Quantisation

Weights carry more precision than inference needs. The interesting part is not the rounding but the error compensation — deciding which weights to round first and how to adjust the rest so the layer’s output survives.