5fa1a76
1
This significantly reduces quantization loss such that you can run models in 4-bit precision without experiencing any performance degradation.
