Deepstream 7.0 adding cuda to the docker image

• Jetson Orin NX
• DeepStream Version 7.0
• Questions

Moving my deepstream application (running in a docker container) from DS 6.0.1 to DS 7.0

I see in DeepStream | NVIDIA NGC

that the cuda libraries are no longer in the supplied Docker images for Deepstream (Are the instructions for 7.0 are the same as 7.1 ?, i can not find the 7.0 instructions, is there a different link for 7.0? )

I’ve started using nvcr.io/nvidia/deepstream-l4t:7.0-triton-multiarch as the base FROM in my Dockerfile.
Is this wrong and should I use the samples (nvcr.io/nvidia/deepstream-l4t:7.0-samples-multiarch) one as the base ?

how would I add the cuda libraries back ?
would i do this via apt installs, or is it another 'FROM" in my docker file ?
is there an example Dockerfile that does this ? I’m trying to install the nvidia-cuda apt meta package, but it is not available.. there is an nvidia-cuda-dev package but even that fails to install..

just wondering how everyone else does this ?

thanks,
Tom

It is recommended to upgrade DS-7.1 directly

Noticed that you are using Orin NX. for jetson, deepstream docker image version must be same as host deepstream version. Because the docker image on jeston shared same libraries and userspace driver. such as CUDA.

Before using the Deepstream 7.1 image, please upgrade the Jetpack version of Orin NX to 6.1.

xxx-samples-multiarch just use for deloyment. It lacks a development environment. there shouldn’t be any other difference. xxx-triton-multiarch includes all features

would love to but apparently the gazedetect model support has been dropped from 7.1 … so here I am trying to get it to work with 7.0

ok, but i can only get the cuda libraries installed on the samples one.

What do I need to install on xxx-triton-multiarch to get cuda support inside ? nvidia-cuda is not available and nvidia-cuda-dev fails to install.

Just for my curiosity, why is xxx-triton-multiarch so much bigger than the -samples one ? seeing that samples has all the development environment in it. I would expect the samples one to be much bigger.

No cuda installation required, as long as the Jetpack version is correct, using xxx-triton-multiarch is enough

I have a cuda dependency on the gazedetection (via a nvdsvideotemplate ) that does use cuda directly though.

libcuda.so.1 => not found
libnvdla_compiler.so => not found
libnvcudla.so => not found
libnvbufsurface.so.1.0.0 => not found
libnvcv_faciallandmarks.so => not found

How would I install these dependencies ?

What is your Jetpack version? If you want to run DS-7.0, you need to flash Jetpack 6.0 GA. What is the command line for starting the docker image?

docker run -it --rm --network=host --runtime nvidia  -e DISPLAY=$DISPLAY -w /opt/nvidia/deepstream/deepstream-7.0 -v /tmp/.X11-unix/:/tmp/.X11-unix nvcr.io/nvidia/deepstream:7.0-triton-multiarch

Ah.. i was starting it without any of those values… I forgot.. once i did all the libraries are resolved.

Thank you very much @junshengy !!! you made my week !!