nvidia-container-cli fails to enable Vulkan in containers about libnvidia-container HOT 8 CLOSED

@stgraber we've started working on this and are hoping to enable it in the next few weeks :)
cc @guptaNswati

from libnvidia-container.

@stgraber thats the goal. Right now I am trying to find all the needed deps across different versions, so that we can make it generic. One of the libs it needs is libnvidia_GLX.so that libnvidia-container already exposes if the graphics option is used but not others. And icd_loader comes from nvidia_utils package that also needs to be added.

from libnvidia-container.

Hello!

Vulkan isn't supported in containers yet. One of the reasons is we haven't discussed with someone familiar with Vulkan.

Ideally I would be interested in having some confirmation that this is the only thing we need to do to enable Vulkan in nvidia containers.

from libnvidia-container.

Oh? That's surprising, since Vulkan is listed under nvidia-container-runtime's supported driver capabilities. Would it not make sense to expose those files now even if they're officially unsupported, since they're part of the same driver package whose other files are already exposed, and since I can confirm that their presence makes things work?

In any case, thanks for the info. I suppose I'll have to continue mounting those three files manually until support is added here.

from libnvidia-container.

It'd be great if the common files could at least get added as that's causing issues for our users who use LXD containers to run steam and other games.
(As can be seen in the linked LXD issue)

from libnvidia-container.

Excellent, thanks for working on this!

from libnvidia-container.

Getting the same error as the OP with libnvidia-container-tools 1.0.0, in EL7 on ppc64le:

$ nvidia-container-cli --version
version: 1.0.0
...
$ cat /etc/redhat-release 
Red Hat Enterprise Linux Server release 7.6 (Maipo)
$ nvidia-container-cli info
NVRM version:   418.87.00
CUDA version:   10.1
...

Looks like 'libnvidia-glvkspirv' isn't being picked up by nvidia-container-cli

$ nvidia-container-cli list | grep libnvidia-glvkspirv | wc -l
0

It might be worth examining the list of libraries in the driver documentation and cross-check those with what nvidia-container-cli list is picking up.

Upon closer inspection of nvidia-container-cli list between systems running ppc64le and x86-64 flavors of EL7, the following libraries are omitted in the ppc64le flavor, but are listed in the x86-64 flavor, despite the libraries themselves being present on the ppc64le system.

/usr/lib64/libnvidia-fbc.so
/usr/lib64/libnvidia-glvkspirv.so
/usr/lib64/libnvidia-ifr.so
/usr/lib64/libnvidia-opticalflow.so

There is also version divergence; the version in the ppc64le repo hasn't been updated since 2018?

$ nvidia-container-cli --version
version: 1.0.0
build date: 2018-09-20T20:54+0000
build revision: 881c88e2e5bb682c9bb14e68bd165cfb64563bb1
build compiler: gcc 4.8.5 20150623 (Red Hat 4.8.5-28)
build platform: ppc64le
...
$ nvidia-container-cli --version
version: 1.0.7
build date: 2020-01-21T19:05+0000
build revision: b71f87c04b8eca8a16bf60995506c35c937347d9
build compiler: gcc 4.8.5 20150623 (Red Hat 4.8.5-39)
build platform: x86_64

from libnvidia-container.

@stgraber I don't think we wrote on this thread but @guptaNswati created an image a while back:

• https://gitlab.com/nvidia/container-images/vulkan
• https://hub.docker.com/r/nvidia/vulkan

@qhaas PPC packages are outdated and haven't been updated in a while (due to the lack of an actual machine) we are targeting publishing some in the current month.

from libnvidia-container.