Without workers: training waits for batch. With workers: prefetched.

What you're seeing

num_workers ~ CPU cores / 2 for CPU training (leave room for compute).

★ KEY TAKEAWAY
Multi-worker dataloaders prefetch batches into a queue. Without workers, training waits on data; with N workers, never starves.
▶ WHAT TO TRY
  • Slide Workers from 0 to 8.
  • For CPU training: num_workers ~ CPU cores / 2 (leave room for compute).
← Back to Transformer Math & CPU SLM