GREP-617 Topology Binding Policy

[enoodle]

What type of PR is this?

/kind api
/kind documentation

What this PR does / why we need it:

GREP for #617 - Adding configurable policy for Topology CR creation per scheduling backend.

Which issue(s) this PR fixes:

References #617

Special notes for your reviewer:

Does this PR introduce a API change?

NONE - this is a GREP only

Additional documentation e.g., enhancement proposals, usage docs, etc.:

[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.

[xianlubird]

I could not agree more. This proposal looks very similar to the schedulerTopologyReferences section described in the topology-aware scheduling GREP: https://github.com/ai-dynamo/grove/tree/main/docs/proposals/244-topology-aware-scheduling.

I think this can greatly simplify scheduler backend implementations. In practice, topology-aware scheduling often involves multiple backend-specific resources, and managing those resources correctly can become quite complex.

That said, I think Grove should still support both modes. Some users may prefer to manage the scheduler backend topology resources explicitly, while others may prefer Grove to derive and manage them automatically from ClusterTopology. Having a per-backend policy such as AutoCreate vs ReferenceRequired sounds like a good way to make that behavior explicit.

[xianlubird]

LGTM

[yankay]

nit: existing GREP directories use the un-padded issue number as the prefix (e.g. 244-, 375-, 436-). Could we rename this to 617-scheduler-backend-topology-binding-policy/ and update the heading to GREP-617 to stay consistent? Thanks!

[enoodle]

This comes from the GREP skill in the repo: https://github.com/ai-dynamo/grove/blob/main/.agents/skills/grove-grep/SKILL.md#step-1--identify-the-grep
so it should be updated there first. currently it is the "new standard" here :)