1 paper
Hengjie Cao, Mengyi Chen, Yifeng Yang +13
This work identifies anisotropy in the singular value spectra of parameters, activations, and gradients as the fundamental barrier to low-bit training of large language models (LLM…