M-ANT: Efficient Low-bit Group Quantization for LLMs via Mathematically Adaptive Numerical Type.
Weiming Hu, Haoyan Zhang, Cong Guo, Yu Feng, Renyang Guan, Zhendong Hua, Zihan Liu, Yue Guan, Minyi Guo, Jingwen Leng
Browse the full HPCA paper archive.