Option --gpus failes with AMD and Intel GPUs

[mviereck]

Description
New option --gpus to provide the GPU to containers failes with AMD and Intel GPUs.

It is specific to systems with NVIDIA GPU and NVIDIA's proprietary driver and NVIDIA's
container runtime setup.

A cli option should be general and not be vendor-specific.

Steps to reproduce the issue:
On a system with an AMD GPU:

$ docker run --rm --gpus all debian echo
docker: Error response from daemon: could not select device driver "" with capabilities: [[gpu]].

Describe the results you received:
Option --gpus all fails on a system with an AMD GPU.
Likely it also fails with Intel GPUs and with NVIDIA GPUs using the nouveau driver.

Describe the results you expected:
Option --gpus all should provide the GPU to the container.

Output of docker --version:
Docker version 19.03.1, build 74b1e89

Discussion
Coming from #1200, #1714 , opening a new ticket:
I want to analyze the current state of --gpus and make proposals:

• Either make GPU support vendor-specific to NVIDIA with docker plugin install and drop the cli option --gpus.
• Or make --gpus work in general for all vendors. I would prefer that.

The current state:

• --gpus works with a specific NVIDIA setup only. Dependencies:
  • NVIDIA GPU
  • NVIDIA proprietary driver on host
  • nvidia-container-toolkit on host
  • nvidia/nvidia-docker image.
• --gpus fails with NVIDIA GPUs not fulfilling above dependencies.
• --gpus fails with AMD GPUs.
• --gpus fails with Intel GPUs.
• --gpus fails with NVIDIA GPUs and nouveau.
• Unknown/to check: Does --gpus work with the combination:
  • NVIDIA GPU
  • NVIDIA proprietary driver on host
  • same NVIDIA proprietary driver in container (arbitrary image)
  • Without nvidia-container-toolkit on host

Desirable:

• Support of other vendors and nouveau and possible existing NVIDIA driver in container. ToDo:
  • --gpus should provide /dev/dri to container.
  • --gpus should provide /dev/nvidia* to container.
  • --gpus maybe should provide /dev/vga_arbiter to container.
  • --gpus maybe should add the container user to groups video and render to support unprivileged users in container.

At least for --gpus all or e.g. --gpus intel this should not be too hard.

Maybe additional, or to be done by the user:

• Driver check on host and in container by docker.
• Driver installation in container by docker. (Probably going too far. However, possible e.g. with NVIDIA's proprietary driver. I am not sure how to accomplish this with MESA drivers, but likely possible, too.)

[tiborvass]

@mviereck thanks, it was always meant to be vendor-neutral. If you want to contribute to the moby/moby repo to add support for other GPU vendors, you are more than welcome! We would need expertise from the various GPU vendors which I don't have. @RenaudWasTaken was providing NVIDIA expertise hence the first implementation.

[mviereck]

If you want to contribute to the moby/moby repo to add support for other GPU vendors, you are more than welcome! We would need expertise from the various GPU vendors which I don't have.

Thanks! I am not experienced with Go, though, so I cannot contribute directly with code.
However, I have some experience in providing GPUs of all vendors to docker containers.
My project x11docker allows to run GUI applications in docker containers, optionally with GPU hardware acceleration.

This works quite well with MESA drivers on host and in image, even if host and container system are quite different (e.g. debian host and alpine image). Given MESA is installed, all you need is described above:

• --gpus should provide /dev/dri to container.
• --gpus should provide /dev/nvidia* to container.
• --gpus maybe should provide /dev/vga_arbiter to container.
• --gpus maybe should add the container user to groups video and render to support unprivileged users in container.

However, this does not cover additional setups like limited GPU memory access. That would indeed need expertise from the vendors. But the probably most common --gpus all would be served.

These packages cover all MESA packages needed in their dependencies:

	Debian and Ubuntu	Arch	Fedora	Alpine
OpenGL	mesa-utils mesa-utils-extra	mesa-demos	glx-utils mesa-dri-drivers	mesa-demos mesa-dri-ati mesa-dri-intel mesa-dri-nouveau mesa-dri-swrast
Video decoding support	libxv1 va-driver-all	libxv libva libva-intel-driver libva-vdpau-driver	libXv libva libva-intel-hybrid-driver libva-vdpau-driver	libxv libva libva-glx libva-intel-driver libva-vdpau-driver

[Alex031544]

You may have a look at this article: The AMD Deep Learning Stack Using Docker by Sam Tehrani

[dimagoltsman]

is there any progress with it?

[thaJeztah]

I don't think contributions have been made to provide support for other GPU vendors, so no change

[MithunKinarullathil]

Checking in, any progress on this? Thank you.

[thaJeztah]

No, no progress (that I'm aware of)

[w8jcik]

It doesn't need to support other vendors, but it should not be crashing like this

$ docker run --gpus all -it ubuntu:20.04
Unable to find image 'ubuntu:20.04' locally
20.04: Pulling from library/ubuntu
Digest: sha256:fd92c36d3cb9b1d027c4d2a72c6bf0125da82425fc2ca37c414d4f010180dc19
Status: Downloaded newer image for ubuntu:20.04
docker: Error response from daemon: could not select device driver "" with capabilities: [[gpu]].
ERRO[0003] error waiting for container: context canceled

Also, look at this error message. It doesn't say Nvidia GPU not detected. It says ERR[0003]. This is not user friendly.