virtual gpu support

[5cat]

Tell us about your request

Supporting the virtual gpu https://github.com/awslabs/aws-virtual-gpu-device-plugin to allocate pods based on their fractional usage of a gpu

Tell us about the problem you're trying to solve. What are you trying to do, and why is it hard?

Instead of requesting a whole GPU for a single pod, we request one instance for multiple pods using virtual gpu device plugin.

Are you currently working around this issue?

Not allowing karpenter to touch GPU related pods and manually allocate each pod to a node.

Additional Context

No response

Attachments

No response

Community Note

• Please vote on this issue by adding a 👍 reaction to the original issue to help the community and maintainers prioritize this request
• Please do not leave "+1" or "me too" comments, they generate extra noise for issue followers and do not help prioritize the request
• If you are interested in working on this issue or have submitted a pull request, please leave a comment

[jonathan-innis]

Another use-case around kubernetes-sigs/karpenter#751. There's still a lot of discussion around the maintainers around the right way to attack the extended resource support issue that are discussed in some detail here: #2390.

[jonathan-innis]

How does the device plugin surface the virtual GPU resource on the node? Say, in this example, what would be the k8s.amazonaws.com/vgpu on the node?

[5cat]

Hey @jonathan-innis ,

Thank you for the info, it seems that my issue is not accurate or reflective what I have wanted, things that are covered in https://github.com/awslabs/aws-virtual-gpu-device-plugin are not what I had in mind.
What I want is support for the resources nvidia.com/mig-<slice_count>g.<memory_size>gb and various strategies mentioned in https://docs.google.com/document/d/1mdgMQ8g7WmaI_XVVRrCvHPFPOMCm5LQD5JefgAh6N8g/edit

I will close this issue and open a new one with more accurate details.