Building and testing distributed ML training across mixed GPU clusters using PyTorch FSDP2, Ray Train, and Kubernetes.