Merge commit '957af3023594810253a0abf06cf51348bfa43a4c' into litellm_team_byok_models_fix_uuid_v2

litellm/proxy/proxy_server.py

@@ -8343,6 +8343,12 @@ async def model_list(
         if hidden_names:
             all_models = [m for m in all_models if m not in hidden_names]
 
+        # Opt-in (default off): surface the public name for team-scoped rows.
+        # Off by default so /v1/models model ids stay backward-compatible for
+        # callers that scripted against the internal routing name.
+        if general_settings.get("use_team_public_model_name", False):
+            all_models = _translate_model_names_for_listing(all_models, llm_router)
+
         # Build response data with all proxy models
         model_data = []
         for model in all_models:
@@ -8380,6 +8386,12 @@ async def model_list(
     if hidden_names:
         all_models = [m for m in all_models if m not in hidden_names]
 
+    # Opt-in (default off): surface the public name for team-scoped rows.
+    # Off by default so /v1/models model ids stay backward-compatible for
+    # callers that scripted against the internal routing name.
+    if general_settings.get("use_team_public_model_name", False):
+        all_models = _translate_model_names_for_listing(all_models, llm_router)
+
     # Build response data
     model_data = []
     for model in all_models:
@@ -12589,6 +12601,34 @@ def _translate_model_name_for_response(model: dict) -> dict:
     return {**model, "model_name": team_public}
 
 
+def _translate_model_names_for_listing(model_names: List[str], llm_router) -> List[str]:
+    """Swap internal team routing keys for their public names in list-style
+    responses (e.g. `/v1/models`, `/models`).
+
+    `/v1/models` builds from bare model-name strings produced by access-group
+    expansion (`get_model_access_groups`), which surfaces the internal routing
+    key `model_name_{team_id}_{uuid}` for team-scoped (BYOK) deployments. This
+    is a presentation-layer swap only -- access-group/auth semantics are
+    unchanged (see issue #28382). Sibling deployments collapse to one public
+    name, so the result is de-duplicated while preserving order.
+    """
+    if llm_router is None:
+        return model_names
+    internal_to_public: Dict[str, str] = {}
+    for m in getattr(llm_router, "model_list", None) or []:
+        model_info = m.get("model_info") or {}
+        if not isinstance(model_info, dict):
+            continue
+        team_id = model_info.get("team_id")
+        team_public = model_info.get("team_public_model_name")
+        name = m.get("model_name") or ""
+        if team_id and team_public and name.startswith(f"model_name_{team_id}_"):
+            internal_to_public[name] = team_public
+    if not internal_to_public:
+        return model_names
+    return list(dict.fromkeys(internal_to_public.get(n, n) for n in model_names))
+
+
 def _get_proxy_model_info(model: dict) -> dict:
     # provided model_info in config.yaml
     model_info = model.get("model_info", {})

