All terms
Optimization
NF4
A 4-bit number format for shrinking model weights with very little quality loss.
Definition
NF4, short for 4-bit NormalFloat, is a way of storing each of a model's weights using only four bits, with the small set of available values chosen to match how weights are normally spread out. It is the format behind QLoRA fine-tuning and lets very large models fit on a single graphics card with little loss in quality.