LoRA Is Slower Than You Think
This work addresses optimization challenges for practitioners fine-tuning LLMs under resource constraints, offering incremental improvements over existing techniques.
The paper tackled the inconsistent speed improvements of Low-Rank Adaptation (LoRA) for fine-tuning large language models by analyzing its limitations and proposing new methods that achieve comparable or superior performance with more consistent training speed gains.
Low-Rank Adaptation (LoRA) is one of the most widely used techniques for fine-tuning large language models (LLMs). By introducing a small number of trainable low-rank weight matrices, LoRA substantially reduces the number of parameters that need to be updated, offering significant advantages in memory consumption and computational efficiency compared to full fine-tuning. However, we observed that LoRA does not consistently provide speed improvements across all model architectures and training setups. Motivated by this inconsistency, we conduct a comprehensive analysis of LoRA's performance and investigate the underlying factors limiting its speedup. Based on our findings, we propose several methods for more efficient fine-tuning of LLMs. We empirically evaluate these methods and compare them to LoRA, demonstrating that our approach achieves comparable or superior performance while delivering more consistent training speed improvements. Our work offers valuable insights and practical guidelines for practitioners seeking to optimize LLM fine-tuning under resource constraints.