LoRITa promotes low-rankness in neural network weights through the composition of linear layers during training, enabling efficient post-training compression without changing the network structure.
