Merge pull request #39220 from BerriAI/litellm_containers_route_model_list_creds

fix(proxy): route container create and list through model_list deployments

litellm/containers/main.py

@@ -1,7 +1,7 @@
 import asyncio
 import contextvars
 import json
-from collections.abc import Coroutine, Mapping
+from collections.abc import Callable, Coroutine, Mapping
 from functools import partial
 from typing import Final, Literal, overload
 
@@ -47,6 +47,13 @@ __all__ = [
 
 
 ##### Container Create #######################
+async def _encode_created_container_id(
+    pending: Coroutine[object, object, ContainerObject],
+    encode: Callable[[ContainerObject], ContainerObject],
+) -> ContainerObject:
+    return encode(await pending)
+
+
 @client
 async def acreate_container(
     name: str,
@@ -256,16 +263,16 @@ def create_container(
             _is_async=_is_async,
         )
 
-        # Encode container_id with provider/model metadata for routing
+        encode: Final = partial(
+            ContainerRequestUtils.encode_container_id_in_response,
+            custom_llm_provider=custom_llm_provider,
+            litellm_metadata=kwargs.get("litellm_metadata"),
+            extra_body=extra_body,
+        )
         if isinstance(container_obj, ContainerObject):
-            container_obj = ContainerRequestUtils.encode_container_id_in_response(
-                response_obj=container_obj,
-                custom_llm_provider=custom_llm_provider,
-                litellm_metadata=kwargs.get("litellm_metadata"),
-                extra_body=extra_body,
-            )
+            return encode(container_obj)
 
-        return container_obj
+        return _encode_created_container_id(pending=container_obj, encode=encode)
 
     except Exception as e:
         raise litellm.exception_type(

litellm/proxy/container_endpoints/endpoints.py

@@ -208,7 +208,7 @@ async def list_containers(
 
     # Read query parameters
     query_params: Final = dict(request.query_params)
-    data: Final[dict[str, Any]] = {"query_params": query_params}
+    data: Final[dict[str, Any]] = {"query_params": query_params, "model": query_params.get("model")}
 
     # Extract custom_llm_provider using priority chain
     custom_llm_provider: Final = (

litellm/router.py

@@ -6714,7 +6714,10 @@ class Router:
         metadata. When present, decode the ID, replace ``container_id`` with the
         upstream value, and route through ``_ageneric_api_call_with_fallbacks`` so
         deployment credentials (e.g. regional ``api_base`` for Azure) match
-        :meth:`_init_responses_api_endpoints`. Otherwise call the handler directly.
+        :meth:`_init_responses_api_endpoints`. Create/list calls carry no container ID, so
+        they route through the deployment named by ``model`` when the caller passes one,
+        falling back to the direct call when no deployment matches. Otherwise call the
+        handler directly with global provider credentials.
         """
         if custom_llm_provider and "custom_llm_provider" not in kwargs:
             kwargs["custom_llm_provider"] = custom_llm_provider
@@ -6746,6 +6749,14 @@ class Router:
                     **kwargs,
                 )
 
+        requested_model: Final = kwargs.get("model")
+        if isinstance(requested_model, str) and requested_model.strip():
+            return await self._ageneric_api_call_with_fallbacks(
+                original_function=original_function,
+                passthrough_on_no_deployment=True,
+                **kwargs,
+            )
+
         return await original_function(**kwargs)
 
     async def _init_responses_api_endpoints(

tests/router_unit_tests/test_router_endpoints.py

@@ -1324,6 +1324,98 @@ async def test_init_containers_api_endpoints_managed_id_without_model_id_applies
     assert call_kw["custom_llm_provider"] == "azure"
 
