Post-LoRA Restoration: Utilizing Transferability of Low-Rank Adapter in Quantized Foundation Models

In this study, we consider the transferability of LoRA adapters in quantized foundation models.Specifically, we investigate whether LoRA adapters trained on a low-bit-width foundation model can still function effectively when merged into a higher-bit-width foundation model.By leveraging this transferability, it becomes possible to construct models with performance comparable to conventional LoRA using QLoRA adapters trained under resource-constrained conditions.Our method can be utilized to not only improve the performance of trained QLoRA models without additional training but also accelerate the construction of LoRA.