# GPTQ

Quantizes weights layer by layer to 4 bits or less, using approximate second-order information to limit error.

## What it is used for

Compressing language models for cheaper GPU inference.