I have a Jetson AGX Orin with Jetpack 5.1, and therefore came with it CUDA 11.4 and cuDNN 8.6
In order to have a working onnxruntime-gpu, I needed to update it to CUDA 11.8
I first mistakenly installed the CUDA 12.2 version, and then was able to install CUDA 11.8.
Now, although I set this variable in my .zshrc:
export LD_LIBRARY_PATH=/usr/lib/llvm-8/lib
export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda-11/lib64
and
export PATH=$PATH:/usr/local/cuda-11/bin
I still see the following when running the famous jetsonInfo.py script:
$ python3 jetsonInfo.py
NVIDIA Jetson AGX Orin
L4T 35.2.1 [ JetPack 5.1 ]
Ubuntu 20.04.6 LTS
Kernel Version: 5.10.104-tegra CUDA 12.2.140
CUDA Architecture: 8.7
OpenCV version: 4.5.4
OpenCV Cuda: YES
CUDNN: 8.6.0.166
TensorRT: 8.5.2.2
Vision Works: NOT_INSTALLED
VPI: 2.2.4
Vulcan: 1.3.204
Could someone help me set my jetson so that it uses CUDA 11.8 please ?
Without it, onnxruntime_gpu still uses CPU as device instead of the GPU.
(/usr/loca/cuda-11 points to /usr/local/cuda-11.8)
However, the link /usr/local/cuda was pointing on /etc/alternatives/cuda, itself pointing back at /usr/local/cuda-12.2
Even though there are no cuda without versions specified in my PATH and LD_LIBRARY_PATH, changing the link /etc/alternatives/cuda back to /usr/local/cuda-11.8 seems to have fixed my problem.
I don’t know where else there could be something pointing to /usr/local/cuda, but it appears to be somewhere.
Anyway, my issue is currently fixed.