NatScale

New NatScale framework optimizes MoE model quantization

Researchers have developed NatScale, a new framework for mixed-precision quantization of Mixture-of-Experts (MoE) language models. This approach aims to reduce storage costs by assigning different bit-widths to experts …