Triton enables write CUDA kernels. To integrate an kernel into TensorRT is necessary create a plugin.
Since TensorRT 9.1 version is possible create plugins using Python API.
In this tutorial we will learn this.
Do you need:
pip install -U torch tensorrt cuda-python onnx onnx_graphsurgeon
python tensorrt_python_plugin.py
For more details please read the poster in medium:
https://vilsonrodrigues.medium.com/writing-tensorrt-plugins-using-triton-and-python-9dfcc7af73cd