diff --git a/litellm/proxy/management_endpoints/usage_endpoints/ai_usage_chat.py b/litellm/proxy/management_endpoints/usage_endpoints/ai_usage_chat.py index 7355d71f8bd..66f7516047a 100644 --- a/litellm/proxy/management_endpoints/usage_endpoints/ai_usage_chat.py +++ b/litellm/proxy/management_endpoints/usage_endpoints/ai_usage_chat.py @@ -534,26 +534,42 @@ def _caller_metadata(user_api_key_dict: "UserAPIKeyAuth") -> Mapping[str, object ) -async def resolve_usage_ai_chat_model(user_api_key_dict: "UserAPIKeyAuth", model: str | None) -> str: +async def usage_ai_chat_router_settings(user_api_key_dict: "UserAPIKeyAuth") -> Mapping[str, object] | None: + from litellm.proxy.proxy_server import llm_router, prisma_client, proxy_config, proxy_logging_obj + + if llm_router is None: + return None + return await proxy_config.get_hierarchical_router_settings( + user_api_key_dict=user_api_key_dict, prisma_client=prisma_client, proxy_logging_obj=proxy_logging_obj + ) + + +def _requested_model(model: str | None, router_settings: Mapping[str, object] | None) -> str: + from litellm.router_utils.common_utils import resolve_model_group_alias + + picked_or_default: Final = _resolve_model(model) + if router_settings is None: + return picked_or_default + return resolve_model_group_alias(router_settings.get("model_group_alias"), picked_or_default) or picked_or_default + + +async def resolve_usage_ai_chat_model( + user_api_key_dict: "UserAPIKeyAuth", model: str | None, router_settings: Mapping[str, object] | None +) -> str: from litellm.proxy.auth.auth_checks import key_model_aliases_for_auth_check from litellm.proxy.common_request_processing import ( _resolve_per_request_model_group_alias, # pyright: ignore[reportPrivateUsage] # the key and team alias rule /chat/completions applies ) from litellm.proxy.litellm_pre_call_utils import resolve_team_model_alias - from litellm.proxy.proxy_server import llm_router, prisma_client, proxy_config, proxy_logging_obj + from litellm.proxy.proxy_server import llm_router - team_aliased_model: Final = resolve_team_model_alias(_resolve_model(model), user_api_key_dict) + team_aliased_model: Final = resolve_team_model_alias(_requested_model(model, router_settings), user_api_key_dict) global_aliased_model: Final = litellm.model_alias_map.get(team_aliased_model, team_aliased_model) key_aliases: Final = key_model_aliases_for_auth_check(user_api_key_dict) key_aliased_model: Final = ( key_aliases.get(global_aliased_model, global_aliased_model) if key_aliases is not None else global_aliased_model ) - if llm_router is None: - return key_aliased_model - router_settings: Final = await proxy_config.get_hierarchical_router_settings( - user_api_key_dict=user_api_key_dict, prisma_client=prisma_client, proxy_logging_obj=proxy_logging_obj - ) - if router_settings is None: + if llm_router is None or router_settings is None: return key_aliased_model group_alias_target: Final = await _resolve_per_request_model_group_alias( requested_model=key_aliased_model, @@ -626,7 +642,9 @@ async def _acompletion( ) -async def check_usage_ai_chat_model_access(user_api_key_dict: "UserAPIKeyAuth", model: str | None) -> None: +async def check_usage_ai_chat_model_access( + user_api_key_dict: "UserAPIKeyAuth", model: str | None, router_settings: Mapping[str, object] | None +) -> None: from litellm.proxy._types import ProxyException from litellm.proxy.auth.auth_checks import ( _personal_key_owner, # pyright: ignore[reportPrivateUsage] # the same owner lookup the resolved-model check uses @@ -642,7 +660,7 @@ async def check_usage_ai_chat_model_access(user_api_key_dict: "UserAPIKeyAuth", user_api_key_cache, ) - resolved_model: Final = _resolve_model(model) + resolved_model: Final = _requested_model(model, router_settings) try: await can_key_call_resolved_model( model=resolved_model, diff --git a/litellm/proxy/management_endpoints/usage_endpoints/endpoints.py b/litellm/proxy/management_endpoints/usage_endpoints/endpoints.py index 88f2cd77dad..fba6de45547 100644 --- a/litellm/proxy/management_endpoints/usage_endpoints/endpoints.py +++ b/litellm/proxy/management_endpoints/usage_endpoints/endpoints.py @@ -54,6 +54,7 @@ async def usage_ai_chat( check_usage_ai_chat_model_access, resolve_usage_ai_chat_model, stream_usage_ai_chat, + usage_ai_chat_router_settings, ) is_admin: Final = user_api_key_has_admin_view(user_api_key_dict) @@ -63,8 +64,13 @@ async def usage_ai_chat( user_id = require_caller_user_id_for_non_admin(user_api_key_dict) messages: Final = [{"role": m.role, "content": m.content} for m in data.messages] - await check_usage_ai_chat_model_access(user_api_key_dict=user_api_key_dict, model=data.model) - routed_model: Final = await resolve_usage_ai_chat_model(user_api_key_dict=user_api_key_dict, model=data.model) + router_settings: Final = await usage_ai_chat_router_settings(user_api_key_dict) + await