Wangcai Zhao

MOSAIC: Adaptive Inter-layer Composition for Efficient Heterogeneous Vision-Language Models

Yuncheng Yang, Feiyang Ye, Shixian Luo, Yinna Zhu, Lianlei Shan, Wangcai Zhao, et al.

Vision-Language Models (VLMs) have achieved success using homogeneous Transformers to process multimedia data. Recent studies show that heterogeneous structures interleaving efficient mechanisms, like linear attention, improve both performance and inference latency over homogeneo…Read more