foreword
The solution that jetson cannot install cuda, cudnn, tensorrt separately is much simpler than downloading the SDK manager and flashing the machine. This
method is to directly download the deb package and install it. The deb package installation website
https://repo.download.nvidia.com/jetson/
Install cuda separately
1
sudo apt-get install cuda-toolkit-10-2
2
After the installation is complete, copy the following to the end of the .bashrc file
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-10.2/lib64
export PATH=$PATH:/usr/local/cuda-10.2/bin
export CUDA_HOME=$CUDA_HOME:/usr/local/cuda-10.2
3 You can manually open the .bashrc file, or you can open it directly with the command
sudo gedit ~/.bashrc
Install cudnn separately
Simple method but less controllable:
1 First check what cuDNN is provided by the warehouse. terminal input
sudo apt-cache policy libcudnn8
2 If not, you can check
sudo apt-cache policy libcudnn7
3 and so on
After I installed cuda-10.2 in JetPack 4.5.1 of NX, there is only one version of libcudnn8. Then type
sudo apt-get install libcudnn8
4 is enough. If you want to choose a version
sudo apt-get install libcudnn8=(*.*.*.**+cuda*** )
Select from the list of versions you have previously viewed in parentheses.
Simple method with high selectivity:
According to your own version, select the required version on the website https://repo.download.nvidia.com/jetson/ Instructions:
jtop
View jetson version:
View the version provided by the warehouse
Then search for the following files: (usually under common) for example
Install the following order to install
dpkg -i libcudnn8_8.0.2.39-1+cuda10.2_arm64.deb
dpkg -i libcudnn8-dev_8.0.2.39-1+cuda10.2_arm64.deb
dpkg -i libcudnn8-doc_8.0.2.39-1+cuda10.2_arm64.deb
Install TensorRT separately
method 1
Similar to installing cudnn,
go to the website
https://repo.download.nvidia.com/jetson/
, select version 4.6 or what you need,
search for the corresponding tensosrrt version file below and install it.
For example, if I want to install version 8.0.1, I will search for it on the
webpage These files are downloaded and installed in the following order:
dpkg -i libnvinfer7_7.2.0-1+cuda10.2_arm64.deb
dpkg -i libnvinfer-dev_7.2.0-1+cuda10.2_arm64.deb
dpkg -i libnvinfer-plugin7_7.2.0-1+cuda10.2_arm64.deb
dpkg -i libnvinfer-plugin-dev_7.2.0-1+cuda10.2_arm64.deb
dpkg -i libnvonnxparsers7_7.2.0-1+cuda10.2_arm64.deb
dpkg -i libnvonnxparsers-dev_7.2.0-1+cuda10.2_arm64.deb
dpkg -i libnvparsers7_7.2.0-1+cuda10.2_arm64.deb
dpkg -i libnvparsers-dev_7.2.0-1+cuda10.2_arm64.deb
dpkg -i libnvinfer-bin_7.2.0-1+cuda10.2_arm64.deb
dpkg -i libnvinfer-doc_7.2.0-1+cuda10.2_all.deb
dpkg -i libnvinfer-samples_7.2.0-1+cuda10.2_all.deb
dpkg -i tensorrt_7.2.0.14-1+cuda10.2_arm64.deb
dpkg -i python-libnvinfer_7.2.0-1+cuda10.2_arm64.deb
dpkg -i python-libnvinfer-dev_7.2.0-1+cuda10.2_arm64.deb
dpkg -i python3-libnvinfer_7.2.0-1+cuda10.2_arm64.deb
dpkg -i python3-libnvinfer-dev_7.2.0-1+cuda10.2_arm64.deb
dpkg -i graphsurgeon-tf_7.2.0-1+cuda10.2_arm64.deb
dpkg -i uff-converter-tf_7.2.0-1+cuda10.2_arm64.deb
Or write a bash file and run it several times:
#!/bin/bash
for FILE in `find . -name "*.deb"`
do
sudo apt install ./$FILE
done
Method 2
sudo apt-cache policy tensorrt
Choose the version to install, generally there is no choice, but it is simple and convenient
sudo apt-get install tensorrt=(*.*.*.**+cuda*** )
Pit point:
The tensorrt version is generally bound to python3.6. If there are other methods that can be bound to other python versions, I hope you can send me a copy. Thank you