diff --git a/litellm/proxy/auth/auth_utils.py b/litellm/proxy/auth/auth_utils.py index ecb37e67c14..644253ceac7 100644 --- a/litellm/proxy/auth/auth_utils.py +++ b/litellm/proxy/auth/auth_utils.py @@ -1432,7 +1432,7 @@ def _extract_models_from_managed_resource_id( ) _append_model_candidates( candidates=candidates, - value=get_model_id_from_unified_batch_id(unified_file_id), + value=_resolve_model_id_with_router(get_model_id_from_unified_batch_id(unified_file_id), llm_router), ) except Exception as e: verbose_proxy_logger.debug("Unable to extract model from managed file/batch ID: %s", str(e)) @@ -1442,7 +1442,10 @@ def _extract_models_from_managed_resource_id( parsed_id = parse_unified_id(resource_id) if parsed_id: - _append_model_candidates(candidates=candidates, value=parsed_id.get("model_id")) + _append_model_candidates( + candidates=candidates, + value=_resolve_model_id_with_router(parsed_id.get("model_id"), llm_router), + ) _append_model_candidates(candidates=candidates, value=parsed_id.get("target_model_names")) except Exception as e: verbose_proxy_logger.debug("Unable to extract model from unified managed resource ID: %s", str(e)) diff --git a/litellm/router.py b/litellm/router.py index 69535d7c74a..0234a8f3424 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -9560,7 +9560,12 @@ class Router: return None # Strategy 1: Check if model_id directly matches a model_name or deployment ID - if model_id in self.model_names or self.has_model_id(model_id): + if model_id in self.model_names: + return model_id + if self.has_model_id(model_id): + deployment = self.get_deployment(model_id=model_id) + if deployment is not None and deployment.model_name: + return deployment.model_name return model_id # Strategy 2: Search through router's model_list to find by litellm_params.model diff --git a/tests/router_unit_tests/test_router_helper_utils.py b/tests/router_unit_tests/test_router_helper_utils.py index a969d21a681..bcc70fae67c 100644 --- a/tests/router_unit_tests/test_router_helper_utils.py +++ b/tests/router_unit_tests/test_router_helper_utils.py @@ -2659,6 +2659,23 @@ def test_resolve_model_name_from_model_id(): result = router.resolve_model_name_from_model_id("gpt-5-mini") assert result == "gpt-5-mini" + # Test case 10: model_id is a deployment ID (hash) that differs from the + # public model_name. Regression for #32580: managed batch/file IDs embed the + # deployment model_id, and it must resolve back to the public model_name so + # team model-access checks compare against the model group, not the hash. + model_list = [ + { + "model_name": "bedrock-batch-model", + "litellm_params": { + "model": "bedrock/global.anthropic.claude-haiku-4-5-20251001-v1:0", + }, + "model_info": {"id": "8d0eaa7e6c6f54a425dfd0062cb6b0dc"}, + }, + ] + router = Router(model_list=model_list) + result = router.resolve_model_name_from_model_id("8d0eaa7e6c6f54a425dfd0062cb6b0dc") + assert result == "bedrock-batch-model" + def test_get_valid_args(): """Test get_valid_args static method returns valid Router.__init__ arguments""" diff --git a/tests/test_litellm/proxy/auth/test_auth_utils.py b/tests/test_litellm/proxy/auth/test_auth_utils.py index 9f24c662581..1610d76efb7 100644 --- a/tests/test_litellm/proxy/auth/test_auth_utils.py +++ b/tests/test_litellm/proxy/auth/test_auth_utils.py @@ -569,6 +569,103 @@ def test_get_model_from_request_resolves_video_id_model_with_router(): ) +_BATCH_DEPLOYMENT_ID = "8d0eaa7e6c6f54a425dfd0062cb6b0dc" + + +def _managed_batch_router(): + from litellm.router import Router + + return Router( + model_list=[ + { + "model_name": "bedrock-batch-model", + "litellm_params": { + "model": "bedrock/global.anthropic.claude-haiku-4-5-20251001-v1:0", + }, + "model_info": {"id": _BATCH_DEPLOYMENT_ID}, + }, + { + "model_name": "some-other-model", + "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "test-key"}, + "model_info": {"id": "a-different-deployment-id"}, + }, + ] + ) + + +def _encode_managed_id(decoded: str) -> str: + return base64.urlsafe_b64encode(decoded.encode()).decode().rstrip("=") + + +_MANAGED_BATCH_ID = _encode_managed_id( + f"litellm_proxy;model_id:{_BATCH_DEPLOYMENT_ID};llm_batch_id:provider-batch-123" +) +_MANAGED_BATCH_OUTPUT_FILE_ID = _encode_managed_id( + f"litellm_proxy;model_id:{_BATCH_DEPLOYMENT_ID};llm_batch_id:provider-batch-123;" + "llm_output_file_id:provider-file-456" +) + + +@pytest.mark.parametrize( + "route, request_data", + [ + ("/v1/batches/{batch_id}", {"batch_id": _MANAGED_BATCH_ID}), + ("/v1/batches/{batch_id}/cancel", {"batch_id": _MANAGED_BATCH_ID}), + ("/v1/files/{file_id}", {"file_id": _MANAGED_BATCH_OUTPUT_FILE_ID}), + ("/v1/files/{file_id}/content", {"file_id": _MANAGED_BATCH_OUTPUT_FILE_ID}), + ], +) +def test_get_model_from_request_resolves_batch_id_deployment_to_model_name(route, request_data): + """Regression for #32580: managed batch retrieve/cancel and managed batch output + file reads encode the deployment model_id into the resource id. The auth layer must + resolve that id back to the public model group name so model-access checks compare + against the model group, not the raw deployment id.""" + assert ( + get_model_from_request( + request_data=request_data, + route=route, + llm_router=_managed_batch_router(), + ) + == "bedrock-batch-model" + ) + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "route, request_data", + [ + ("/v1/batches/{batch_id}", {"batch_id": _MANAGED_BATCH_ID}), + ("/v1/batches/{batch_id}/cancel", {"batch_id": _MANAGED_BATCH_ID}), + ("/v1/files/{file_id}/content", {"file_id": _MANAGED_BATCH_OUTPUT_FILE_ID}), + ], +) +async def test_managed_batch_routes_pass_team_model_access_check(route, request_data): + """End-to-end regression for #32580: a team scoped to the batch model group got + ``team_model_access_denied`` on retrieve/cancel because the deployment id, not the + model group, was authorized. Fails pre-fix with the deployment id in the message.""" + from litellm.proxy._types import LiteLLM_TeamTable + from litellm.proxy.auth.auth_checks import can_team_access_model + + llm_router = _managed_batch_router() + model = get_model_from_request(request_data=request_data, route=route, llm_router=llm_router) + + assert ( + await can_team_access_model( + model=model, + team_object=LiteLLM_TeamTable(team_id="team-batch", models=["bedrock-batch-model"]), + llm_router=llm_router, + ) + is True + ) + + with pytest.raises(Exception, match="team not allowed to access model"): + await can_team_access_model( + model=model, + team_object=LiteLLM_TeamTable(team_id="team-other", models=["some-other-model"]), + llm_router=llm_router, + ) + + def test_get_model_from_request_resolves_character_id_model_with_router(): from litellm.types.videos.utils import encode_character_id_with_provider