Paper ID: 2409.01162

Balancing Performance and Efficiency: A Multimodal Large Language Model Pruning Method based Image Text Interaction

Gaotong Yu, Yi Chen, Jian Xu

Recently, multimodal large language models (MM-LLMs) have achieved great success in many multimodal tasks, but their high computational costs limit their further promotion and application. In the MM-LLMs framework, the main computational consumption step is the processing of concatenated text and visual tokens at the LLM layer. The length of the input token for LLM directly affects the overall training and inference efficiency. In response to this issue, we further studied the visual tokens of MM-LLMs. We found that the similarity between visual and CLS tokens in the visual encoder follows a long-tail distribution. In other words, only a few visual tokens are highly similar to CLS tokens. Therefore, we designed a dynamic pruning algorithm to address this issue. Firstly, for different input samples, we search for the inflection point of their visual CLS token similarity curve and use it as the corresponding segmentation point to trim the visual markers. This process mainly reduces the output of the visual encoder to accelerate the model. Then, in the LLM layer, the concatenated visual text tokens are pruned for the second time. During this process, due to the interaction between visual and textual features, visual and textual tokens with low text correlation are further filtered, achieving a balance between efficiency and performance. The results on multiple datasets show that our proposed method can achieve performance that competes with the original performance when using an average of 22% of the original token quantity. Our source code will be made publicly available following acceptance.

Submitted: Sep 2, 2024