Summary of Fast Matrix Multiplications For Lookup Table-quantized Llms, by Han Guo et al.
Fast Matrix Multiplications for Lookup Table-Quantized LLMs
by Han Guo, William Brandon, Radostin Cholakov, Jonathan Ragan-Kelley, Eric P. Xing, Yoon Kim
First submitted to arxiv on: 15 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computation and Language (cs.CL); Distributed, Parallel, and Cluster Computing (cs.DC)
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 paper presents FLUTE, a flexible lookup table engine for large language models (LLMs) with non-uniform, lookup table (LUT) quantization. It addresses the challenge of developing high-performance kernels for weight-quantized LLMs by minimizing bit manipulations and mitigating shared memory bandwidth constraints. The FLUTE kernel can be 2-4x faster than existing GEMM kernels at batch sizes < 32 and a quantization group size of 128, which is typical in LLM inference. As an application, the paper explores a simple extension to lookup table-based NormalFloat quantization and applies it to quantize LLaMA3 to various configurations, achieving competitive quantization performance while increasing end-to-end throughput by 1.5-2 times. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper talks about how we can make computers work faster with big language models. Right now, these models are slowed down because of how they store and use information. The researchers created a new way to do this called FLUTE, which helps the computer process information more quickly. They tested it on a famous language model called LLaMA3 and found that it can do its job up to 2 times faster than before! This is important for making computers smarter and more useful. |
Keywords
* Artificial intelligence * Inference * Language model * Quantization
