SUMO: Subspace-Aware Moment-Orthogonalization for Accelerating Memory-Efficient LLMTraining

Open in new window