Hybrid Cache

HuggingFace open KV-cache implementation for sliding-window and long-context LLMs.