Search papers, labs, and topics across Lattice.
Nanjing University
1
0
1
One-sided matrix normalization leaves performance on the table: equilibrating both rows and columns accelerates LLM pretraining convergence while completely shedding AdamW's second-moment memory footprint.
