Crusoe engineers show how they keep large training runs alive through GPU failures with automatic detection, checkpointing and node replacement.