Skip OCI Object Storage download for Ready same-path artifacts

[op109lvb]

What this PR does

Speeds up copied model reconciliation when the copied CR uses OCI Object Storage and points to the same storage path as a model artifact that is already Ready on the same node.

For same-path object storage reuse, model-agent skips the expensive download/integrity-validation path instead of re-running file validation.

The implementation looks at this node’s Ready model entries and reuses only when the storage URI, path, schema path, storage key, parameters, and resolved destination path match exactly, and the local destination path already exists.

Why we need it

Copied model CRs can point to the same object storage path as an already Ready model on the same node. In that case, re-downloading or re-validating all files adds unnecessary reconciliation latency.

This speeds up copied model CR reconciliation while preserving safety:

• reuse only happens for OCI Object Storage models
• reuse only happens when the storage URI/path/params match exactly
• reuse only happens if the local destination path exists on that node
• nodes without a local Ready artifact continue through the normal download/validation flow
• Hugging Face continues using its existing SHA-based artifact reuse logic

How to test

go test -count=1 ./pkg/modelagent ./pkg/controller/v1beta1/basemodel
ok  	sigs.k8s.io/ome/pkg/modelagent	1.715s
ok  	sigs.k8s.io/ome/pkg/controller/v1beta1/basemodel	0.493s

Checklist

• Tests added/updated (if applicable)
• Docs updated (if applicable)
• make test passes locally

[truddy0]

do we need to handle the case where the getDestPath is empty?

[op109lvb]

No need to add.
getDestPath already falls back to modelRootDir + "/" + storageUri when storage.path is empty, and reuse still requires os.Stat(destPath) to pass.

[truddy0]

nit: maybe we can extract reusable part between cluster basemodel and basemodel.

[op109lvb]

Thanks. I kept the BaseModel and ClusterBaseModel lookup loops separate because that matches the existing repo pattern: they are different CR types with different listers and namespace/key semantics. The shared comparison logic is already extracted into sameModelStoragePath, so I’d avoid a broader refactor in this focused PR.

[pallasathena92]

we have this before f8d263f
which enable HuggingFace to leverage sha for artifact reuse.
we can use it in oci object storage as well

[op109lvb]

we have this before f8d263f which enable HuggingFace to leverage sha for artifact reuse. we can use it in oci object storage as well

For the existing HF reuse path: It works because HF has a model-level commit SHA that model-agent can fetch before download and compare with the SHA recorded in the node ConfigMap. OCI Object Storage does not have the same model-level SHA in the current implementation.

To prove an existing OCI local copy matches, the current path has to do per-object size/MD5 validation, which is the expensive work this PR is trying to skip for copied CRs.
So I kept this PR scoped to OCI same-path Ready reuse: if the storage identity and resolved local destination are exactly the same and already Ready on this node, we skip downloadModel and avoid re-running per-file validation.