 
+@pytest.mark.asyncio
+async def test_init_containers_api_endpoints_create_with_model_uses_deployment_credentials(monkeypatch):
+    """
+    ``POST /v1/containers`` carries no container ID, so a ``model`` in the request
+    body is the only way to pick a deployment. The upstream call must receive that
+    deployment's ``api_key``/``api_base`` instead of falling back to the global
+    ``OPENAI_API_KEY`` (which may be unset on the proxy).
+    """
+    monkeypatch.delenv("OPENAI_API_KEY", raising=False)
+    router = Router(
+        model_list=[
+            {
+                "model_name": "gpt-5.4",
+                "litellm_params": {
+                    "model": "openai/gpt-5.4",
+                    "api_key": "sk-model-list-key",
+                    "api_base": "https://custom.openai.example/v1",
+                },
+            }
+        ]
+    )
+    mock_original_function = AsyncMock(return_value={"id": "cntr_test", "name": "Test Container"})
+
+    await router._init_containers_api_endpoints(
+        original_function=mock_original_function,
+        custom_llm_provider="openai",
+        name="Test Container",
+        model="gpt-5.4",
+    )
+
+    mock_original_function.assert_called_once()
+    call_kw = mock_original_function.call_args.kwargs
+    assert call_kw["api_key"] == "sk-model-list-key"
+    assert call_kw["api_base"] == "https://custom.openai.example/v1"
+    assert call_kw["model"] == "openai/gpt-5.4"
+    assert call_kw["name"] == "Test Container"
+
+
+@pytest.mark.asyncio
+async def test_init_containers_api_endpoints_create_without_model_calls_directly():
+    """
+    Without ``model`` (or with ``model=None`` as the proxy forwards it), create/list
+    must keep calling the handler directly with global provider credentials.
+    """
+    router = Router(model_list=[])
+    router._ageneric_api_call_with_fallbacks = AsyncMock()
+    mock_original_function = AsyncMock(return_value={"id": "cntr_test"})
+
+    await router._init_containers_api_endpoints(
+        original_function=mock_original_function,
+        custom_llm_provider="openai",
+        name="Test Container",
+        model=None,
+    )
+
+    router._ageneric_api_call_with_fallbacks.assert_not_called()
+    mock_original_function.assert_called_once_with(custom_llm_provider="openai", name="Test Container", model=None)
+
+
+@pytest.mark.asyncio
+async def test_init_containers_api_endpoints_create_with_unknown_model_passes_through(monkeypatch):
+    """
+    A ``model`` that names no configured deployment must not turn into a 400. The call
+    falls through to the handler with the caller's model and no injected deployment
+    credentials, matching the behaviour before model-based routing existed.
+    """
+    monkeypatch.delenv("OPENAI_API_KEY", raising=False)
+    router = Router(
+        model_list=[
+            {
+                "model_name": "gpt-5.4",
+                "litellm_params": {"model": "openai/gpt-5.4", "api_key": "sk-model-list-key"},
+            }
+        ]
+    )
+    mock_original_function = AsyncMock(return_value={"id": "cntr_test"})
+
+    await router._init_containers_api_endpoints(
+        original_function=mock_original_function,
+        custom_llm_provider="openai",
+        name="Test Container",
+        model="does-not-exist",
+    )
+
+    mock_original_function.assert_called_once()
+    call_kw = mock_original_function.call_args.kwargs
+    assert call_kw["model"] == "does-not-exist"
+    assert call_kw["name"] == "Test Container"
+    assert "api_key" not in call_kw
+    assert "api_base" not in call_kw
+
+
 def test_router_model_group_encrypted_content_affinity_callback_registration():
     from litellm.router_utils.pre_call_checks.deployment_affinity_check import (
         DeploymentAffinityCheck,

tests/test_litellm/containers/test_container_api.py

@@ -152,6 +152,42 @@ class TestContainerAPI:
             assert response.id == "cntr_async_123"
             assert response.name == "Async Test Container"
 
+    @pytest.mark.asyncio
+    async def test_acreate_container_encodes_router_model_id(self):
+        """
+        The async handler returns a coroutine, so the managed-ID encoding must run
+        after it resolves. Otherwise follow-up calls (retrieve/delete/files) lose the
+        deployment and fall back to global provider credentials.
+        """
+        upstream_response = ContainerObject(
+            id="cntr_upstream_123",
+            object="container",
+            created_at=1747857508,
+            status="running",
+            expires_after={"anchor": "last_active_at", "minutes": 20},
+            last_active_at=1747857508,
+            name="Routed Container",
+        )
+
+        async def _resolve_upstream():
+            return upstream_response
+
+        with patch.object(  # test-quality-ok: create_container exposes no client seam, only the handler
+            base_llm_http_handler,
+            "container_create_handler",
+            side_effect=lambda **kwargs: _resolve_upstream() if kwargs["_is_async"] else upstream_response,
+        ):
+            response = await acreate_container(
+                name="Routed Container",
+                custom_llm_provider="openai",
+                litellm_metadata={"model_info": {"id": "deployment-abc"}},
+            )
+
+        decoded = ResponsesAPIRequestUtils._decode_container_id(response.id)
+        assert decoded["model_id"] == "deployment-abc"
+        assert decoded["custom_llm_provider"] == "openai"
+        assert decoded["response_id"] == "cntr_upstream_123"
+
     @pytest.mark.asyncio
     async def test_alist_containers_basic(self):
         """Test basic async container listing functionality."""