Nvidia-container with GPU

[HuiW]

Hi

Is there a way to know if nvidia-container works with GPU?

We installed nvidia-container by sdkmanager.

And check nvidia-container is installed with below command.

#sudo dpkg --get-selections | grep nvidia
libnvidia-container-tools install
libnvidia-container0:arm64 install
nvidia-container-runtime install
nvidia-docker2 install

However, as running
nvidia-container-cli command and got GPU UUID and Bus Location are null. ( see below)

aae@aae-desktop:~$ nvidia-container-cli info
NVRM version: (null)
CUDA version: 10.2Device Index: 0
Device Minor: 0
Model: Xavier
Brand: (null)
GPU UUID: (null)
Bus Location: (null)
Architecture: 7.2
aae@aae-desktop:~$ nvidia-container-cli --load-kmods info
NVRM version: (null)
CUDA version: 10.2Device Index: 0
Device Minor: 0
Model: Xavier
Brand: (null)
GPU UUID: (null)
Bus Location: (null)
Architecture: 7.2

Is nvidia-container-cli is proper way to know if the container works with GPU or not?

Is there a way to know if nvidia-container works with GPU?

Thank you for any advice,

[dusty_nv]

Hi @HuiW, I haven’t used that nvidia-container-cli tool before, but it does seem to be detecting your GPU. On Jetson, normally you would use --runtime nvidia when you start the container. What version of JetPack-L4T are you on?

[HuiW]

Hi Dusty_nv,

Thank you for your prompt support.
Would you address more " it does seem to be detecting your GPU?"

We did start the container with --runtime nvidia.
However, should I use nvidia-docker to replace docker to enable gpu in a container?

We did test Jetpack 4.6.2 and Jetpack5.0.2.

Thank you,

[dusty_nv]

@HuiW

However, should I use nvidia-docker to replace docker to enable gpu in a container?

You should start the containers like the documentation shows on the container’s NGC page - you don’t need to use nvidia-docker, just docker with --runtime nvidia flag. For example, to check GPU in l4t-pytorch container on JetPack 5.0.2:

$ sudo docker run -it --rm --runtime nvidia --network host nvcr.io/nvidia/l4t-pytorch:r35.1.0-pth1.12-py3
# python3 -c 'import torch; print(torch.cuda.is_available())'
# cd /usr/local/cuda/samples/1_Utilities/deviceQuery
# make
# ./deviceQuery

You could also run this test script in l4t-pytorch container that allocates some data on the GPU and runs some test kernels: https://github.com/dusty-nv/jetson-containers/blob/master/test/test_pytorch.py

[HuiW]

Hi dusty_nv,

Thank you for providing the container.
And deviceQuery does show the information of CUDA and GPU on the container.

Thanks,