Mixture-of-Experts (MoE) models scale more effectively than dense models due to sparse computation through expert routing, selectively activating only a small subset of expert modules. However, sparse computation challenges traditional training practices, as discrete expert routing hinders standard backpropagation and thus gradient-based optimization, which are the cornerstone of deep learning. To better pursue the scaling power of MoE, we introduce GRIN (GRadient-INformed MoE training), which incorporates sparse gradient estimation for expert routing and configures model parallelism to avoid token dropping. Applying GRIN to autoregressive language modeling, we develop a top-2 16×3.8B MoE model. Our model, with only 6.6B activated parameters, outperforms a 7B dense model and matches the performance of a 14B dense model trained on the same data. Extensive evaluations across diverse tasks demonstrate the potential of GRIN to significantly enhance MoE efficacy, achieving 79.4 on MMLU, 83.7 on HellaSwag, 74.4 on HumanEval, and 58.9 on MATH.
@article{arxiv.2409.12136,
title = {GRIN: GRadient-INformed MoE},
author = {Liyuan Liu and Young Jin Kim and Shuohang Wang and Chen Liang and Yelong Shen and Hao Cheng and Xiaodong Liu and Masahiro Tanaka and Xiaoxia Wu and Wenxiang Hu and Vishrav Chaudhary and Zeqi Lin and Chenruidong Zhang and Jilong Xue and Hany Awadalla and Jianfeng Gao and Weizhu Chen},
journal= {arXiv preprint arXiv:2409.12136},
year = {2024}
}