Search papers, labs, and topics across Lattice.
This paper introduces Omni-Prune, a novel query-aware token pruning framework designed for efficient processing of audio-video sequences in omnimodal large language models (OmniLLMs). By leveraging adaptive time windows and a scoring mechanism that integrates encoder attention with text-query relevance, Omni-Prune effectively reduces redundancy while preserving critical cross-modal information. Experimental results show that this approach achieves up to 3.25x speedup in prefill latency and a 1.3x reduction in memory usage, all while maintaining over 99% of the performance of the full model.
Achieving a 3.25x speedup in prefill latency without sacrificing model performance could redefine efficiency benchmarks for omnimodal AI systems.
Omnimodal large language models (OmniLLMs) are rapidly extending multimodal reasoning to cover synchronized audio and video. However, the resulting audio-video token sequences are long, leading to high prefill latency and GPU memory usage at inference time. Existing token pruning methods, designed mainly for vision-only inputs, miss both the cross-modal links between audio and video and the user query that decides which content matters. To bridge this gap, we present Omni-Prune, a training-free, query-aware audio-visual token pruning framework that jointly removes redundancy from both modalities while keeping task-relevant cross-modal evidence. Specifically, Omni-Prune first splits the token sequence into adaptive time windows placed at audio saliency peaks, then scores audio and video tokens on a single scale that combines encoder attention with text-query relevance, and pairs related audio-video tokens so that they are kept together. Within each window, a final K-medoids step then selects a few representative tokens, adding diverse cues that score-based selection alone would miss. Extensive experiments demonstrate that Omni-Prune outperforms established baseline methods, delivering up to 3.25x prefill speedup and 1.3x memory reduction while retaining over 99% of full-model performance.