LoRASuite: Efficient LoRA Adaptation Across Large Language Model Upgrades

TL;DR

LoRASuite tackles the problem of adapting LoRA weights when LLM backbones upgrade, which otherwise necessitates costly retraining. It introduces a modular pipeline that uses a transfer matrix for dimensional mismatches, a CKA-based layer-mapping scheme, and a Hungarian-based head-mapping approach, followed by a light fine-tuning stage for numerical stability. Across diverse backbones and tasks, LoRASuite outperforms small-scale LoRA and, in several cases, matches or exceeds full-scale LoRA retraining, while achieving notable memory and compute savings. The approach also demonstrates favorable transfer to other PEFT methods, particularly improving DoRA performance, and offers practical benefits for maintaining up-to-date, task-specific capabilities on growing LLM ecosystems.

Abstract

As Large Language Models (LLMs) are frequently updated, LoRA weights trained on earlier versions quickly become obsolete. The conventional practice of retraining LoRA weights from scratch on the latest model is costly, time-consuming, and environmentally detrimental, particularly as the diversity of LLMs and downstream tasks expands. This motivates a critical question: "How can we efficiently leverage existing LoRA weights to adapt to newer model versions?" To address this, we propose LoRASuite, a modular approach tailored specifically to various types of LLM updates. First, we compute a transfer matrix utilizing known parameters from both old and new LLMs. Next, we allocate corresponding layers and attention heads based on centered kernel alignment and cosine similarity metrics, respectively. A subsequent small-scale, skillful fine-tuning step ensures numerical stability. Experimental evaluations demonstrate that LoRASuite consistently surpasses small-scale vanilla LoRA methods. Notably, on backbone LLMs such as MiniCPM and Qwen, LoRASuite even exceeds the performance of full-scale LoRA retraining, with average improvements of +1.4 and +6.6 points on math tasks, respectively. Additionally, LoRASuite significantly reduces memory consumption by 5.5 GB and computational time by 78.23%.

Figures (15)

• Figure 1: Memory and time comparison between LoRASuite and LoRA retraining.
• Figure 2: Average performance comparison on math tasks for different types of LLM upgrades.
• Figure 3: Average performance comparison on common tasks for different types of LLM upgrades.
• Figure 4: Average performance on math tasks under different settings for the MiniCPM-S-1B to MiniCPM-2B upgrade.
• Figure 5: Performance comparison on math tasks for different LLMs with up_proj and down_proj as target modules.