# NVIDIA Adds Python Controls for Jetson AGX Orin DLA Inference

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/57000
Published: 2026-11-13T20:11:59.000Z
Updated: 2026-11-13T20:11:59.000Z
Section: Technology

> The `cuda.bindings.cudla` module coordinates supported DLA inference with GPU preprocessing and postprocessing through CUDA streams.

NVIDIA has added a Python interface for running supported inference on the Deep Learning Accelerator in Jetson AGX Orin. It lets developers coordinate DLA work with GPU tasks through CUDA streams, assigning supported neural-network operations to the accelerator while the GPU handles preprocessing and postprocessing.

The interface, cuda.bindings.cudla, is part of the CUDA Python project and is available in cuda-bindings version 13.3.0. It gives Python applications access to cuDLA runtime functions, allowing developers to manage this inference workflow in Python without writing C++ for the orchestration.

Before running a model, developers use TensorRT's trtexec command-line tool to compile it into a file the DLA can load. The --useDLACore=0 option selects the first DLA engine, while --buildDLAStandalone produces a standalone loadable. That file can run on the DLA without TensorRT installed at runtime.

The application initializes CUDA, creates a DLA device in hybrid mode and loads the compiled file. It then allocates GPU memory buffers and registers them with cuDLA before submitting inference work through a CUDA stream. After execution is synchronized, the application retrieves the results. The stream coordinates the inference work with GPU processing in the same workflow.

Python 3.10 or later is required. NVIDIA's Jetson AGX Orin example uses JetPack 7.2.2 or later, which includes the required libcudla.so runtime library. Developers can install the bindings with pip install "cuda-bindings>=13.3.0".

The workflow assigns supported operations to the DLA, which can leave more GPU capacity for other work. It targets edge AI applications where compute and power are constrained. The cuDLA API also applies to NVIDIA RTX Spark, extending the same interface beyond Jetson AGX Orin.