Summary of Quick: Quantization-aware Interleaving and Conflict-free Kernel For Efficient Llm Inference, by Taesu Kim et al.
QUICK: Quantization-aware Interleaving and Conflict-free Kernel for efficient LLM inference
by Taesu Kim, Jongho Lee, Daehyun Ahn, Sarang Kim, Jiwoong Choi, Minkyu Kim, Hyungjun Kim
First submitted to arxiv on: 15 Feb 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
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 QUICK is a group of optimized CUDA kernels designed for efficient inference of quantized Large Language Models (LLMs). The approach addresses the shared memory bank-conflict problem in state-of-the-art mixed precision matrix multiplication kernels. By interleaving the quantized weight matrices offline, QUICK skips the shared memory write-back after dequantization. This results in up to 1.91x speedup over existing kernels of AutoAWQ for larger batches and up to 1.94x throughput gain on representative LLM models across various NVIDIA GPU devices. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary QUICK is a new way to make Large Language Models (LLMs) work faster on computers. It solves a problem with how other methods store information in memory. QUICK does this by rearranging the data beforehand, so it doesn’t need to write things back and forth as much. This makes it up to 1.91 times faster than other methods for big batches of data and up to 1.94 times faster for specific types of LLM models on different computer chips. |
Keywords
* Artificial intelligence * Inference * Precision
