Zeroth-Order Federated Fine-Tuning for Large AI Models in Resource-Constrained Wireless Networks

Large artificial intelligence (AI) models have demonstrated impressive performance in a wide range of fields. Despite their versatility, adapting large AI models to specific downstream applications often requires fine-tuning on decentralized and privacy-sensitive data, posing significant challenges in resource-constrained wireless networks. In this paper, we propose a novel zeroth-order federated fine-tuning framework for efficient fine-tuning large AI models to alleviate the computation, communication, and memory bottleneck issues. Specifically, to address the computation limitation on edge devices, we adopt the split learning architecture, hosting the most computation-intensive component of the large AI model on the edge server. Besides, we employ a memory-efficient zeroth-order fine-tuning algorithm to further reduce the GPU memory consumption. Furthermore, we conduct a rigorous convergence analysis to illustrate how device scheduling influences the learning performance. Based on the analysis, we formulate a global loss minimization problem that jointly optimizes device scheduling, transmit power, and receive beamforming under the average transmission latency constraint. To tackle this complex mixed-integer nonlinear programming problem, we apply Lyapunov theory to break down the long-term optimization problem into multiple subproblems, followed by designing an effective online algorithm. Simulation results show that the proposed framework can lower the GPU memory consumption by up to 84% and GPU hours by 39% compared to the baselines, while achieving comparable accuracy.