ReAx: Resource-Efficient Asynchronous Execution for Accelerating LLM Fine-Tuning at the Edge

초록

With the widespread use of large language models (LLMs), there is an increasing demand for personalized models that meet diverse needs of users. To enable private, network-independent personalization of LLMs, on-device fine-tuning is receiving much attention. However, on-device fine-tuning faces efficiency and scalability challenges as sequential execution of compute- and memory-intensive operations often underuses resources. In this letter, we propose ReAx, a framework that accelerates on-device fine-tuning through resource-efficient asynchronous parallel execution of memory- and compute-intensive operations. Without increasing memory usage, ReAx improves the average fine-tuning performance and energy consumption by 10.42% and 5.55%, respectively, compared with the baseline. As a positive side effect, asynchronous parameter updates induce gradient noise due to slight delays between streams, which act as a regularizer for adverse updates minimizing accuracy drops.

키워드