check_usage_ai_chat_model_access( + user_api_key_dict=user_api_key_dict, model=data.model, router_settings=router_settings + ) + routed_model: Final = await resolve_usage_ai_chat_model( + user_api_key_dict=user_api_key_dict, model=data.model, router_settings=router_settings + ) return StreamingResponse( wrap_sse_stream_with_keepalive_pings( diff --git a/tests/unit/proxy/management_endpoints/usage_endpoints/test_ai_usage_chat.py b/tests/unit/proxy/management_endpoints/usage_endpoints/test_ai_usage_chat.py index c4cd3ce1dfc..f78afb0fc21 100644 --- a/tests/unit/proxy/management_endpoints/usage_endpoints/test_ai_usage_chat.py +++ b/tests/unit/proxy/management_endpoints/usage_endpoints/test_ai_usage_chat.py @@ -1148,6 +1148,7 @@ class TestUsageAiChatModelAccess: await check_usage_ai_chat_model_access( user_api_key_dict=UserAPIKeyAuth(models=["shared-model"], api_key="sk-test"), model="team-a-secret", + router_settings=None, ) assert exc_info.value.code == "403" @@ -1174,6 +1175,7 @@ class TestUsageAiChatModelAccess: result: Final = await check_usage_ai_chat_model_access( user_api_key_dict=UserAPIKeyAuth(models=["shared-model"], api_key="sk-test"), model="shared-model", + router_settings=None, ) assert result is None @@ -1199,6 +1201,7 @@ class TestUsageAiChatModelAccess: result: Final = await check_usage_ai_chat_model_access( user_api_key_dict=UserAPIKeyAuth(models=[], api_key="sk-test"), model="team-a-secret", + router_settings=None, ) assert result is None @@ -1230,6 +1233,7 @@ class TestUsageAiChatModelAccess: api_key="sk-test", ), model="team-a-secret", + router_settings=None, ) assert exc_info.value.code == "403" @@ -1257,6 +1261,7 @@ class TestUsageAiChatModelAccess: await check_usage_ai_chat_model_access( user_api_key_dict=UserAPIKeyAuth(models=["shared-model"], api_key="sk-test"), model=None, + router_settings=None, ) assert exc_info.value.code == "403" @@ -1292,11 +1297,18 @@ class TestUsageAiChatModelAccess: personal_key: Final = UserAPIKeyAuth(models=[], user_id="u1", api_key="sk-test") if denied: with pytest.raises(ProxyException) as exc_info: - await check_usage_ai_chat_model_access(user_api_key_dict=personal_key, model=model) + await check_usage_ai_chat_model_access( + user_api_key_dict=personal_key, model=model, router_settings=None + ) assert exc_info.value.code == "403" assert "No model was picked" in exc_info.value.message else: - assert await check_usage_ai_chat_model_access(user_api_key_dict=personal_key, model=model) is None + assert ( + await check_usage_ai_chat_model_access( + user_api_key_dict=personal_key, model=model, router_settings=None + ) + is None + ) class TestUsageAiChatTeamModelRouting: @@ -1579,6 +1591,39 @@ class TestUsageAiChatTeamModelRouting: assert exc_info.value.code == "403" + @pytest.mark.asyncio + @pytest.mark.parametrize( + ("key_models", "user_models", "denied"), + [(["prod-model"], None, False), ([], ["prod-alias"], True)], + ) + async def test_group_alias_is_authorized_by_its_target_for_the_key_and_the_user( + self, key_models: list[str], user_models: list[str] | None, denied: bool + ): + from litellm.proxy._types import LiteLLM_UserTable, ProxyException, UserAPIKeyAuth + from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache + + user_cache: Final = UserApiKeyCache() + if user_models is not None: + await user_cache.async_set_cache( + key="u1", value=LiteLLM_UserTable(user_id="u1", models=user_models), model_type=LiteLLM_UserTable + ) + personal_key: Final = UserAPIKeyAuth( + user_id="u1", + user_role=LitellmUserRoles.PROXY_ADMIN, + api_key="sk-test", + models=key_models, + router_settings={"model_group_alias": {"prod-alias": "prod-model"}}, + ) + with patch( # test-quality-ok: the check reads the proxy_server.user_api_key_cache module global; no injection seam + "litellm.proxy.proxy_server.user_api_key_cache", user_cache + ): + if denied: + with pytest.raises(ProxyException) as exc_info: + await self._endpoint_body(personal_key) + assert exc_info.value.code == "403" + else: + assert b'"content": "from the aliased deployment"' in await self._endpoint_body(personal_key) + @pytest.mark.asyncio async def test_router_calls_receive_sanitized_caller_auth_for_access_group_filtering(self): from litellm.proxy._types import UserAPIKeyAuth @@ -1674,8 +1719,10 @@ class TestUsageAiChatTeamModelRouting: "litellm.proxy.proxy_server.prisma_client", None ), ): - await check_usage_ai_chat_model_access(user_api_key_dict=team_key, model=None) - routed_model: Final = await resolve_usage_ai_chat_model(user_api_key_dict=team_key, model=None) + await check_usage_ai_chat_model_access(user_api_key_dict=team_key, model=None, router_settings=None) + routed_model: Final = await resolve_usage_ai_chat_model( + user_api_key_dict=team_key, model=None, router_settings=None + ) events: Final = [ json.loads(raw.removeprefix("data: ")) async for raw in stream_usage_ai_chat(