ResLoRA: Identity Residual Mapping in Low-Rank Adaption

TL;DR

ResLoRA is the first work that combines the residual path with LoRA, an improved framework of LoRA that can achieve better results in fewer training steps without any extra trainable parameters or inference cost compared to LoRA.

Abstract

As one of the most popular parameter-efficient fine-tuning (PEFT) methods, low-rank adaptation (LoRA) is commonly applied to fine-tune large language models (LLMs). However, updating the weights of LoRA blocks effectively and expeditiously is challenging due to the long calculation path in the original model. To address this, we propose ResLoRA, an improved framework of LoRA. By adding residual paths during training and using merging approaches to eliminate these extra paths during inference, our method can achieve better results in fewer training steps without any extra trainable parameters or inference cost compared to LoRA. The experiments on NLG, NLU, and text-to-image tasks demonstrate the effectiveness of our method. To the best of our knowledge, ResLoRA is the first work that combines the residual path with LoRA. The code of our method is available at https://github.com/microsoft/LMOps/tree/main/reslora .

Figures (7)

• Figure 1: An illustration of ResLoRA method on accuracy for SVAMPsvamp. ResLoRA achieves a 2.5x faster convergence speed and improves performance by 14.3%.
• Figure 2: Structures of LoRA and ResLoRA
• Figure 3: Results of text-to-image task. We compare images generated by LoRA and ResLoRA$_{is}$.
• Figure 4: Training loss with different $pre\_num$ values on SVAMP. $pre\_num=-1$ means each ResLoRA block uses all previous ResLoRA blocks.
• Figure 5: Difference of the weights of trained matrices between LoRA and ResLoRA$_{bs}$ blocks. We fine-tune models on SVAMP both for 20 epochs, and observe their difference. The ResLoRA$_{bs}$ blocks have been merged.