Is TensorRT-YOLO open source?

Yes — laugh12321/TensorRT-YOLO is open source, released under the GPL-3.0 license.

What language is TensorRT-YOLO written in?

laugh12321/TensorRT-YOLO is primarily written in C++.

How popular is TensorRT-YOLO?

laugh12321/TensorRT-YOLO has 1.9k stars on GitHub.

Where can I find TensorRT-YOLO?

laugh12321/TensorRT-YOLO is on GitHub at https://github.com/laugh12321/TensorRT-YOLO.

← all repositories

laugh12321/TensorRT-YOLO

Deploy every YOLO variant on NVIDIA without linking TensorRT

It packages the sprawling YOLO family—from v3 through YOLO26—into a single deployable NVIDIA runtime that claims zero third-party C++ dependencies.

★1.9k stars C++ Inference · Serving Computer Vision

View on GitHub ↗ Homepage ↗

Not currently ranked — collecting fresh signals.

star history

What it does

TensorRT-YOLO is a C++/Python inference toolkit that optimizes and runs YOLO-family models on NVIDIA hardware. It covers the full task spectrum—detection, segmentation, classification, pose estimation, oriented bounding boxes, and video analysis—using TensorRT plugins, CUDA kernels, and CUDA Graphs for acceleration. A companion export tool converts ONNX models into engines tailored for its runtime.

The interesting bit

The project tries to be aggressively dependency-averse: for C++ consumers, it claims you can get away with including a single trtyolo.hpp header and using the module without linking against CUDA or TensorRT libraries at all. That is an unusual flex for a stack that otherwise lives and breathes NVIDIA-specific optimizations like multi-context parallel inference and Jetson Zero-Copy memory tricks.

Key highlights

Supports a sprawling list of YOLO variants, from YOLOv3 through YOLO26, plus YOLO-World, YOLO-Master, and YOLOE.
Exposes a unified C++ and Python API across detection, segmentation, pose, OBB, and classification tasks.
Uses CUDA Graphs and custom TensorRT plugins to fuse and accelerate both pre- and post-processing.
Offers preprocessing knobs like SwapRB, custom normalization, and border padding, with a claim of pixel-perfect LetterBox reproduction in CUDA.
Targets x86, ARM, Windows, Linux, and TensorRT 10.x, with specific memory optimizations for Jetson edge devices.

Caveats

You cannot simply feed it a standard ONNX export; you must first convert models through the separate trtyolo-export companion project to get a compatible engine.
The README is written in Chinese, with an English version linked but not shown in the provided source, so non-Chinese speakers may need to hunt for details.

Verdict

Grab this if you are shipping YOLO inference on NVIDIA edge or datacenter hardware and want a broad model compatibility layer. Skip it if you are looking for a pure, framework-agnostic ONNX runtime or cannot stomach an extra export toolchain.

Frequently asked

What is laugh12321/TensorRT-YOLO?: It packages the sprawling YOLO family—from v3 through YOLO26—into a single deployable NVIDIA runtime that claims zero third-party C++ dependencies.
Is TensorRT-YOLO open source?: Yes — laugh12321/TensorRT-YOLO is open source, released under the GPL-3.0 license.
What language is TensorRT-YOLO written in?: laugh12321/TensorRT-YOLO is primarily written in C++.
How popular is TensorRT-YOLO?: laugh12321/TensorRT-YOLO has 1.9k stars on GitHub.
Where can I find TensorRT-YOLO?: laugh12321/TensorRT-YOLO is on GitHub at https://github.com/laugh12321/TensorRT-YOLO.