OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Infrastructure & Serving

Onnx Tensorrt

Onnx Tensorrt is an open-source C++ parser library and Python backend that parses ONNX models for execution with NVIDIA TensorRT. It also documents command-line workflows for checking whether ONNX models can be parsed and built into TensorRT engines.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Onnx Tensorrt — OpenViST project profile in AI Infrastructure & Serving.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use it when an ONNX model needs to be parsed for NVIDIA TensorRT, when a Python workflow needs to prepare and run a TensorRT engine, or when you want to check a model's engine-build path with the documented command-line tools.

Who it is for

  • Developers integrating ONNX model parsing into C++ applications.
  • Python developers preparing TensorRT engines and running model inference.
  • Teams checking ONNX model parsing and engine-build compatibility in a TensorRT environment.

Documented uses

  • Parse an ONNX model for execution with TensorRT.
  • Prepare a TensorRT engine from an ONNX model and run input data through the Python backend.
  • Use trtexec or Polygraphy to check whether a model can be parsed and built into an engine.
  • Build the C++ parser libraries and executables from source.
Before you choose

Requirements & limits

The public repository is licensed under Apache-2.0.

  • The main branch targets TensorRT 11.4; earlier TensorRT versions have separate branches.
  • The documented dependencies include TensorRT 11.4 and CUDA; Protobuf 3.20.3 or later is optional.
  • TensorRT 11.4 supports ONNX 1.21.0 according to the README.
  • Check the operator support matrix for the ONNX operators used by a model.
  • The InstanceNormalization plugin implementation cannot be used to build version-compatible or hardware-compatible engines.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Check whether an ONNX model builds with TensorRT

Use a documented command-line tool to check whether an ONNX file can be parsed and built into a TensorRT engine.

Before you start

  • A local ONNX model file.
  • A TensorRT environment with either the trtexec binary or the Polygraphy tool available.

Steps

  1. For the C++ tool, run trtexec --onnx=model.onnx.

  2. For the Python tool, run polygraphy run model.onnx --trt.

Expected outcome

The selected tool attempts to parse the ONNX file and build a TensorRT engine.

Guide limitations

  • The README directs users to each tool's help for additional command-line options.
  • Supported ONNX operators are listed in the project's operator support matrix.
Documentation 1 ↗
Source-based practical guide

Build the parser libraries on Linux

Build the ONNX-TensorRT parser libraries and executables from the repository.

Before you start

  • A cloned ONNX-TensorRT repository.
  • TensorRT 11.4 and its open-source libraries.
  • CUDA toolkit; the default lookup path is /usr/local/cuda.
  • CMake and Make.

Steps

  1. From the repository, create and enter a build directory with mkdir build && cd build.

  2. Run cmake .. -DTENSORRT_ROOT= && make -j.

  3. Add the built library location to LD_LIBRARY_PATH using export LD_LIBRARY_PATH=$PWD:$LD_LIBRARY_PATH.

Expected outcome

The documented build process produces parser libraries and executables.

Guide limitations

  • If CUDA is installed somewhere other than /usr/local/cuda, the CMake command needs -DCUDA_TOOLKIT_ROOT_DIR=.
  • The project recommends separate TensorRT repository build instructions for Windows and Docker.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
13 October 2026
Original website title
onnx/onnx-tensorrt · Official homepage title · retained in its source language
How we review sourcesSuggest a correction