Summary of Stun: Structured-then-unstructured Pruning For Scalable Moe Pruning, by Jaeseong Lee et al.
STUN: Structured-Then-Unstructured Pruning for Scalable MoE Pruning
by Jaeseong Lee, seung-won hwang, Aurick Qiao, Daniel F Campos, Zhewei Yao, Yuxiong He
First submitted to arxiv on: 10 Sep 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computation and Language (cs.CL)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper proposes a novel approach to prune Mixture-of-Experts (MoEs) in Large Language Models (LLMs), aiming to reduce inference costs while maintaining performance. MoEs have been used to reduce inference costs by sparsely activating experts, but the massive number of experts still makes them expensive to serve. The authors study how to address this issue by pruning MoEs using unstructured and structured pruning methodologies. They find that expert pruning, a form of structured pruning, can actually outperform unstructured-only pruning. To scale expert pruning for recent MoEs, the authors propose a scalable alternative with O(1) complexity that leverages latent structure between experts based on behavior similarity. The proposed method is highly effective and achieves nearly no loss in performance with 40% sparsity using only one H100 and two hours. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper helps to reduce the cost of running large language models by pruning them, making them more efficient. It shows that a special kind of pruning called expert pruning can work better than other methods. The authors also suggest a way to make this method faster and more practical for bigger models. This could be useful for people who want to use these models in real-life applications. |
Keywords
» Artificial intelligence » Inference » Mixture of experts » Pruning
