Support optional subsystem ID matching in device filters - vGPU

[mattwittwer]

Summary

• add optional subsystem ID matching for device-filter

Why

Some GPUs share the same PCI vendor/device ID but differ by subsystem ID, which can cause the wrong config block to match and fail validation. This allows config authors to disambiguate those devices without breaking existing configs.

Changes

• update go-nvlib PCI discovery to read subsystem vendor/device IDs
• extend DeviceID parsing and matching logic to optionally include subsystem IDs

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[karthikvetrivel]

Hey @mattwittwer! Did you get the chance to test this on two GPUs with identical PCI device IDs? Did you get the chance to test this change on an A2/A16?

The core logic looks sound and reasonable to me. Do you think you'd also be able to add some tests? Namely, to DeviceID construction, NewDeviceIDFromString, and most importantly, Matches()?

[mattwittwer]

Hey @mattwittwer! Did you get the chance to test this on two GPUs with identical PCI device IDs? Did you get the chance to test this change on an A2/A16?

The core logic looks sound and reasonable to me. Do you think you'd also be able to add some tests? Namely, to DeviceID construction, NewDeviceIDFromString, and most importantly, Matches()?

Hi @karthikvetrivel! I have not been able to try this out directly with A2 and A16 GPUs on the same cluster. The user who reported this issue was able to return their nodes to production by manually setting the labels.

[karthikvetrivel]

As @rajathagasthya said on the mig-parted PR, this change needs to be made in https://github.com/NVIDIA/go-nvlib and vendored here with a go mod dependency update.