New optimizer speeds up AI language model training with less memory

Original: MALT: Lightweight Curvature-Aware Muon via Diagonal Preconditioning

Source: arxiv.org ↗