Explain how you would optimize a data pipeline to prevent GPU starvation during the training of a multi-billion parameter model.
