Nvidia-docker2

[cyberwillis]

any way to use it with nvidia-docker2 alpha ?

[cyberwillis]

My problem was solved!

[eywalker]

Hi @cyberwillis, if you figured out a good way to integrate it, would you mind sharing? I'm curious how you resolved this question.

[cyberwillis]

Hi again...
Sorry the delay in get back to you.
At first my problem was that I had a composer file that wasn't building or running by nvidia-docker-compose, but the problem is not nvidia-docker-composer at all, because the nvidia-docker now has changed the way it works with introduction of nvidia-docker 2 alpha.
Now it needs to pass to the docker daemon the argument --runtime=nvidia that tell the daemon to run the specific runtime for NVidia and by this way even docker-compose wasn't able to do things. At least I don't know a way to send runtime arguments to dockerd by docker-compose.

So investigating a little here, I found that the specific runtime can be set in /etc/docker/daemon.json

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

by using this the docker-compose started to work again, because eliminates the necessity of pass this argument all the time. Probably nvidia-docker-compose will also work now ... I don't get a time to test. I do have a big chain of images to build and run to be sure about this solution.
But apparently its working... I will see at the end.

Also to complement nvidia-docker-plugin does not exist anymore.

[UPDATE] : fixed the syntax of daemon.json

[cyberwillis]

Yes I confirm, my problem is solved and I get back on track and docker-compose alone is doing everything now. I am not using nvidia-docker-compose because probably it's trying to do something else involving nvidia-docker-plugin that is not needed anymore.

[eywalker]

@cyberwillis thanks a lot for sharing this!

[anibali]

It looks like docker-compose will be gaining a "runtime" option for the 2.3 schema next release, which will enable an alternative to setting default-runtime in daemon.json. See docker/compose#5405.

[Xin-cqu]

Hi again...
Sorry the delay in get back to you.
At first my problem was that I had a composer file that wasn't building or running by nvidia-docker-compose, but the problem is not nvidia-docker-composer at all, because the nvidia-docker now has changed the way it works with introduction of nvidia-docker 2 alpha.
Now it needs to pass to the docker daemon the argument --runtime=nvidia that tell the daemon to run the specific runtime for NVidia and by this way even docker-compose wasn't able to do things. At least I don't know a way to send runtime arguments to dockerd by docker-compose.

So investigating a little here, I found that the specific runtime can be set in /etc/docker/daemon.json

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

by using this the docker-compose started to work again, because eliminates the necessity of pass this argument all the time. Probably nvidia-docker-compose will also work now ... I don't get a time to test. I do have a big chain of images to build and run to be sure about this solution.
But apparently its working... I will see at the end.

Also to complement nvidia-docker-plugin does not exist anymore.

[UPDATE] : fixed the syntax of daemon.json

That works well for me. Thanks.