Summary of Fairness in Serving Large Language Models, by Ying Sheng et al.
Fairness in Serving Large Language Models
by Ying Sheng, Shiyi Cao, Dacheng Li, Banghua Zhu, Zhuohan Li, Danyang Zhuo, Joseph E. Gonzalez, Ion Stoica
First submitted to arxiv on: 31 Dec 2023
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: Machine Learning (cs.LG); Performance (cs.PF)
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 ensuring fairness in serving large language models (LLMs) while processing client requests. The current approach to fair scheduling, which involves rate limits, leads to underutilization and poor client experience when there is spare capacity. To address this issue, the authors propose a novel scheduling algorithm called Virtual Token Counter (VTC), which takes into account the unpredictable request lengths and batching characteristics of LLMs on parallel accelerators. The VTC algorithm ensures fairness by minimizing the difference in service between two backlogged clients while adhering to the work-conserving requirement. Experimental results demonstrate the superior performance of VTC compared to baseline methods under various conditions. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about making sure that when you use big language models like ChatGPT, all users get treated fairly. Right now, some services have limits on how many requests they can process at once, which means some people might not be able to use the model as much as others. This isn’t very good for anyone involved! To fix this problem, the researchers came up with a new way of scheduling these big language models that takes into account how long each request is and when they’re all being processed together at once. Their new method, called Virtual Token Counter (VTC), makes sure everyone gets treated fairly and uses the model efficiently. |
Keywords
* Artificial intelligence * Token
