How would you design a cluster for large-scale model training, including checkpointing and data loading for Enigma?
