Module prefetch 

Predictive Data Prefetching

Uses phase timing data from the sched_ext scheduler to predict when the GpuCompute phase will end and signal the DataLoader to start prefetching the next batch BEFORE the GPU finishes.

This eliminates the “GPU idle waiting for data” gap that causes 5-30% throughput loss in data-bound training.

Structs

PrefetchPredictor

Tracks phase durations to predict when to trigger prefetch.