PhD student, The Chinese University of Hong Kong
3 papers at NeurIPS 2025
We propose Fira, a plug-and-play memory-efficient training framework of LLMs to enable full-rank training consistently under the low-rank constraint.
