This is a simple, easy-to-use, high-performance, and multi-end AI inference deployment framework. It is designed based on directed acyclic graphs, abstracting preprocessing, inference, and postprocessing as nodes of the graph, and supporting optimization methods such as pipeline parallelism and task parallelism. It is compatible with multiple inference backends such as TensorRT, OpenVINO, and MNN, and is adapted to mainstream text-to-image, large language, detection and other models, realizing one-code multi-end deployment.