Holoscan error on Jetson AGX Xavier: ImportError: libnvcudla.so

Hello guys,

I have tried to use Holoscan on Jetson AGX Xavier DevKit. I pulled a docker image: nvcr.io/nvidia/clara-holoscan/holoscan:v2.0.0-igpu and tried to import tensorrt in python. Nothing else.

I run the container with the command: docker run -it --runtime=nvidia --gpus all --cap-add CAP_SYS_PTRACE --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 …

I have got an error: ImportError: libnvcudla.so: cannot open shared object file: No such file or directory

On my Jetson I have JetPack of version 5.1.1-b56 with Ubuntu 20.04 and CUDA 11.4, but in the Holoscan image there are Ubuntu 22.04 and CUDA 12. May it be because of the older version of CUDA on my Jetson?

If it is the case, how should I run Holoscan on Xavier? Should I choose the older version of Holoscan (0.6.0) with CUDA 11.4 ?

Or is it possible to flash Xavier with something having newer CUDA 12? I was not able to find JetPack for Xavier with the newer version of CUDA and Ubuntu 22.

With the best wishes,

Hi there, from SDK Installation - NVIDIA Docs, it shows the current SDK v2.0 on NVIDIA Jetson AGX Orin and Orin Nano is compatible with Jetpack 6.0. Since you have Jetpack 5.1.1, maybe you could try Holoscan v0.6 like you said. SDK Installation - NVIDIA Docs

Thank you very much. With the 0.6.0 version it works. But I have the last question: if I would like to switch to the newer version of Holoscan, can I install CUDA 12.2 on Xavier, or is it impossible because there is not corresponding version of L4T for Xavier to work with the newer CUDA?

Great to hear version 0.6 worked! From JetPack SDK | NVIDIA Developer I see that

Jetson AI stack packaged with this JetPack 6 release includes CUDA 12.2, Tensorrt 8.6, cuDNN 8.9 and VPI 3.1.

so Jetpack 6.0 will give you CUDA 12.2.
Even though the AGX Xavier is not listed in the matrix in https://docs.nvidia.com/holoscan/sdk-user-guide/sdk_installation.html, you could still try (since on your AGX Xavier with Jetpack 5.1.1 the Holoscan SDK v0.6 worked).