nvidia-docker2 support

[niXta1]

Is your feature request related to a problem? Please describe.
Passing through nvidia GPU to a container.

Describe the solution you'd like
Be able to use docker compose v2.3 or find a better way to get GPU to a container

Describe alternatives you've considered

1. Override (needs to be using compose v2.3 i think?).
2. Set the nvidia GPU to be used in all containers by modifying docker config.

This is on Ubuntu 18 with latest stabile docker.

[open-collective-bot]

Hey @niXta1 👋,

Thank you for opening an issue. We will get back to you as soon as we can. Also, check out our Open Collective and consider backing us.

https://opencollective.com/DockSTARTer

PS.: We offer priority support for all backers. Don't forget to add priority label when you start backing us 😄

[nemchik]

If you're taking about Plex we've got documentation about adding the device to the container https://dockstarter.com/apps/plex/#hardware-transcoding that same method should at least be similar with other containers of you're not talking about Plex. Those other containers might also have their own documentation about how to get the device in. All that is just done via overrides.

That covers adding it to specific containers as needed, but not globally to all containers. I'm curious what your reasoning is for wanting it on ALL containers?

Also about the compose versions, what example do you have that if specific to version 2.x that does not work on 3.x?

[niXta1]

If you're taking about Plex we've got documentation about adding the device to the container https://dockstarter.com/apps/plex/#hardware-transcoding that same method should at least be similar with other containers of you're not talking about Plex. Those other containers might also have their own documentation about how to get the device in. All that is just done via overrides.

That covers adding it to specific containers as needed, but not globally to all containers. I'm curious what your reasoning is for wanting it on ALL containers?

Also about the compose versions, what example do you have that if specific to version 2.x that does not work on 3.x?

Yes, this is regarding Plex.

AFAIK nvidia GPU does not work with /dev/dri. It hasn't worked with any other docker setup, why would it work here?

AFAIK nvidia-docker2 'only' (I just read, don't shoot the messenger) works with v2.3 (https://github.com/NVIDIA/nvidia-docker/wiki#do-you-support-docker-compose).

If I can't use v2.3 with dockstarter, maybe it's possible to enable in the docker conf? It's just an idea.

Thanks for the quick reply!

[nemchik]

https://github.com/NVIDIA/nvidia-docker/blob/master/README.md#quickstart according to their readme nvidia-docker2 is now deprecated. There should be an alternative way to achieve using their runtime, but I haven't dug that far into it.

[niXta1]

https://github.com/NVIDIA/nvidia-docker/blob/master/README.md#quickstart according to their readme nvidia-docker2 is now deprecated. There should be an alternative way to achieve using their runtime, but I haven't dug that far into it.

Sorry, my bad, it's not called docker-2 anymore.
There's not really a name for it (unless nvidia-docker?) as far as I can see, but what you need to install is 'nvidia-container-toolkit' to get it running. So let's call it nvidia-container-toolkit for now.

At the end it says:

Do you support Docker Compose?
Yes, use Compose format 2.3 and add runtime: nvidia to your GPU service. Docker Compose must be version 1.19.0 or higher. You can find an example here.

So, to get it working you need to:

1. Install 'nvidia-container-toolkit' according to the instructions.
2. use Compose format 2.3.
3. add 'runtime: nvidia' in the compose file.

[nemchik]

It looks like the new GPU support in docker 19.03 just needs to be added to compose for version 3.x. there's a few open issues on their GH you can follow.

docker/compose#6691
docker/compose#4657
docker/compose#6239

The first one is more promising and the last one seems less likely given the GPU support added to engine.

[niXta1]

It looks like the new GPU support in docker 19.03 just needs to be added to compose for version 3.x. there's a few open issues on their GH you can follow.

docker/compose#6691
docker/compose#4657
docker/compose#6239

The first one is more promising and the last one seems less likely given the GPU support added to engine.

As far as I can see, the newest of those is 3 months old with no reply and was closed... It doesn't look too promising?

[nemchik]

I have tried in the past to allow the compose file version to be set by the user but have not been successful. I have some ideas about how I could do it, but this requires the compose templates we use to be made a little more version agnostic than they currently are. There's not a big demand for this at the moment, and some apps require features from newer compose versions (ex: labels introduced in compose file version 3.3). It makes more sense to request older apps to update than to restrict newer apps from functionality.

[daveb92]

I can confirm i have this working, override file for Plex is as below:

version: "3.4" # this must match the version in docker-compose.yml
services:
plex:
environment:
- NVIDIA_VISIBLE_DEVICES=all
- NVIDIA_DRIVER_CAPABILITIES=compute,video,utility

I set it up a while ago, but i do know that i had to change the docker to use the nvidia runtime as default in /etc/docker/daemon.json

{
"default-runtime": "nvidia",
"runtimes": {
"nvidia": {
"path": "nvidia-container-runtime",
"runtimeArgs": []
}
}
}

[niXta1]

I can confirm i have this working, override file for Plex is as below:

version: "3.4" # this must match the version in docker-compose.yml
services:
plex:
environment:

• NVIDIA_VISIBLE_DEVICES=all
• NVIDIA_DRIVER_CAPABILITIES=compute,video,utility

I set it up a while ago, but i do know that i had to change the docker to use the nvidia runtime as default in /etc/docker/daemon.json

{
"default-runtime": "nvidia",
"runtimes": {
"nvidia": {
"path": "nvidia-container-runtime",
"runtimeArgs": []
}
}
}

Sweet! Option nr. 2 it is!

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.