Mixed Precision Computing

Master Weights

Copy of model weights maintained in FP32 (or FP64) during mixed precision training, serving as a precision reference for weight updates while forward/backward computations are performed in FP16.

← Zurück