ZXZuan Xie

Zuan Xie

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Zuan Xie has a total of 5 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Lightweight and Post-Training Structured Pruning for On-Device Large Language Models

    by Zihuai Xu, Yang Xu, Hongli Xu, Yunming Liao, Zhiwei Yao, Zuan Xie

    Considering the hardware-friendly characteristics, structured pruning has emerged as an effective solution to reduce the resource requirements of large language models (LLMs) on resource-constrained devices. Since pruning a certain number of parameters will often reduce the model accuracy, fine-tuning is usually required to recover performance loss. However, fine-tuning relies on high device power and substantial data, making it unsuitable for on-device applications. Recent approaches propose post-training structured pruning techniques that do not require fine-tuning, following three granularities, i.e., neurons, attention heads, and layers. Unfortunately, the previous solutions still suffer from the high memory overhead of substructures’ importance evaluation at different granularities and the problem of only applying to models with specific structures, which limit their scope of applications. This paper proposes COMP, a lightweight and general post-training structured hybrid-granularity pruning method designed for on-device applications. COMP begins by assessing the importance of model layers and neurons individually using distributional distance and the matrix condition number. Subsequently, COMP performs fine-grained neuron pruning and simultaneously determines whether to apply coarse-grained layer pruning or not by comparing the outcomes of the two granularity strategies. Furthermore, COMP implements mask tuning to restore model accuracy without additional fine-tuning, minimizing dependency on external data resources and device power. Experimental results demonstrate that COMP performs well across models with various architectures. When pruning 20% of LLaMA-2-13B parameters, COMP requires less than 8 GB of memory, maintaining more than 90% model accuracy. Meanwhile, COMP improves performance by 3.54% compared to LLM-Pruner, while reducing memory overhead by 90%.