Summary of Rotated Runtime Smooth: Training-free Activation Smoother For Accurate Int4 Inference, by Ke Yi et al.
Rotated Runtime Smooth: Training-Free Activation Smoother for accurate INT4 inference
by Ke Yi, Zengke Liu, Jianwei Zhang, Chengyuan Li, Tong Zhang, Junyang Lin, Jingren Zhou
First submitted to arxiv on: 30 Sep 2024
Categories
- Main: Machine Learning (cs.LG)
- 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 proposes a novel activation smoother, Rotated Runtime Smooth (RRS), to address the limitations of existing quantization methods for large language models. The authors observe that outliers in activations can be classified into channel-wise and spike outliers, which hinder the development of INT4 weight-activation quantization. RRS consists of two components: Runtime Smooth (RS) and Rotation operation. RS eliminates channel-wise outliers by smoothing activations with maximums during runtime, while the rotation operation narrows the gap between spike outliers and normal values. The proposed method outperforms state-of-the-art methods in the LLaMA and Qwen families and improves WikiText-2 perplexity for INT4 inference. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps us understand how to make big language models more efficient. Right now, these models are really powerful but they also use a lot of computer power and memory. To fix this, researchers have been trying to shrink the models down while keeping their performance good. They found that some parts of the model get really big and noisy, which makes it hard to shrink them without losing too much information. The authors propose a new way to smooth out these noisy areas, called Rotated Runtime Smooth (RRS). It works by looking at the maximum value in each part of the model and using that to calm down the noise. This helps the model work better when it’s shrunk down. |
Keywords
* Artificial intelligence * Inference * Llama * Perplexity * Quantization
