We provide the easy-to-use Docker development environment, rich auxiliary tools and sample packages for each component, which makes it easy for you to quickly start experiencing the development process of model training/conversion, deployment, validation, inference, etc

Accelerated mixed-accuracy computation, greatly improved applicability in the PTQ scheme, and industry-leading quantization accuracy

In scenarios where PTQ scheme accuracy does not meet the requirements, the QAT scheme makes the model accuracy more secure

Richer range of operators, completely covering the common algorithms of Smart Drive, and more flexible operator supporting

Optimization at the model graph encompasses over a hundred improvements, primarily evident in operator substitution, operator splitting, substructure fusion, redundant operator elimination, constant folding, and other aspects

A unified heterogeneous programming interface abstracting and encapsulating the functional hardware on the SoC, while providing function-based API access