E8 lattice quantization preserves model accuracy at two-bit precision

The authors jointly quantize eight-weight groups without codebook search, reducing memory while matching the full-precision model’s four-task average.

Chinese Tech
Xinyu Lei · Hong Huang · Jiebin Zhang · Hongyu Li · Jinlong Song · Decheng Wu · +3 more

Peking University · City University of Hong Kong · Tencent

Research Digest··2 min read
Lei et al.

E8-QAT divides weights into groups of eight and projects each group onto the E8 lattice, a regular arrangement of points in eight-dimensional space.

§