cuDLA
A low-level DLA programming interface provided by NVIDIA that allows developers to organize DLA tasks in a near-native manner and invoke them through Python bindings, bypassing the high-level TensorRT wrapper to achieve fine-grained control over the computational graph.
Timeline (last 90 days)
cuDLA provides a lower-level programming model than the TensorRT DLA backend, allowing developers to organize DLA tasks in a near-native manner, and it can be invoked through Python bindings.
cuDLA directly exposes the DLA task submission interface (such as loading an executable graph, setting input and output tensors, triggering inference and synchronization), and developers decide for themselves how to partition and submit the computation graph, bypassing the abstraction layer of TensorRT engine building.
NVIDIA's official developer blog published the technical article "Run Native DLA Inference in Python on NVIDIA Jetson AGX Orin Using cuDLA".
All Facts (3)
cuDLA provides a lower-level programming model than the TensorRT DLA backend, allowing developers to organize DLA tasks in a near-native manner, and it can be invoked through Python bindings.
50%UnverifiedcuDLA directly exposes the DLA task submission interface (such as loading an executable graph, setting input and output tensors, triggering inference and synchronization), and developers decide for themselves how to partition and submit the computation graph, bypassing the abstraction layer of TensorRT engine building.
50%UnverifiedNVIDIA's official developer blog published the technical article "Run Native DLA Inference in Python on NVIDIA Jetson AGX Orin Using cuDLA".
50%