Summary of Maskllm: Learnable Semi-structured Sparsity For Large Language Models, by Gongfan Fang et al.
MaskLLM: Learnable Semi-Structured Sparsity for Large Language Models
by Gongfan Fang, Hongxu Yin, Saurav Muralidharan, Greg Heinrich, Jeff Pool, Jan Kautz, Pavlo Molchanov, Xinchao Wang
First submitted to arxiv on: 26 Sep 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: Computation and Language (cs.CL); Machine Learning (cs.LG)
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 MaskLLM, a novel learnable pruning method, is introduced to establish Semi-structured (N:M) Sparsity in Large Language Models (LLMs). By modeling N:M patterns as a learnable distribution through Gumbel Softmax sampling, MaskLLM facilitates end-to-end training on large-scale datasets. This approach yields two key advantages: high-quality masks and transferability across domains or tasks. Empirical results demonstrate substantial improvements over state-of-the-art methods, achieving a perplexity (PPL) of 6.72 on Wikitext compared to the dense model’s 5.12 PPL. MaskLLM can be used for customized masks in downstream tasks or domains, and its learnable nature enables lossless application of 2:4 sparsity. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper introduces a new way to make Large Language Models (LLMs) more efficient without losing their abilities. The approach, called MaskLLM, helps reduce the number of calculations needed during inference by finding and removing redundant information. This method is unique in that it can be trained along with the model’s weights, allowing for better masks and the ability to transfer this efficiency to different tasks or domains. Tests show that MaskLLM outperforms current methods, achieving a score of 6.72 on a specific task compared to the original model’s score of 5.12. |
Keywords
» Artificial intelligence » Inference » Perplexity » Pruning » Softmax » Transferability
