Docker runtime to access the GPU Jetson NX

Autonomous Machines Jetson & Embedded Systems Jetson Xavier NX
#1

I’m using a customized BSP which uses L4T32.4.3, to flash the emmc on Jetson NX, attached to a customized board and having issues with the docker runtime to access the GPU with the standard L4T container: The d I’m having issues with the docker runtime to access the GPU with the standard L4T container: nvcr.io/nvidia/l4t-base:r32.5.0 The docker engine cannot access the device. I get the following error:
Docker engine cannot access the device. I get the following error:

Cannot start service : could not select device driver “” with capabilities: [[gpu]]

This exact same compose file works on a NX with an NVidia dev board. Please advise. is the Docker Image and NX image mismatch causing the issue?

#3

This problem may be due to different JetPack versions:

Screenshot from 2021-03-24 10-45-08
Screenshot from 2021-03-24 10-45-081054×721 99.2 KB

#4

Hi,

As mehmetdeniz mentioned, could you give nvcr.io/nvidia/l4t-base:r32.4.3 a try?
Thanks.

#5

I had a similar thought previously but cannot run the following with multiple versions of L4T:

$ docker run --rm --gpus all -it nvcr.io/nvidia/l4t-base:r32.4.3
docker: Error response from daemon: could not select device driver “” with capabilities: [[gpu]].

The deprecated GPU support doesn’t work either:

$ docker run --rm --runtime nvidia -it nvcr.io/nvidia/l4t-base:r32.4.3
docker: Error response from daemon: Unknown runtime specified nvidia.

What should /etc/docker/daemon.json contain?

Should I be using other flags?

What else can I try?

#6

I could not found enough solution about this problem in Jetson platform, but you can try this solution:

sudo systemctl daemon-reload
sudo systemctl restart docker
#7

If you are getting that error, then the nvidia-docker runtime from JetPack appears not to have been properly installed. Did you install those parts of JetPack in your customized BSP?

#8

Hi we have tried to install them after flashing the BSP

#9

Is there a way to install the docker runtime without flashing the BSP image?

#10

Normally it would get installed by SDK Manager. These are the Docker-related packages that get installed:

$ sudo dpkg-query -l | grep nvidia

ii  libnvidia-container-tools                     0.9.0~beta.1                                     arm64        NVIDIA container runtime library (command-line tools)
ii  libnvidia-container0:arm64                    0.9.0~beta.1                                     arm64        NVIDIA container runtime library
ii  nvidia-container-csv-cuda                     10.2.89-1                                        arm64        Jetpack CUDA CSV file
ii  nvidia-container-csv-cudnn                    8.0.0.180-1+cuda10.2                             arm64        Jetpack CUDNN CSV file
ii  nvidia-container-csv-tensorrt                 7.1.3.0-1+cuda10.2                               arm64        Jetpack TensorRT CSV file
ii  nvidia-container-csv-visionworks              1.6.0.501                                        arm64        Jetpack VisionWorks CSV file
ii  nvidia-container-runtime                      3.1.0-1                                          arm64        NVIDIA container runtime
ii  nvidia-container-toolkit                      1.0.1-1                                          arm64        NVIDIA container runtime hook
ii  nvidia-docker2

So you could try installing those packages manually. This list above was gathered from JetPack 4.5.

#11

@dusty_nv thanks for those details! For my case with L4T 32.4.3 I was able to get

docker run --rm --runtime nvidia -it nvcr.io/nvidia/l4t-base:r32.4.3

to run correctly after installing the following (I skipped cudnn and visionworks):

sudo apt install libnvidia-container-tools libnvidia-container0:arm64 nvidia-container-csv-cuda nvidia-container-csv-tensorrt nvidia-container-runtime nvidia-container-toolkit nvidia-docker2

Thank you!