Energy-Efficient Split Learning for Fine-Tuning Large Language Models in Edge Networks
Zuguang Li, Shaohua Wu, Liang Li, Songge Zhang
TL;DR
This work tackles the challenge of energy- and time-efficient fine-tuning of large language models (LLMs) in geo-distributed edge networks under privacy constraints. It introduces an energy-efficient split-learning framework that uses LoRA-based fine-tuning and an adaptive CARD algorithm to jointly select the split layer and allocate edge-server compute resources. Key contributions include a detailed delay-energy model, a weak Pareto objective for joint optimization, and a decomposition into upper- and lower-layer problems with a convex solution for the server side and brute-force search for the cut layer; simulations on a 5-device platform demonstrate substantial improvements, achieving up to a 70.8% reduction in training delay and a 53.1% reduction in server energy compared with baselines. The approach enables privacy-preserving, scalable, and energy-conscious LLM fine-tuning in practical edge deployments, with potential for robust operation in dynamic wireless environments.
Abstract
In this letter, we propose an energy-efficient split learning (SL) framework for fine-tuning large language models (LLMs) using geo-distributed personal data at the network edge, where LLMs are split and alternately across massive mobile devices and an edge server. Considering the device heterogeneity and channel dynamics in edge networks, a \underline{C}ut l\underline{A}yer and computing \underline{R}esource \underline{D}ecision (CARD) algorithm is developed to minimize training delay and energy consumption. Simulation results demonstrate that the proposed approach reduces the average training delay and server's energy consumption by 70.8% and 53.1%, compared to the benchmarks, respectively.