VisCache prunes visual KV caches for efficient VLLM inference

VisCache reduces visual KV cache memory and computation overhead in Vision LLMs, enabling more efficient long-context multimodal inference. The paper is on arXiv (2608.24063).