The sparsely activated mixture of experts (MoE) model presents an effective alternative to densely activated (dense) models, combining improved accuracy with computational efficiency. However, training MoE models from scratch requires extensive data and computational resources, a challenge that limits their widespread adoption. To address this, we introduce MoE Jetpack, a framework designed to fine-tune the abundant and easily accessible dense checkpoints into MoE models.

The sparsely activated mixture of experts (MoE) model presents an effective alternative to densely activated (dense) models, combining improved accuracy with computational efficiency. However, training MoE models from scratch requires extensive data and computational resources, a challenge that limits their widespread adoption. To address this, we introduce MoE Jetpack, a framework designed to fine-tune the abundant and easily accessible dense checkpoints into MoE models. MoE Jetpack incorporates two key techniques: (1) checkpoint recycling, which initializes MoE models with dense checkpoints to accelerate convergence and enhance accuracy, minimizing the need for extensive pre-training; (2) the hyperspherical adaptive MoE (SpheroMoE) layer, which optimizes the MoE architecture to enhance fine-tuning performance and efficiency.Experimental results indicate that MoE Jetpack doubles the convergence speed and enhances accuracy by 2.8% on ImageNet-1K.

Scaling large language models (LLMs) significantly improves performance but comes with prohibitive computational costs. Mixture-of-Experts (MoE) models offer an efficient alternative, increasing capacity without a proportional rise in compute requirements. However, training MoE models from scratch poses challenges like overfitting and routing instability. We present an efficient training recipe leveraging pre-trained dense checkpoints, training an 8-Expert Top-2 MoE model from Llama 3-8B with less than $1\%$ of typical pre-training compute. Our approach enhances downstream performance on academic benchmarks, achieving a $\textbf{2%}$ improvement in 0-shot accuracy on MMLU, while reaching a Model FLOPs Utilization (MFU) of $\textbf{46.8%}$ during training using our framework. We also integrate online upcycling in NeMo for seamless use of pre-trained weights, enabling cost-effective development of high-capacity MoE models.

Training large, deep neural networks to convergence can be prohibitively expensive. As a result, often only a small selection of popular, dense models are reused across different contexts and tasks. Increasingly, sparsely activated models, which seek to decouple model size from computation costs, are becoming an attractive alternative to dense models. Although more efficient in terms of quality and computation cost, sparse models remain data-hungry and costly to train from scratch in the large scale regime. In this work, we propose sparse upcycling -- a simple way to reuse sunk training costs by initializing a sparsely activated Mixture-of-Experts model from a dense checkpoint. We show that sparsely upcycled T5 Base, Large, and XL language models and Vision Transformer Base and Large models, respectively, significantly outperform their dense counterparts on SuperGLUE and ImageNet, using only ~50% of the initial dense pretraining sunk cost. The upcycled models also outperform sparse models trained from scratch on 100% of the initial dense pretraining computation budget.