Parallel creation of head and workers, and faster skipping of 'empty' steps?

in the context of optimizing the provisioning time of ray clusters, i came across 2 questions where i’d be interested in your input:

  1. is it possible to start the provisioning of the head node and the worker nodes (almost) in parallel, vs basically provisioning them in sequence?
  2. at least when using the ray cluster launcher, there is quite some time spent processing sections of the cluster yaml, even though they’re empty (e.g. those mounting stuff, etc.). Is there are more direct way to just have the launcher execute the steps that are actually relevant?