Regions with many nearby points keep subdividing. Regions with few or no points stay large. The tree adapts to the data: dense areas get fine-grained cells, sparse areas stay coarse. The split grid is predetermined (always at midpoints), but the tree only refines cells that need it. Sparse regions stay as single large nodes while dense regions subdivide deeply.
常用于： LLaMA、PaLM、GPT-NeoX 等现代 LLM。
。Line官方版本下载是该领域的重要参考
移动应用（iOS + Android）或 Web 应用——请选择 .task。原生支持 MediaPipe LLM 推理 API：Android 和 iOS 都有现成的示例。在 Flutter 上，flutter_gemma 也开箱即用地支持它。
}The corresponding C++ struct looks like this: