-name }}-${{ matrix.accelerator_type }}-${{ env.SHA }}
        run: |
          if [[ "${{ matrix.accelerator_type }}" == v4* ]]; then
            gcloud compute tpus tpu-vm create "$JOB_NAME" \
              --accelerator-type=${{ matrix.accelerator_type }} \
              --version="tpu-vm-v4-pt-$XLA_VER" \
              --preemptible
          fi          

      - name: Run tests
        if: steps.tpu-create.outcome == 'success'
        env:
          JOB_NAME: ${{ env.PR_NUMBER }}-${{ matrix.pkg-name }}-${{ matrix.accelerator_type }}-${{ env.SHA }}
        run: |
          set -uex

          # zip-copy-unzip the repository
          zip -q -r repo.zip . -x .git/
          gcloud compute tpus tpu-vm scp --worker=all repo.zip "$JOB_NAME":~
          gcloud compute tpus tpu-vm ssh "$JOB_NAME" --worker=all --command="cd ~; unzip -q -o repo.zip"

          # run script
          gcloud compute tpus tpu-vm ssh "$JOB_NAME" --worker=all --command="cd ~; bash tests/tests_${{ matrix.pkg-name }}/run_tpu_tests.sh"
          exit_code=$?

          # pull out the coverage file
          gcloud compute tpus tpu-vm scp "$JOB_NAME":~/coverage.xml .

          exit $exit_code          

      - name: Cleanup job
        if: always()
        env:
          JOB_NAME: ${{ env.PR_NUMBER }}-${{ matrix.pkg-name }}-${{ matrix.accelerator_type }}-${{ env.SHA }}
        run: |
          if ! gcloud compute tpus tpu-vm list | grep -q "$JOB_NAME"; then
            echo "$JOB_NAME wasn't created"
            exit 0
          fi

          # diagnostics
          gcloud compute tpus tpu-vm describe "$JOB_NAME"

          # delete the job
          gcloud compute tpus tpu-vm delete "$JOB_NAME" --async
          sleep 5

          # diagnostics
          gcloud compute tpus tpu-vm list          

      - name: Upload coverage to Codecov
        uses: codecov/codecov-action@v4
        continue-on-error: true
        with:
          token: ${{ secrets.CODECOV_TOKEN }}
          file: coverage.xml
          flags: tpu,pytest
          name: TPU-coverage
          fail_ci_if_error: false