How would you design a scalable and efficient system for training a large language model, considering both computational and data constraints?