all: 
  children:
    cluster:
      children:
        manager:
          hosts:
            compute000:
        workers:
          children:
            compute:
              hosts:
                compute001:
            gpus:
              hosts:
                compute002:
                compute003:
                compute004: