I just flashed an AGX Orin with JP 5.0 Dev Preview. I expected Docker to be installed by default but it does not appear to be. Is there a step I missed to the flashing process? How did this get missed during the flashing process?
Do you also install the components as well?
Or please run the following command to install all the packages:
$ sudo apt update $ sudo apt install nvidia-jetpack
@AastaLLL I installed nvidia-jetpack as recommended. It installed without problems.
Now I have my Docker environment setup however my containers cannot find the libcublas.so.10 library. The exact error is:
“ImportError: libcublas.so.10: cannot open shared object file: No such file or directory”
However, running “ldconfig -v” shows me the following which indicates a new version of the library is installed currently:
user@jetson:~/user/Benchmarking/-scripts$ ldconfig -v | grep cublas
libcublasLt.so.11 → libcublasLt.so.126.96.36.199
libcublas.so.11 → libcublas.so.188.8.131.52
Clearly the software in my container was compiled with the older version of JP and libcublas however I can’t recompile it right now. I need some substitution method to introduce libcublas.so.10 into the environment alongside libcublas.so.11.
So the question becomes:
- Can I install libcublas.so.10 alongside the libcublas.so.11 already installed? If so, how?
- If I cannot install libcublas.so.10 in parallel, can I create a symlink from this name to libcublas.so.11?
- Would it be better to remove nvidia-jetpack and install only specific libraries/packages required to get libcublas.so.10?
- What is the Nividia recommended path forward?
Please noted that JetPack 5.0DP includes CUDA 11.
So there are no CUDA 10 libraries (libcublas.so.10) installed.
Could you check which app/library has the dependency on CUDA 10?
If it is a re-compilable app, you can just recompile it in the container with CUDA 11 directly.