Publications
Make Your LVLM KV Cache More Lightweight
Key-Value (KV) cache has become a de facto component of modern Large Vision-Language Models (LVLMs) for inference. While it enhances decoding efficiency in Large Language Models (LLMs), its direct adoption in LVLMs introduces substantial GPU memory overhead due to the large number of vision tokens processed during the prefill stage. To tackle this problem, we propose LightKV, a novel approach that reduces KV cache size by exploiting the redundancy among vision-token embeddings. Guided by text prompts, LightKV employs cross-modality message passing to aggregate informative messages across vision tokens and progressively compress them during prefill. This prompt-aware guidance distinguishes our method from prior vision-only compression strategies. We evaluate LightKV on eight open-source LVLMs across eight public benchmark datasets, e.g., MME and SeedBench. Experimental results demonstrate that with only 55% of the original vision tokens, LightKV (a) halves the vision-token KV cache size, (b) reduces computation by up to 40%, and (c) preserves general-purpose performance while significantly outperforming existing baselines.
Efficient Corset Selection for Accelerated Vision Instruction-Tuning
Large Vision-Language Models (LVLMs) have demonstrated remarkable performance across a range of cross-modal understanding tasks. However, their supervised fine-tuning (SFT) stage often requires extensive data, leading to substantial challenges given limited resource budgets. In this work, we focus specifically on visual instruction SFT, where models are trained on multimodal instruction–response pairs rather than task-specific adaptation datasets. To address this bottleneck, recent efforts in data efficiency have exclusively relied on coreset selection to produce a reduced dataset of informative samples. All these methods, as found in this work, incur significant resource burdens in both time and additional storage required for coreset selection. Therefore, we propose a novel, resource-light coreset selection method for alleviating this bottleneck. Our method adopts a two-stage design: First, an LLM estimates the linguistic difficulty of each sample without visual input to identify high-language-prior samples. Second, we introduce a biased sampling distribution that favors challenging samples while maintaining data diversity. We evaluate our method on three representative models: LLaVA-1.5-7B, Qwen2-VL-7B, and InternVL2-8B, trained on two general-purpose datasets for visual instruction SFT. Our method consistently outperforms existing state-of-the-art baselines at the same coreset size budgets. More importantly, our approach delivers significant benefits in coreset selection efficiency than these baselines. These results together demonstrate the effectiveness and lightweight nature of our approach for efficient LVLM SFT, especially in resource-limited settings.