Drop-Upcycling：通过部分重初始化训练稀疏专家混合模型

摘要

混合专家（MoE）架构相较于同等容量的密集模型，显著降低了训练和推理成本。升级再利用（Upcycling）是一种利用预训练密集模型初始化和训练MoE模型的方法。尽管升级再利用在初期能带来性能提升，但其训练进度较从零开始训练更为缓慢，导致长期性能欠佳。我们提出了一种名为“丢弃式升级再利用”（Drop-Upcycling）的方法，有效解决了这一问题。该方法巧妙结合了两种看似矛盾的方式：既利用预训练密集模型的知识，又对部分权重进行统计意义上的重新初始化。这一策略性地促进了专家专业化，显著提升了MoE模型在知识获取上的效率。大规模实验表明，Drop-Upcycling在长期训练中，特别是在处理数千亿乃至更多标记时，显著优于以往的MoE构建方法。因此，我们的MoE模型仅需5.9亿活跃参数，就能实现与同一模型家族中130亿参数密集模型相当的性能，同时训练所需的浮点运算量（FLOPs）大约仅为后者的四分之一。所有实验资源，包括源代码、训练数据、模型检查点和日志，均已公开，以促进MoE研究的可重复性和未来探索。

English

The Mixture of Experts (MoE) architecture reduces the training and inference cost significantly compared to a dense model of equivalent capacity. Upcycling is an approach that initializes and trains an MoE model using a pre-trained dense model. While upcycling leads to initial performance gains, the training progresses slower than when trained from scratch, leading to suboptimal performance in the long term. We propose Drop-Upcycling - a method that effectively addresses this problem. Drop-Upcycling combines two seemingly contradictory approaches: utilizing the knowledge of pre-trained dense models while statistically re-initializing some parts of the weights. This approach strategically promotes expert specialization, significantly enhancing the MoE model's efficiency in knowledge acquisition. Extensive large-scale experiments demonstrate that Drop-Upcycling significantly outperforms previous MoE construction methods in the long term, specifically when training on hundreds of billions of tokens or more. As a result, our MoE model with 5.9B active parameters achieves comparable performance to a 13B dense model in the same model family, while requiring approximately 1/4 of the training FLOPs. All experimental resources, including source code, training data, model checkpoints and logs, are publicly available to promote reproducibility and future research on MoE.

Drop-Upcycling：通过部分重初始化训练稀疏专家混合模型

Drop-Upcycling: Training Sparse Mixture of Experts with Partial Re-initialization

摘要

Summary

Support