# NVIDIA introduces cuDLA Python bindings for native DLA inference on Jetson Orin

> **Key takeaway:** The new cuda.bindings.cudla module allows Python developers to orchestrate Deep Learning Accelerator (DLA) tasks alongside GPU workloads without writing C++.

**Published:** 2026-11-13T18:57:35+00:00 
**Source:** NVIDIA Technical & Open Source Blog 
**Category:** bigtech-foss 
**Canonical URL:** https://fosswire.org/news/nvidia-introduces-cudla-python-bindings-for-native-dla-infer.html 

## Summary
NVIDIA has released the cuda.bindings.cudla module, a Cython-generated Python layer that exposes the cuDLA C++ API. This allows developers to manage the NVIDIA Deep Learning Accelerator (DLA) directly within Python applications. The DLA is a fixed-function hardware component designed to execute specific neural network operations like convolutions and pooling more efficiently than a general-purpose GPU.

By using these bindings, developers on Jetson AGX Orin can implement hybrid-mode workflows where CUDA streams coordinate DLA inference with GPU-based preprocessing and postprocessing. This offloading mechanism improves power efficiency and preserves GPU capacity for other concurrent tasks. The same API is also applicable to NVIDIA RTX Spark, extending DLA capabilities to Windows-based AI development laptops.

## Technical implementation of cuDLA bindings
- **API Exposure:** The cuda.bindings.cudla module provides thirteen functions, nine data types, and eight enums to facilitate hybrid-mode workflows via CUDA streams.
- **Inference Lifecycle:** The Python workflow involves nine distinct steps, including DLA device initialization, loading compiled binaries from memory, tensor descriptor introspection, and buffer registration.
- **Hardware Decoupling:** Using the `--buildDLAStandalone` flag with TensorRT's trtexec tool allows the creation of raw DLA loadables that can be executed using only the libcudla.so runtime, removing the requirement for TensorRT at inference time.
- **Error Handling:** The bindings map cuDLA calls to standard Python exceptions, utilizing the cudla.Status enum for granular error reporting and standard try/finally cleanup patterns.
- **Memory Management:** The API supports registering GPU buffers with the DLA, enabling efficient data movement between the GPU and the accelerator during hybrid execution.

## Why it matters
This release lowers the barrier for robotics and edge AI developers to utilize specialized hardware without the complexity of C++ runtime management. By enabling Python-native orchestration of the DLA, it simplifies the deployment of power-efficient multi-model pipelines on Jetson platforms and RTX Spark laptops.