TGV-KV: Text-Grounded KV Eviction for Vision-Language Models

Vision-Language Models (VLMs) inherit the auto-regressive generation paradigm and cache the keys and values (KV) of all previous tokens to accelerate inference, resulting in memory consumption that scales linearly with context length. This issue is particularly pronounced in VLMs due to substantial redundancy in the visual modality. Although KV cache eviction approaches can effectively reduce inference memory, they often incur significant performance degradation in VLMs, as most are designed for language models and overlook the inherent gap between text and vision. By systematically analyzing the modality gap in VLMs in this work, we argue that the importance of visual information should be grounded in textual guidance and accordingly propose a Text-Grounded KV Eviction method for VLMs (TGV-KV). TGV-KV comprises three submodules: (1) Text-Vision Budgeting (TVB) assigns budget to each layer based on the mutual information interaction. (2) Text-Weighted Ranking (TWR) assesses the priority of text and ranks vision importance based on weighted text-image attention. (3) Text-Prioritised Retention (TPR) policy strategically preserves text KV to avoid acute information loss. We evaluate TGV-KV across five models with different sizes and architectures, showing that TGV-KV preserves 99.2% full-KV accuracy on the VizWiz-VQA task with LLaVA-NeXT and boosts end-to-end throughput by 52.6% with an extreme retention budget of 5%. Code is available at https://github.com/Danielement321/TGV-KV.

Straightforward Layer-wise Pruning for More Efficient Visual Adaptation

Parameter-efficient transfer learning (PETL) aims to adapt large pre-trained models using limited parameters. While most PETL approaches update the added parameters and freeze pre-trained weights during training, the minimal impact of task-specific deep layers on cross-domain data poses a challenge as PETL cannot modify them, resulting in redundant model structures. Structural pruning effectively reduces model redundancy; however, common pruning methods often lead to an excessive increase in stored parameters due to varying pruning structures based on pruning rates and data. Recognizing the storage parameter volume issue, we propose a Straightforward layer-wise pruning method, called SLS, for pruning PETL-transferred models. By evaluating parameters from a feature perspective of each layer and utilizing clustering metrics to assess current parameters based on clustering phenomena in low-dimensional space obtained through t-SNE, SLS facilitates informed pruning decisions. Our study reveals that layer-wise pruning, with a focus on storing pruning indices, addresses storage volume concerns. Notably, mainstream Layer-wise pruning methods may not be suitable for assessing layer importance in PETL-transferred models, where the majority of parameters are pre-trained and have limited relevance to downstream datasets. Comparative analysis against state-of-the-art PETL methods demonstrates that the pruned model achieved a notable balance between model throughput and accuracy. Moreover, SLS effectively reduces storage overhead arising from varying pruned structures while enhancing the accuracy and speed of pruned models compared to conventional pruning methods.