Paper ID: 2502.14504 • Published Feb 20, 2025

PLPHP: Per-Layer Per-Head Vision Token Pruning for Efficient Large Vision-Language Models

Yu Meng, Kaiyuan Li, Chenran Huang, Chen Gao, Xinlei Chen, Yong Li, Xiaoping Zhang
Shenzhen International Graduate School, Tsinghua UniversityTsinghua UniversityTongji University
TL;DR
Get AI-generated summaries with premium
Get AI-generated summaries with premium
Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across a range of multimodal tasks. However, their inference efficiency is constrained by the large number of visual tokens processed during decoding. To address this challenge, we propose Per-Layer Per-Head Vision Token Pruning (PLPHP), a two-level fine-grained pruning method including Layer-Level Retention Rate Allocation and Head-Level Vision Token Pruning. Motivated by the Vision Token Re-attention phenomenon across decoder layers, we dynamically adjust token retention rates layer by layer. Layers that exhibit stronger attention to visual information preserve more vision tokens, while layers with lower vision attention are aggressively pruned. Furthermore, PLPHP applies pruning at the attention head level, enabling different heads within the same layer to independently retain critical context. Experiments on multiple benchmarks demonstrate that PLPHP delivers an 18% faster decoding speed and reduces the Key-Value Cache (KV Cache) size by over 50%, all at the cost of 0.46% average performance drop, while also achieving notable performance improvements in multi-image tasks. These results highlight the effectiveness of fine-grained token pruning and contribute to advancing the efficiency and scalability of LVLMs. Our source code will be made publicly available.

Figures & Tables

Unlock access to paper figures and tables to enhance your research experience.