FLM-TopK: Expediting Federated Large Language Model Tuning by Sparsifying Intervalized Gradients

The past few years have witnessed the unprecedented capability of large language models (LLMs). To adapt LLMs with various downstream tasks, fine-tuning methods, e.g., Low-Rank Adaptation (LoRA), are proposed to efficiently tune LLMs. Meanwhile, federated LLM tuning emerges for refining LLMs with clients owning private data. In the federated tuning process, the server and clients frequently exchange fine-tune gradients via Internet, giving the rise of the communication challenge. To overcome this challenge, most existing works employ quantization methods for compressing gradients because sparsification methods like TopK incur heavy overhead for transmitting position IDs (PIDs) of sparsified gradients. In this work, to expedite federated LLM tuning with a higher compression rate, we design the Federated LLM Tuning with TopK (FLM-TopK) algorithm. Specifically, FLM-TopK intervalizes gradients before compression. Then, TopK is separately applied for gradients in each interval so that the overhead representing PIDs is constrained. To optimize our algorithm, we empirically study the distribution of gradients, which obeys the Gaussian distribution. Based on the Gaussian distribution, we establish an optimization problem to minimize the compression error by jointly optimizing the interval size and the sparsification rate per interval. We prove that the non-convex problem can be approximately solved by alternating optimization. To demonstrate the superiority of FLM-TopK, we conduct extensive experiments on nine public datasets. The results demonstrate that FLM-TopK significantly outperforms SOTA baselines, achieving 6.42%-18.87% improvement in accuracy and 17.07%-44.44% reduction in communication traffic.