Go to ICLR 2025 Conference homepageRethinking the Impact of Heterogeneous Sublayers in Transformers
Keywords: Transformer, Layer Pruning
Abstract: Large language models (LLMs) using transformers have achieved state-of-the-art performance across a wide array of tasks.
However, the sheer size and complexity of these models present both theoretical and practical challenges, e.g., interpretation of the model behavior and deployment of edge devices.
In this work, we revisit the architecture of transformers and propose a more granular understanding of the impacts of individual sublayers, i.e., Multi-Head Attention (MHA) and Feed-Forward Network (FFN).
We introduce a novel metric, normalized relative impact factor, that allows for progressive, heterogeneous layer pruning.
This metric calculates the relative impact factor of each sublayer on the overall performance, normalized by the number of parameters.
Our experiments demonstrate that our approach can lead to a 20\% reduction in parameters and a 37\% inference speedup, while maintaining minimal performance loss.
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: 6047
Loading