Summary of Turbo Sparse: Achieving Llm Sota Performance with Minimal Activated Parameters, by Yixin Song et al.
Turbo Sparse: Achieving LLM SOTA Performance with Minimal Activated Parameters
by Yixin Song, Haotong Xie, Zhengyan Zhang, Bo Wen, Li Ma, Zeyu Mi, Haibo Chen
First submitted to arxiv on: 10 Jun 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: 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 Exploiting activation sparsity is a promising approach to accelerate large language models (LLMs) without compromising performance. However, commonly used activation functions like SwiGLU and GeGLU exhibit limited sparsity. The authors propose a novel dReLU function designed to improve LLM activation sparsity, along with a high-quality training data mixture ratio to facilitate effective sparsification. Additionally, the authors leverage sparse activation patterns within Feed-Forward Network (FFN) experts of Mixture-of-Experts (MoE) models to further boost efficiency. By applying the neuron sparsification method to Mistral and Mixtral models, only 2.5 billion and 4.3 billion parameters are activated per inference iteration, respectively, while achieving even more powerful model performance. Evaluation results demonstrate that this sparsity achieves a 2-5x decoding speedup. Remarkably, on mobile phones, the TurboSparse-Mixtral-47B model achieves an inference speed of 11 tokens per second. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about making language models run faster without losing their ability to understand and generate text. The authors found that some activation functions used in these models don’t make them very sparse, which means they use a lot of energy to process information. They created a new kind of activation function called dReLU that makes the model more efficient. By using this function along with high-quality training data, they were able to reduce the number of calculations needed by 2-5 times. This could make it possible for language models to be used on mobile devices, where processing power is limited. |
Keywords
» Artificial intelligence » Inference » Mixture of experts