tests/test_litellm/proxy/proxy_server/test_team_model_name_translation.py

@@ -593,3 +593,132 @@ async def test_model_info_v1_litellm_model_id_team_id_applies_team_filter(monkey
     team_filter.assert_awaited_once()
     assert team_filter.await_args.kwargs["team_id"] == "other-team"
     assert team_filter.await_args.kwargs["all_models"] == [team_row]
+
+
+@pytest.mark.asyncio
+async def test_v1_models_translates_team_model_for_access_group_key(monkeypatch):
+    """Regression (#28382 sibling leak): a virtual key whose model access group
+    resolves to a team BYOK deployment must list the PUBLIC name in /v1/models,
+    not the internal routing key model_name_{team_id}_{uuid}.
+
+    The /model/info read-path fix did not cover /v1/models, which builds from
+    bare model-name strings via access-group expansion.
+    """
+    team_dep = {
+        "model_name": "model_name_teamX_uuid9",
+        "litellm_params": {"model": "azure/gpt-4.1"},
+        "model_info": {
+            "id": "id1",
+            "team_id": "teamX",
+            "team_public_model_name": "tushar-gpt-4.1",
+            "access_groups": ["grp-a"],
+        },
+    }
+    router = MagicMock()
+    router.get_model_names.return_value = ["model_name_teamX_uuid9"]
+    router.get_model_access_groups.return_value = {"grp-a": ["model_name_teamX_uuid9"]}
+    router.get_fully_blocked_model_names.return_value = set()
+    router.model_list = [team_dep]
+    router.get_model_list.return_value = [team_dep]
+
+    monkeypatch.setattr(ps, "llm_router", router)
+    monkeypatch.setattr(ps, "user_model", None)
+    # opt-in flag ON -> listing surfaces public names
+    monkeypatch.setattr(ps, "general_settings", {"use_team_public_model_name": True})
+
+    # virtual key granted access via the access group (no team membership)
+    key = UserAPIKeyAuth(
+        user_id="u", api_key="sk-test", models=["grp-a"], team_models=[]
+    )
+    resp = await ps.model_list(user_api_key_dict=key)
+
+    ids = [d["id"] for d in resp["data"]]
+    assert "tushar-gpt-4.1" in ids
+    assert "model_name_teamX_uuid9" not in ids
+
+
+@pytest.mark.asyncio
+async def test_v1_models_keeps_internal_names_when_flag_off(monkeypatch):
+    """Default (flag off): /v1/models stays backward-compatible and lists the
+    internal routing name, so consumers that scripted against those ids are not
+    broken. Translation is opt-in via
+    general_settings['use_team_public_model_name'].
+    """
+    team_dep = {
+        "model_name": "model_name_teamX_uuid9",
+        "litellm_params": {"model": "azure/gpt-4.1"},
+        "model_info": {
+            "id": "id1",
+            "team_id": "teamX",
+            "team_public_model_name": "tushar-gpt-4.1",
+            "access_groups": ["grp-a"],
+        },
+    }
+    router = MagicMock()
+    router.get_model_names.return_value = ["model_name_teamX_uuid9"]
+    router.get_model_access_groups.return_value = {"grp-a": ["model_name_teamX_uuid9"]}
+    router.get_fully_blocked_model_names.return_value = set()
+    router.model_list = [team_dep]
+    router.get_model_list.return_value = [team_dep]
+
+    monkeypatch.setattr(ps, "llm_router", router)
+    monkeypatch.setattr(ps, "user_model", None)
+    monkeypatch.setattr(ps, "general_settings", {})  # flag absent -> default off
+
+    key = UserAPIKeyAuth(
+        user_id="u", api_key="sk-test", models=["grp-a"], team_models=[]
+    )
+    resp = await ps.model_list(user_api_key_dict=key)
+
+    ids = [d["id"] for d in resp["data"]]
+    assert "model_name_teamX_uuid9" in ids  # internal id preserved (backward-compat)
+    assert "tushar-gpt-4.1" not in ids
+
+
+def test_translate_model_names_for_listing_swaps_and_dedupes():
+    """Internal team routing keys -> public name; sibling deployments sharing a
+    public name collapse to one entry (order preserved); globals untouched."""
+    from litellm.proxy.proxy_server import _translate_model_names_for_listing
+
+    router = MagicMock()
+    router.model_list = [
+        {
+            "model_name": "model_name_teamX_uuidA",
+            "model_info": {
+                "team_id": "teamX",
+                "team_public_model_name": "tushar-gpt-4.1",
+            },
+        },
+        {
+            "model_name": "model_name_teamX_uuidB",  # sibling: same public name
+            "model_info": {
+                "team_id": "teamX",
+                "team_public_model_name": "tushar-gpt-4.1",
+            },
+        },
+        {"model_name": "gpt-4o", "model_info": {"db_model": False}},
+    ]
+
+    out = _translate_model_names_for_listing(
+        ["model_name_teamX_uuidA", "model_name_teamX_uuidB", "gpt-4o"], router
+    )
+    assert out == ["tushar-gpt-4.1", "gpt-4o"]
+
+
+def test_translate_model_names_for_listing_leaves_unmapped_names():
+    """Names with no team mapping (globals, access-group keys) pass through."""
+    from litellm.proxy.proxy_server import _translate_model_names_for_listing
+
+    router = MagicMock()
+    router.model_list = [{"model_name": "gpt-4o", "model_info": {"db_model": False}}]
+    assert _translate_model_names_for_listing(["gpt-4o", "beta-group"], router) == [
+        "gpt-4o",
+        "beta-group",
+    ]
+
+
+def test_translate_model_names_for_listing_none_router():
+    """No router -> return the input list unchanged."""
+    from litellm.proxy.proxy_server import _translate_model_names_for_listing
+
+    assert _translate_model_names_for_listing(["a", "b"], None) == ["a", "b"]