Summary of On the Efficacy Of Eviction Policy For Key-value Constrained Generative Language Model Inference, by Siyu Ren et al.
On the Efficacy of Eviction Policy for Key-Value Constrained Generative Language Model Inference
by Siyu Ren, Kenny Q. Zhu
First submitted to arxiv on: 9 Feb 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
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 tackles the challenge of deploying Large Language Models (LLMs) in resource-constrained environments, where excessive memory and computational demands are a significant hurdle. The key-value cache, which grows linearly with batch size and sequence length, is a major contributor to these demands. To address this issue, recent works have proposed various eviction policies, but they are limited by their importance score calculation and eviction scope construction methods. This paper identifies the deficiencies of these prior policies and introduces RoCo, a robust cache omission policy based on temporal attention scores and robustness measures. Experimental results demonstrate the superiority of RoCo in prefilling and auto-regressive decoding stages. Additionally, the paper releases EasyKV, a user-friendly software package for key-value constrained generative inference. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about making it possible to use powerful language models on devices with limited resources, like computers or phones. These language models are very good at understanding and generating text, but they require a lot of memory and computing power to work properly. The problem is that these devices don’t have enough resources to handle the demands of these models. To solve this problem, researchers have been working on ways to reduce the amount of memory and computing power needed by these models. This paper looks at some of the methods they’ve tried and proposes a new method called RoCo that works better than the others. The paper also releases a software package called EasyKV that makes it easy for people to use this new method. |
Keywords
» Artificial intelligence » Attention » Inference
