[ZeXi0308]

Description

In NvFractions GPU sharing mode, a pod that requests gpu-fraction together with kai.scheduler/container.<c>.gpu-memory.portion.limit passes admission and gets scheduled, but every bind fails:

admission webhook denied the request: cannot combine nvidia.com/container. limit annotation with gpu-fraction or gpu-memory annotation

The two halves disagree:

1. The binder's nvfractions plugin translates the portion limit into nvidia.com/container.<c>.gpu-memory.limit (portion × the node's nvidia.com/gpu.memory) at bind time and keeps gpu-fraction on the pod.
2. ValidateGPUFractionRequest rejects any NvFractions limit next to gpu-fraction. It already exempts the analogous gpu-memory + translated gpu-memory.limit case (hasMemoryLimitShorthand), but had no exemption for a translated portion limit.

Since validateGpuMemoryPortionLimitAnnotation requires gpu-fraction to be present, a portion limit could never bind in NvFractions mode; the pod cycles between Scheduled and BindingError.

This PR exempts the translated limit when the pod carries a gpu-memory.portion.limit annotation, mirroring the existing shorthand exemption. The gpu-fraction.limit shorthand is covered too, because the admission mutation translates it into gpu-memory.portion.limit before validation runs.

Related Issues

Ref kai-scheduler/KAI-Scheduler#2490

Checklist

• Self-reviewed
• Added/updated tests (if needed)
• Updated documentation (if needed) — n/a
• Added a changelog fragment via make changelog (or applied the skip-changelog label). Do not edit CHANGELOG.md directly — pending fragments are folded into it at release time.

Breaking Changes

None.

Additional Notes

Tests added (TestValidateGPUFractionRequest_TranslatedLimits): gpu-fraction with/without a portion limit next to an NvFractions limit, plus the existing gpu-memory with/without gpu-memory.limit precedent as a control. The "without" cases still fail with cannot combine.

Verified against KAI-Scheduler with a local replace github.com/kai-scheduler/api => ../api: all pkg/admission/webhook/... packages (including nvfractions and gpusharing, whose "rejects nvfractions limit combined with gpu-fraction" case still passes) and pkg/binder/plugins/{nvfractions,gpusharing} pass.

Reviewer note: keeping the fix in admission rather than having the binder strip gpu-fraction seemed right, since the portion-limit validation requires gpu-fraction and the binder should not rewrite the user's request. One consequence to be aware of: a user-supplied nvidia.com/container.<c>.gpu-memory.limit next to gpu-fraction + gpu-memory.portion.limit is now accepted at create time as well, which is the same latitude the gpu-memory + gpu-memory.limit precedent already has. Happy to tighten this if you prefer.

Once this is tagged I'll open the follow-up PR in KAI-Scheduler to bump the dependency and close the issue.

🤖 Generated with Claude Code