Fast and Effective Weight Update for Pruned Large Language Models

Vladimír Boža

Abstract:

Pruning large language models (LLMs) is a challenging task due to their enormous size. The primary difficulty is fine-tuning the model after pruning, which is needed to recover the lost performance caused by dropping weights. Recent approaches have either ignored fine-tuning entirely, focusing on efficient pruning criteria, or attempted layer-wise weight updates, preserving the behavior of each layer. However, even layer-wise weight updates can be costly for LLMs, and previous works have resorted to various approximations.

In our paper, we propose a fast and effective weight update algorithm for pruned layers based on the Alternating Direction Method of Multipliers (ADMM). We further extend it with a simple gradual pruning mask selection and achieve state-of-the-art pruning performance across a wide range of LLMs.

Submission Length: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission:

We reworked sections 2.3 and 3 and added formal statements of theoretical claims.

We added a short subsection about explicit differences between our approach and SparseGPT/Wanda.

We replaced word "accurate" with word "effective" in the title.

We also updated abstract and introduction.

Code: https://github.com/fmfi-compbio/admm-pruning

Assigned Action Editor: ~Gintare_Karolina_Dziugaite1

Submission Number: 2631