1 paper across 1 session
We propose Fira, a plug-and-play memory-efficient training framework of LLMs to enable full-rank training consistently under the low-rank constraint.
