Confidant: Customizing Transformer-based LLMs via Collaborative Training on Mobile Devices

摘要

Large language models (LLMs) have emerged as a cornerstone for advancing AI technologies. It revolutionizes the way we interact with devices, websites, and information, and paves the way for the development of highly intuitive and capable virtual assistants. Training of today's LLMs happens in cloud data centers due to the requirement of enormous data and a significant amount of computing power. Despite extensive research in mobile edge computing, fine-tuning pre-trained LLMs using resource-constrained devices like commodity smartphones remains highly under-explored. In this paper, we propose Confidant, a practical collaborative training framework that allows modern LLMs to be fine-tuned across multiple off-the-shelf mobile devices. To this end, Confidant partitions an LLM into several sub-models, allowing each of them to fit in the memory of a mobile device. Multiple mobile devices then collaborate to train the LLM by employing a novel pipeline parallel training approach. In specific, Confidant encompasses a memory-aware dynamic model partitioning and intra-device multi-processor scheduler to minimize the training time across heterogeneous platforms. To ensure resilient distributed training, a hybrid fault tolerance mechanism is devised to proactively manage potential device and network failures. We fully implemented Confidant in C++/Python, and built a cross-framework adapter, enabling collaborative training on a variety of mobile platforms. Experimental results show that Confidant excels in achieving computation-, memory-efficient, and robust customization of LLMs - it manages to train state-of-the-art billion-sized LLMs including BERT, GPT-2, Phi2, and LLaMA3, and fine-tunes Phi2-2.7B on Alpaca in just 40.1 hours using three consumer-grade mobile devices.