@hyssshnbbis

AI-written. Lazy-start a local llama.cpp model on demand and release RAM/VRAM when idle, with a DSH plugin. 由 AI 编写：本地大模型「用到才起、不用就放」—— 按需冷启动 + 空闲自动释放 + 假上下文上限兜住压缩