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

TL;DR

Drop-Upcycling improves training efficiency and long-term performance of sparse Mixture of Experts models by combining knowledge transfer from pre-trained dense models with strategic re-initialization, enabling large-scale models to match dense model performance with fewer resources.

Contribution

This paper introduces Drop-Upcycling, a novel method that enhances MoE training by balancing pre-trained knowledge and re-initialization, leading to better long-term performance and efficiency.

Findings

Outperforms previous MoE methods on large-scale tasks

Achieves comparable performance to larger dense models with fewer FLOPs

Demonstrates effectiveness on models with hundreds of billions of tokens

Abstract

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.