Summary of A Deeper Look at Depth Pruning Of Llms, by Shoaib Ahmed Siddiqui et al.
A deeper look at depth pruning of LLMs
by Shoaib Ahmed Siddiqui, Xin Dong, Greg Heinrich, Thomas Breuel, Jan Kautz, David Krueger, Pavlo Molchanov
First submitted to arxiv on: 23 Jul 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 Large Language Models (LLMs) require significant resources and deployment costs in production. Recent work has attempted to prune LLMs by removing unimportant blocks, achieving a 10% reduction without degrading downstream metrics in well-trained models like LLaMa-2 and Mistral 7b. This paper explores different block importance metrics, including adaptive Shapley value, which exhibits a trade-off between tasks due to differences in computed block influences. The analysis is extended to individual self-attention and feed-forward layers, showing that self-attention layers are amendable to pruning, allowing up to 33% removal without performance degradation on MMLU for Mistral 7b. Simple performance recovery techniques using lightweight additive bias or low-rank linear adapters are also explored, achieving a 5% absolute improvement on MMLU. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large Language Models need lots of resources and money to work well in real-world applications. Some researchers have tried to make them more efficient by removing parts that aren’t important. This paper looks at different ways to figure out what’s important and how much it matters. It shows that some methods are better than others, depending on the task you’re trying to do. It also explores how to recover performance after pruning certain parts of the model, which is useful for reducing costs. |
Keywords
» Artificial intelligence » Llama » Pruning » Self attention
