Llama 3 Meets MoE: Efficient Upcycling
This enables cost-effective development of high-capacity models for AI researchers and practitioners, though it is incremental as it builds on existing MoE and pre-training methods.
The paper tackles the high computational cost of scaling large language models by efficiently upcycling a pre-trained dense model into a Mixture-of-Experts model, achieving a 2% improvement in 0-shot accuracy on MMLU with less than 1% of typical pre-training compute.
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.