optimizer-selection

OmniOpt presents a unified framework for optimizer selection in large-scale model training, combining meta-pipeline transformations, norm-constrained linear minimization oracles, and a cross-domain benchmark to systematically analyze optimizer families and their trade-offs.