DeAOff: Dependence-aware offloading of decoder-based generative models for edge computing
Ning Jiahong, Yang Tingting, Ce Zheng, Xinghan Wang, Feng Ping, Zhang Xiufeng · China Communications · 2025
This paper presents an algorithm named the dependency-aware offloading framework (DeAOff), which is designed to optimize the deployment of Gen-AI decoder models in mobile edge computing (MEC) environments. These models, such as decoders, pose significant challenges due to their interlayer dependencies and high computational demands, especially under edge resource constraints. To address these challenges, we propose a two-phase optimization algorithm that first handles dependency-aware task allocation and subsequently optimizes energy consumption. By modeling the inference process using directed acyclic graphs (DAGs) and applying constraint relaxation techniques, our approach effectively reduces execution latency and energy usage. Experimental results demonstrate that our method achieves a reduction of up to 20% in task completion time and approximately 30% savings in energy consumption compared to traditional methods. These outcomes underscore our solution's robustness in managing complex sequential dependencies and dynamic MEC conditions, enhancing quality of service. Thus, our work presents a practical and efficient resource optimization strategy for deploying models in resource-constrained MEC scenarios.