[Question]: is 26.3.0 nvidia-peermem issue tracked / fixed in newer releases?

[nicdumz]

https://docs.nvidia.com/datacenter/cloud-native/gpu-operator/latest/release-notes.html#v26-3-0-known-issues documents a known gpu-operator issue (which, coincidentally, we've just hit).

• Is this fixed in newer releases? (from which releases?) Should we document this in this page :-)
• If this is not fixed is there an existing bug we can track to follow resolution?

Thanks.

cc @jgehring

[rajathagasthya]

I believe this is fixed in newer driver images by NVIDIA/gpu-driver-container@446061a

@karthikvetrivel Could you confirm if we can remove this as a known issue in the documentation?

cc @mikemckiernan

[nicdumz]

in newer driver images

Thank you! From which driver image version are things fixed?

I believe we've experienced that on v26.3 + 595.71.05

[rajathagasthya]

@nicdumz Can you try with 595.91.07 and see if you hit the same issue?

[nicdumz]

If you confirm that 446061a is in 595.91.07 then we could try yes

[rajathagasthya]

@nicdumz 446061a is in 595.91.07, and it is also in 595.71.05. So I'm curious why you're seeing it.

Note that the fix only handles the case where nvidia-peermem is still loaded when the sidecar starts. If the module gets unloaded afterwards, the sidecar has to reload it, and on a fast-path restart there are no module files to load (please see #2619).

Could you confirm you're seeing the issue in 595.71.05?