English

Llama 3 Meets MoE: Efficient Upcycling

Machine Learning 2024-12-16 v1

Abstract

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%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.

Keywords

Cite

@article{arxiv.2412.09952,
  title  = {Llama 3 Meets MoE: Efficient Upcycling},
  author = {Aditya Vavre and Ethan He and Dennis Liu and Zijie Yan and June Yang and Nima Tajbakhsh and Ashwath Aithal},
  journal= {arXiv preprint arXiv:2412.09952},
  year   = {2024}
}
R2 v1 2026-06-28T20:33:36.067Z