Summary of All-in-one Tuning and Structural Pruning For Domain-specific Llms, by Lei Lu et al.
All-in-One Tuning and Structural Pruning for Domain-Specific LLMs
by Lei Lu, Zhepeng Wang, Runxue Bao, Mengbing Wang, Fangyi Li, Yawen Wu, Weiwen Jiang, Jie Xu, Yanzhi Wang, Shangqian Gao
First submitted to arxiv on: 19 Dec 2024
Categories
- Main: Computation and Language (cs.CL)
- 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 The proposed All-in-One Tuning and Structural Pruning (ATP) approach is a one-stage structural pruning and fine-tuning method that dynamically identifies the optimal substructure throughout the fine-tuning phase via a trainable pruning decision generator. This approach addresses limitations in existing pruning techniques for large language models, which often combine pruning decisions derived from pretrained weights with fine-tuned weights, leading to potential performance degradation. ATP introduces LoRA-aware forward and sparsity regularization to ensure that learned pruning decisions can be directly removed after the process. The method outperforms state-of-the-art two-stage pruning methods on tasks in legal and healthcare domains, recovering up to 88% and 91% of dense model performance when pruning LLaMA2-7B and LLaMA3-8B models, respectively. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary ATP is a new way to make language models smaller and work better for specific tasks. Right now, people usually take a general-purpose language model and then make it smaller by removing some parts. Then, they fine-tune the smaller model for the specific task. However, this process can be suboptimal because the decisions about what parts to remove are made before the model is fine-tuned. ATP changes this process by making the decision of which parts to remove at the same time as the fine-tuning. This approach works well on tasks in legal and healthcare domains. |
Keywords
» Artificial intelligence » Fine tuning » Language model » Lora » Pruning » Regularization
