Summary of Self-selected Attention Span For Accelerating Large Language Model Inference, by Tian Jin et al.
Self-Selected Attention Span for Accelerating Large Language Model Inference
by Tian Jin, Wanzin Yazar, Zifei Xu, Sayeh Sharify, Xin Wang
First submitted to arxiv on: 14 Apr 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 The authors optimize the inference computation of large language models (LLMs) on modern GPUs, which is inefficient due to the increasing number of tokens attended. They leverage the problem-solving capabilities of LLMs to fine-tune an LLM for two specific tasks: evaluating complex arithmetic expressions and summarizing news articles. The goal is twofold: learn to solve the task and identify minimal attention spans required. The fine-tuned model uses these self-identified minimal attention spans to create sparse attention masks during inference, improving throughput by 28% with a custom CUDA kernel. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models can do many things, but they’re not very efficient when it comes to doing them on computers. That’s because they have to look at more and more words as they generate new ones. To make this process faster, the authors used these same language models to teach themselves how to be more efficient. They showed that by doing this, the models can solve tasks like evaluating math problems or summarizing news articles much faster. This is important because it could help computers do many things faster and more efficiently. |
Keywords
» Artificial intelligence » Attention » Inference
