How do you handle data pipeline bottlenecks when training models on billions of tokens of source code?
