airockchip/rknn-toolkit2

RKNN-Toolkit2 is a software development kit for converting trained AI models into a format optimized for Rockchip neural processing units and deploying them on Rockchip chips.

The toolkit addresses the challenge of running deep learning models efficiently on Rockchip NPU hardware. It works by accepting trained models on a development computer, converting them into RKNN format through the toolkit, and then enabling inference on Rockchip development boards using either C or Python APIs. The stack includes RKNN-Toolkit2 for model conversion and performance evaluation on PC, RKNN-Toolkit-Lite2 for Python-based deployment on Rockchip NPU platforms, and RKNN Runtime providing C and C++ interfaces for production deployment.

Adoption makes sense for teams building AI applications targeting Rockchip processors across the RK3588, RK3576, RK3566, RK3568, RK3562 series and RV1103, RV1106, RV1126B, and RK2118 chips. The toolkit supports a wide range of Python versions and includes model zoo examples to accelerate development.

The project maintains active development with regular updates to operator support across NPU, GPU, and CPU execution paths. Recent work has expanded capabilities for sparse inference, model encryption, and multi-input output optimization, alongside bug fixes for memory management and runtime stability. The changelog documents improvements to graph optimization and fusion rules that enhance conversion quality for various model architectures.