Automatic Truncation Position Selection in Singular Value Decomposition for Large Language Models
Keywords: Model decomposition; Large Language Model; Optimization
Abstract: Model decomposition in large language models has drawn much attention due to its superiority and good interpretability, where activation-aware singular value decomposition (SVD) can achieve competitive performance by mitigating reconstruction errors brought by outliers in activation. However, the performance of the state-of-the-art SVD-based LLM compression method is limited to the selection of truncation positions. No work meticulously examines the details of this problem theoretically and empirically tests its correlation with model performance. To fill the research gap, we propose an efficient method that can automatically select truncation positions, namely AutoTrunc. In our work, we first analyze the correlation between truncation positions and the model performance. Then, the model layer importance is modeled based on the correlation, followed by mathematical proof to illustrate how to reach and obtain the optimal truncation position configuration for different layer types. Extensive experiments are carried out to verify our presumption and evaluate our proposed method. Our proposed AutoTrunc outperforms the state-of-the-art SVD-based LLM compression method, with perplexity scores dropping by 24.65% and 38.63% at the compression ratio of 50% in LLaMA-2-7B and LLaMA-2-13B, respectively. The code will be released upon acceptance.
Primary Area: foundation or frontier models, including LLMs
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8816
Loading