Every FLOP Counts: Scaling a 300B Mixture-of-Experts LING LLM without Premium GPUs
Ling Team, Binwei Zeng, Chao Huang, Chao Zhang, Changxin Tian, Cong Chen, Dingnan Jin, Feng Yu, Feng Zhu, Feng Yuan, Fakang Wang, Gangshan Wang, Guangyao Zhai, Haitao Zhang, Huizhong Li, Jun Zhou, Jia Liu, Junpeng Fang, Junjie Ou, Jun Hu, Ji Luo, Ji Zhang, Jian Liu, Jian Sha, Jianxue Qian, Jiewei Wu, Junping Zhao, Jianguo Li, Jubao Feng, Jingchao Di, Junming Xu, Jinghua Yao, Kuan Xu, Kewei Du, Longfei Li, Lei Liang, Lu Yu, Li Tang, Lin Ju, Peng Xu, Qing Cui, Song Liu, Shicheng Li, Shun Song, Song Yan, Tengwei Cai, Tianyi Chen, Ting Guo, Ting Huang, Tao Feng, Tao Wu, Wei Wu, Xiaolu Zhang, Xueming Yang, Xin Zhao, Xiaobo Hu, Xin Lin, Yao Zhao, Yilong Wang, Yongzhen Guo, Yuanyuan Wang, Yue Yang, Yang Cao, Yuhao Fu, Yi Xiong, Yanzhe Li, Zhe Li, Zhiqiang Zhang, Ziqi Liu, Zhaoxin Huan, Zujie Wen, Zhenhang Sun, Zhuoxuan Du, Zhengyu He
TL;DR
The work tackles the cost and resource barriers of training 300B-scale Mixture-of-Experts LLMs by introducing Ling-Lite and Ling-Plus, two open-source MoE models designed for efficiency and accessibility. It combines architectural innovations (fine-grained experts, NormHead, dropless routing, and Stochastic Routing Warmup) with system-level optimizations (DLRover, XPUTimer, EDiT, PCache, Babel, Flood) to enable training on heterogeneous, commodity hardware while maintaining competitive performance. Through a comprehensive pre-training on a 9T-token corpus and extensive post-training alignment (SFT, DPO, RL optimization, long-context, and tool-use data synthesis), the authors demonstrate cost savings of around 20% on pre-training and robust tool-use capabilities. The results show Ling-Lite and Ling-Plus achieving competitive benchmarks across languages, coding/math tasks, and safety metrics, with strong cross-platform robustness and significant accelerations in evaluation and inference workloads, highlighting the practical potential of resource-constrained MoE AI development.
Abstract
In this technical report, we tackle the challenges of training large-scale Mixture of Experts (MoE) models, focusing on overcoming cost inefficiency and resource limitations prevalent in such systems. To address these issues, we present two differently sized MoE large language models (LLMs), namely Ling-Lite and Ling-Plus (referred to as "Bailing" in Chinese, spelled Bǎilíng in Pinyin). Ling-Lite contains 16.8 billion parameters with 2.75 billion activated parameters, while Ling-Plus boasts 290 billion parameters with 28.8 billion activated parameters. Both models exhibit comparable performance to leading industry benchmarks. This report offers actionable insights to improve the efficiency and accessibility of AI development in resource-constrained settings, promoting more scalable and sustainable technologies. Specifically, to reduce training costs for large-scale MoE models, we propose innovative methods for (1) optimization of model architecture and training processes, (2) refinement of training anomaly handling, and (3) enhancement of model evaluation efficiency. Additionally, leveraging high-quality data generated from knowledge graphs, our models demonstrate superior capabilities in tool use compared to other models. Ultimately, our experimental findings demonstrate that a 300B MoE LLM can be effectively trained on lower-performance devices while achieving comparable performance to models of a similar scale, including dense and MoE models. Compared to high-performance devices, utilizing a lower-specification hardware system during the pre-training phase demonstrates significant cost savings, reducing computing costs by approximately 20%. The models can be accessed at https://huggingface.co/inclusionAI.