Saved in:
| Main Authors: | , , , , , , , , , , , , , , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2502.16982 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866915169769095168 |
|---|---|
| author | Liu, Jingyuan Su, Jianlin Yao, Xingcheng Jiang, Zhejun Lai, Guokun Du, Yulun Qin, Yidao Xu, Weixin Lu, Enzhe Yan, Junjie Chen, Yanru Zheng, Huabin Liu, Yibo Liu, Shaowei Yin, Bohong He, Weiran Zhu, Han Wang, Yuzhi Wang, Jianzhou Dong, Mengnan Zhang, Zheng Kang, Yongsheng Zhang, Hao Xu, Xinran Zhang, Yutao Wu, Yuxin Zhou, Xinyu Yang, Zhilin |
| author_facet | Liu, Jingyuan Su, Jianlin Yao, Xingcheng Jiang, Zhejun Lai, Guokun Du, Yulun Qin, Yidao Xu, Weixin Lu, Enzhe Yan, Junjie Chen, Yanru Zheng, Huabin Liu, Yibo Liu, Shaowei Yin, Bohong He, Weiran Zhu, Han Wang, Yuzhi Wang, Jianzhou Dong, Mengnan Zhang, Zheng Kang, Yongsheng Zhang, Hao Xu, Xinran Zhang, Yutao Wu, Yuxin Zhou, Xinyu Yang, Zhilin |
| contents | Recently, the Muon optimizer based on matrix orthogonalization has demonstrated strong results in training small-scale language models, but the scalability to larger models has not been proven. We identify two crucial techniques for scaling up Muon: (1) adding weight decay and (2) carefully adjusting the per-parameter update scale. These techniques allow Muon to work out-of-the-box on large-scale training without the need of hyper-parameter tuning. Scaling law experiments indicate that Muon achieves $\sim\!2\times$ computational efficiency compared to AdamW with compute optimal training.
Based on these improvements, we introduce Moonlight, a 3B/16B-parameter Mixture-of-Expert (MoE) model trained with 5.7T tokens using Muon. Our model improves the current Pareto frontier, achieving better performance with much fewer training FLOPs compared to prior models.
We open-source our distributed Muon implementation that is memory optimal and communication efficient. We also release the pretrained, instruction-tuned, and intermediate checkpoints to support future research. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_16982 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Muon is Scalable for LLM Training Liu, Jingyuan Su, Jianlin Yao, Xingcheng Jiang, Zhejun Lai, Guokun Du, Yulun Qin, Yidao Xu, Weixin Lu, Enzhe Yan, Junjie Chen, Yanru Zheng, Huabin Liu, Yibo Liu, Shaowei Yin, Bohong He, Weiran Zhu, Han Wang, Yuzhi Wang, Jianzhou Dong, Mengnan Zhang, Zheng Kang, Yongsheng Zhang, Hao Xu, Xinran Zhang, Yutao Wu, Yuxin Zhou, Xinyu Yang, Zhilin Machine Learning Artificial Intelligence Computation and Language Recently, the Muon optimizer based on matrix orthogonalization has demonstrated strong results in training small-scale language models, but the scalability to larger models has not been proven. We identify two crucial techniques for scaling up Muon: (1) adding weight decay and (2) carefully adjusting the per-parameter update scale. These techniques allow Muon to work out-of-the-box on large-scale training without the need of hyper-parameter tuning. Scaling law experiments indicate that Muon achieves $\sim\!2\times$ computational efficiency compared to AdamW with compute optimal training. Based on these improvements, we introduce Moonlight, a 3B/16B-parameter Mixture-of-Expert (MoE) model trained with 5.7T tokens using Muon. Our model improves the current Pareto frontier, achieving better performance with much fewer training FLOPs compared to prior models. We open-source our distributed Muon implementation that is memory optimal and communication efficient. We also release the pretrained, instruction-tuned, and intermediate checkpoints to support future research. |
| title | Muon is Scalable for LLM Training |
| topic | Machine Learning Artificial Intelligence Computation and Language |
| url | https://arxiv.org/abs/2502.16982 |