Mixed Precision Training

Using lower precision (FP16) for some computations while keeping FP32 for stability, speeding up training.

This concept is essential for understanding training & optimization and forms a key part of modern AI systems.