ReLoRA: High-Rank Training Through Low-Rank Updates
This work addresses the problem of reducing training costs for large-scale AI models, offering a practical solution for researchers and practitioners, though it is incremental as it builds on existing parameter-efficient techniques.
The paper tackles the high computational cost of training large neural networks by introducing ReLoRA, a parameter-efficient training method using low-rank updates, which achieves comparable performance to regular training while saving up to 5.5Gb of GPU RAM and improving training speed by 9-40% for transformer models up to 1.3B parameters.
Despite the dominance and effectiveness of scaling, resulting in large networks with hundreds of billions of parameters, the necessity to train overparameterized models remains poorly understood, while training costs grow exponentially. In this paper, we explore parameter-efficient training techniques as an approach to training large neural networks. We introduce a novel method called ReLoRA, which utilizes low-rank updates to train high-rank networks. We apply ReLoRA to training transformer language models with up to 1.3B parameters and demonstrate comparable performance to regular neural network training. ReLoRA saves up to 5.5Gb of RAM per GPU and improves training speed by 9-40% depending on the model size and hardware setup. Our findings show the potential of parameter-efficient techniques for large-scale pre-training.