Prompt-Ladder: Memory-efficient prompt tuning for vision-language models on edge devices

Siqi Cai, Xuan Liu, Jingling Yuan, Qihua Zhou

Abstract: Highlights•We propose a memory-efficient prompt tuning method that unleash the adaptation capability of VLMs.•We theoretically analyze the factors that affect memory usage.•We improve the optimization quality by utilizing existing knowledge of Vlms.•Extensive experiment results indicate the effectiveness of our method.