Does this work with vGPU?

[bsteinb]

We would like to use this project to configure MIG backed vGPUs on A100s in an OpenStack cloud. We are using vGPU host driver version 525.60.12. As I understand it, I have to enable SR-IOV on the A100s in order to expose the vGPUs to OpenStack / KVM using the /usr/lib/nvidia/sriov-manage included with the vGPU driver. However, once I do that, mig-parted starts acting up:

# nvidia-mig-parted export
FATA[0000] Error checking MIG capable: error getting device handle: Invalid Argument 

Is this scenario not supported or am I doing something wrong?

[cdesiniotis]

This use case is not officially supported as we do not test it. Can you provide the debug output so we can get more information:

nvidia-mig-parted -d export

[bsteinb]

I'm afraid that does not provide any additional info:

# nvidia-mig-parted -d export
FATA[0000] Error checking MIG capable: error getting device handle: Invalid Argument 

[cdesiniotis]

Okay. I have tried using mig-parted in the past for this use case and recall needing this change in one of mig-parted's dependencies to make it work: https://gitlab.com/nvidia/cloud-native/go-nvlib/-/merge_requests/13. mig-parted is still using an older version of this dependency without this change, so it is not expected to work today. I am tracking this as a feature request. cc @klueska

[bsteinb]

Thanks for the pointer. I can confirm that nvidia-mig-parted export works after bumping go-nvlib to the latest commit. I will try applying a configuration tomorrow.

[sdake]

@cdesiniotis Hi. I took a look at the PR https://gitlab.com/nvidia/cloud-native/go-nvlib/-/merge_requests/1 where you authored the PR.

In my case, I want to bind a virtual function to vfio-pci and pass through the virtual function. I don't want to use vGPU. It is unclear if this is something that would ever function. Would you clarify if you could?

Thank you,
-steve