NVIDIA samples separate Python model export from C++ inference

DIN Deploy provides C++ examples for bringing speech recognition, segmentation and image generation into local applications. Python exporters convert model checkpoints into ONNX files; the application then uses ONNX Runtime’s session and tensor APIs. TensorRT RTX supplies acceleration, while keeping most application code separate from the vendor-specific operations used by optional accelerated paths.

NVIDIA · Oct 1, 2026

Why it matters

Native application developers get concrete examples of separating model preparation from deployment.

More context

Try one sample using the Windows or Linux CMake presets and a compatible runtime and GPU. CMake downloads ONNX Runtime and TensorRT RTX by default; model export uses checkpoints from Hugging Face.

Keep in mind

These are sample implementations, not production guarantees. Acceleration and quantization depend on hardware; NVIDIA’s reported measurements use DGX Spark. 1min has not built or benchmarked the samples.

Original read