How would you optimize a PyTorch training pipeline that is bottlenecked by CPU data loading for AMD GPU training?
