svd

The paper introduces CERSA, a novel parameter-efficient fine-tuning method that uses singular value decomposition to retain principal components, significantly reducing memory usage while outperforming existing methods like LoRA.