一句话:针对商品化 GPU 服务器上低通信带宽与通信争用问题,提出大规模模型微调系统 Mobius。
论文关注在低成本商品化 GPU 服务器上微调大规模深度学习模型这一场景,其目标是让更多人能够负担并使用大规模模型。作者指出,这类服务器上 GPU 间的通信带宽较低,同时通信争用严重,直接导致微调训练效率低下,因此需要专门的系统设计来缓解这一瓶颈。所提供的摘要仅阐述了动机与问题,未披露 Mobius 的具体技术方案、系统设计以及实验评估结果,故方法细节与量化收益无法在此概述。
Fine-tuning on cheap commodity GPU servers makes large-scale deep learning models benefit more people. However, the low inter-GPU communication bandwidth and pressing communication contention on the commodity GPU server obstruct training efficiency.
@inproceedings{paperbot10035,
title = {Mobius: Fine Tuning Large-Scale Models on Commodity GPU Servers},
author = {Yangyang Feng and Minhui Xie and Zijie Tian and Shuo Wang and Youyou Lu and Jiwu Shu},
booktitle = {ASPLOS},
year = {2023},
doi = {10.1145/3575693.3575703}
}