Summary of Modegpt: Modular Decomposition For Large Language Model Compression, by Chi-heng Lin et al.
MoDeGPT: Modular Decomposition for Large Language Model Compression
by Chi-Heng Lin, Shangqian Gao, James Seale Smith, Abhishek Patel, Shikhar Tuli, Yilin Shen, Hongxia Jin, Yen-Chang Hsu
First submitted to arxiv on: 19 Aug 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computation and Language (cs.CL); Machine Learning (stat.ML)
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 A novel structured compression framework, MoDeGPT, is introduced for Large Language Models (LLMs) that do not require recovery fine-tuning while addressing the drawbacks of existing methods. MoDeGPT partitions Transformer blocks into modules with reduced hidden dimensions and applies three matrix decomposition algorithms: Nyström approximation, CR decomposition, and SVD. This framework matches or surpasses previous structured compression methods in terms of compute costs, achieving 98% savings on a 13B model. Experiments show that MoDeGPT maintains 90-95% zero-shot performance with 25-30% compression rates on LLaMA-2/3 and OPT models. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary MoDeGPT is a new way to make Large Language Models work better on devices with limited resources. It does this by breaking down the model into smaller pieces, reducing the amount of information it needs to remember. This makes the model smaller and faster, but doesn’t affect how well it works. MoDeGPT can compress models by up to 30% without sacrificing performance. It’s a big improvement over other methods that require fine-tuning or sacrifice accuracy. |
Keywords
» Artificial intelligence » Fine tuning » Llama » Transformer » Zero shot
