Training with lower precision formats to save memory and increase throughput, with loss‑scaling.