[FEATURE] Native SDK support for reading and writing MLflow tags on Unity Catalog model versions

[asanglard]

Problem Statement

When using Unity Catalog registered models in Databricks, MLflow tags set via MlflowClient.set_model_version_tag() are stored in an undocumented internal endpoint (/api/2.0/mlflow/unity-catalog/model-versions/set-tag and .../get). The ModelVersionsAPI and RegisteredModelsAPI in the SDK (w.model_versions, w.registered_models) only expose the Unity Catalog 2.1 REST layer, which does not expose these MLflow-layer tags.

As a result, developers who want to read or write MLflow tags on UC model versions have no first-class SDK method and are forced to call the internal, undocumented endpoint directly:

# Currently required workaround — undocumented endpoint:
client.api_client.do(
    "GET",
    "/api/2.0/mlflow/unity-catalog/model-versions/get",
    query={"name": "catalog.schema.model", "version": "1"},
)
client.api_client.do(
    "POST",
    "/api/2.0/mlflow/unity-catalog/model-versions/set-tag",
    body={"name": "catalog.schema.model", "version": 1, "key": "k", "value": "v"},
)

Proposed Solution

Add get_tags(), set_tag(), and delete_tag() methods to ModelVersionsAPI (and equivalently to RegisteredModelsAPI) that wrap the MLflow Unity Catalog tag endpoints, similar to the existing set_model_version_tag / delete_model_version_tag methods already present in ModelRegistryAPI (workspace registry):

# Desired SDK interface:
w.model_versions.set_tag(full_name="catalog.schema.model", version=1, key="k", value="v")
tags = w.model_versions.get(full_name="catalog.schema.model", version=1).tags

Additional Context

• SDK version: 0.97.0
• The equivalent workspace-registry API (ml.ModelRegistryAPI) already has set_model_version_tag / delete_model_version_tag, so the pattern is established — it just hasn't been carried forward to the UC ModelVersionsAPI.
• The tags are visible in the Databricks UI on the model version page and are set by standard MLflow training code (MlflowClient.set_model_version_tag()), so reading/writing them is a common operational need (e.g. flagging a model version as archived, setting deployment metadata).
• Related undocumented endpoints: POST /api/2.0/mlflow/unity-catalog/model-versions/set-tag, GET /api/2.0/mlflow/unity-catalog/model-versions/get.