CATP: Cross-Attention Token Pruning for Accuracy Preserved Multimodal Model Inference

arXiv:2404.08567v1 Announce Type: new
Abstract: In response to the rising interest in large multimodal models, we introduce Cross-Attention Token Pruning (CATP), a precision-focused token pruning method. Our approach leverages cross-attention layers in multimodal models, exemplified by BLIP-2, to extract valuable information for token importance determination. CATP employs a refined voting strategy across model heads and layers. In evaluations, CATP achieves up to 12.1X higher accuracy compared to existing token pruning methods, addressing the trade-off between computational efficiency and model precision.

abstract accuracy arxiv attention blip-2 cs.ai cs.cl extract importance inference information large multimodal models multimodal multimodal model multimodal models precision pruning token type voting