Realistically, most people barely get to multiple GPUs, let alone multiple machines. You're more likely to do hyperparameter tuning across machines before you do distributed training.