Skywork-moe: A Deep Dive Into Training Techniques For Mixture-of-experts Language Models

Wei Tianwen, Zhu Bo, Zhao Liang, Cheng Cheng, Li Biye, Lü Weiwei, Cheng Peng, Zhang Jianhao, Zhang Xiaoyu, Zeng Liang, Wang Xiaokun, Ma Yutuan, Hu Rui, Yan Shuicheng, Fang Han, Zhou Yahui. Arxiv 2024

[Paper]    
RAG Training Techniques

In this technical report, we introduce the training methodologies implemented in the development of Skywork-MoE, a high-performance mixture-of-experts (MoE) large language model (LLM) with 146 billion parameters and 16 experts. It is initialized from the pre-existing dense checkpoints of our Skywork-13B model. We explore the comparative effectiveness of upcycling versus training from scratch initializations. Our findings suggest that the choice between these two approaches should consider both the performance of the existing dense checkpoints and the MoE training budget. We highlight two innovative techniques: gating logit normalization, which improves expert diversification, and adaptive auxiliary loss coefficients, allowing for layer-specific adjustment of auxiliary loss coefficients. Our experimental results validate the effectiveness of these methods. Leveraging these techniques and insights, we trained our upcycled Skywork-MoE on a condensed subset of our SkyPile corpus. The evaluation results demonstrate that our model delivers strong performance across a wide range of benchmarks.

Similar Work