Docker scripts for building ONNX Runtime with TensorRT and OpenVINO in manylinux environment.
Supports x86_64 and aarch64 (JetPack) architectures.
- CUDA 11.6 (and CUDA 11.1 for tests)
- cuDNN 8.4
- TensorRT 8.4
Place CUDA (.run), cuDNN (tar.gz) and TensorRT (tar.gz) files into distrib folder.
Simply type the following command in your terminal and press Enter:
bash docker-run.sh
Wheels will be placed into wheelhouse folder.
- To specify
Pythonversions for which wheels will be built, editPYTHON_TARGETSvariable indocker-run.sh - To change number of parallel threads edit
THREADS_NUMvariable indocker-run.sh
Wheels compiled for x86_64 architecture depend on the following packages from NVIDIA repository:
nvidia-cuda-runtime-cu116 (11.6)nvidia-cublas-cu116 (11.9)nvidia-cudnn-cu116 (8.4)nvidia-cufft-cu116 (10.7)nvidia-curand-cu116 (10.2)tensorrt (8.5)
and openvino (2022.2) from standard PyPI repository.
Compiled wheels do not explicitly depend on NVIDIA packages, you can install them by the following commands:
pip install --extra-index-url https://pypi.ngc.nvidia.com \
nvidia-cuda-runtime-cu116==11.6.55 \
nvidia-cudnn-cu116==8.4.0.27 \
nvidia-cufft-cu116==10.7.2.124 \
nvidia-curand-cu116==10.2.9.124 \
nvidia-cublas-cu116==11.9.2.110
pip install --no-deps --extra-index-url https://pypi.ngc.nvidia.com tensorrt==8.5.2.2
pip install openvino==2022.2 openvino-dev==2022.2
The recommended way to install this ONNX Runtime package is to use our install.sh script,
which installs ONNX Runtime with all dependencies automatically.
Install GPU version (with all NVIDIA dependencies):
wget -O - https://raw.githubusercontent.com/ENOT-AutoDL/ONNX-Runtime-with-TensorRT-and-OpenVINO/master/install.sh | bash
Install CPU-only version (without NVIDIA packages, use this version if your target device has no GPU):
wget -O - https://raw.githubusercontent.com/ENOT-AutoDL/ONNX-Runtime-with-TensorRT-and-OpenVINO/master/install.sh | bash -s -- -t CPU