Summary of Lexico: Extreme Kv Cache Compression Via Sparse Coding Over Universal Dictionaries, by Junhyuck Kim et al.
Lexico: Extreme KV Cache Compression via Sparse Coding over Universal Dictionaries
by Junhyuck Kim, Jongho Park, Jaewoong Cho, Dimitris Papailiopoulos
First submitted to arxiv on: 12 Dec 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
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 This paper introduces Lexico, a novel key-value cache compression method that leverages sparse coding with a universal dictionary. The authors demonstrate that a small input-agnostic dictionary of approximately 4k atoms can accurately approximate the KV cache in modern language models, enabling efficient compression across different input prompts, tasks, and models. This is achieved through the use of orthogonal matching pursuit for sparse approximation, allowing for flexible compression ratios through direct sparsity control. The authors evaluate Lexico on GSM8K and multiple model families, including Mistral, Llama 3, and Qwen2.5, and demonstrate that it maintains 90-95% of the original performance while using only 15-25% of the full KV-cache memory, outperforming both quantization and token eviction methods. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper creates a new way to make language models use less memory while keeping their performance high. It’s called Lexico, and it uses a special dictionary that can be used for many different tasks and models. This helps language models like Mistral, Llama 3, and Qwen2.5 remember things more efficiently. The results show that Lexico is better than other methods at compressing data without losing too much information. It’s especially good in situations where there isn’t enough memory to use the usual ways of compressing data. |
Keywords
» Artificial intelligence » Llama » Quantization » Token
