Comment by ece

This would be interesting on a Spark or two. I don't know how seamless the GPU and networking cluster setup would be when resource sharing between agent and post-training VMs.