monitor-vms.yml

# Regularly updates the CI container
name: Reboots VMs in a controlled way
on:
  schedule:
    - cron: 0/15 * * * *
  workflow_dispatch:
  
jobs:
  pre-flight:
    runs-on: ubuntu-latest
    outputs:
      list-of-vms: ${{ steps.main.outputs.main }}
    environment: main
    steps:
      - name: Get list of VMs
        id: main
        env:
          GITHUB_TOKEN: ${{ secrets.PAT }}
        run: |
          RUNNERS=$(curl -L \
            -H "Accept: application/vnd.github+json" \
            -H "Authorization: Bearer $GITHUB_TOKEN" \
            -H "X-GitHub-Api-Version: 2022-11-28" \
            https://api.github.com/repos/NVIDIA/NeMo/actions/runners)

          MATRIX=$(echo $RUNNERS \
            | jq -c '[
                .runners[] 
                | select(.status == "online")
                | select(.name | contains("gpu"))
                | {
                  "vm": .name, 
                  "n_gpus": [
                    .labels[] 
                    | select(.name | endswith("gpu")) | .name
                  ][0][:1]
                }
              ]
            '
          )
          echo main=$MATRIX | tee -a "$GITHUB_OUTPUT"

  maintenance:
    needs: pre-flight
    strategy:
      fail-fast: false
      matrix:
        include: ${{ fromJSON(needs.pre-flight.outputs.list-of-vms )}}
    uses: ./.github/workflows/monitor-single-vm.yml
    with:
      vm: ${{ matrix.vm }}
      n_gpus: ${{ matrix.n_gpus }}
    secrets: inherit  # pragma: allowlist secret