You have a native synchronous implementation for model training and observe GPU utilization issues. Your training data is split across multiple files. To reduce input pipeline execution time, what should you do?