Iterative Structured Pruning for Large Language Models with Multi-Domain Calibration
Guangxin Wu, Hao Zhang, Zhang Zhibin, Jiafeng Guo, Xueqi Cheng
TL;DR
This work tackles the challenge of deploying large language models by proposing a structured pruning framework that combines a domain-diverse calibration set with an iterative pruning strategy. By using a bias-compensation mechanism and a variance-based channel importance criterion, the method identifies and removes redundant channels while preserving performance across diverse downstream tasks. Empirical results on Qwen2.5 models demonstrate substantial compression with minimal accuracy degradation, and ablations confirm the value of multi-domain calibration and iterative pruning, especially under aggressive pruning. The approach promises practical gains for real-world deployment by aligning pruning decisions with heterogeneous input distributions and progressive, stable optimization.
Abstract
Large Language Models (LLMs) have achieved remarkable success across a wide spectrum of natural language processing tasks. However, their ever-growing scale introduces significant barriers to real-world deployment, including substantial computational overhead, memory footprint, and inference latency. While model pruning presents a viable solution to these challenges, existing unstructured pruning techniques often yield irregular sparsity patterns that necessitate specialized hardware or software support. In this work, we explore structured pruning, which eliminates entire architectural components and maintains compatibility with standard hardware accelerators. We introduce a novel structured pruning framework that leverages a hybrid multi-domain calibration set and an iterative calibration strategy to effectively identify and remove redundant channels. Extensive experiments on various models across diverse downstream tasks show that our approach achieves significant compression with minimal performance degradation.