From 7015bf37bb3830dab79b3efd886eee9b5b7ffad6 Mon Sep 17 00:00:00 2001 From: ryan-crabbe-berri Date: Sat, 5 Sep 2026 17:32:17 -0700 Subject: [PATCH 1/8] fix(proxy): apply team model aliases on the JWT auth path Team reads on the auth path never loaded the team's alias table, and the team-based JWT branch copied a hand-picked subset of team fields onto UserAPIKeyAuth, so aliases (and a few other team grants) never reached JWT callers: restricted teams 403'd alias requests and open teams 400'd them Load the alias relation where the team row is read and cached, project the team onto every team_* token field through one shared team_grants helper used by both JWT returns, and keep the relation when team model add/delete rewrites the cached team. ui_sso reuses the shared alias table model Resolves LIT-5858 Claude-Session: https://claude.ai/code/session_01EX13mWex6RaBo9PYnkAtFW --- basedpyright-code-budget.json | 2 +- litellm/proxy/auth/auth_checks.py | 9 +- litellm/proxy/auth/handle_jwt.py | 5 +- litellm/proxy/auth/team_grants.py | 122 +++++++++++++++++ litellm/proxy/auth/user_api_key_auth.py | 22 +-- .../management_endpoints/team_endpoints.py | 4 +- litellm/proxy/management_endpoints/ui_sso.py | 23 +--- .../proxy/auth/test_auth_checks.py | 64 +++++++++ .../proxy/auth/test_handle_jwt.py | 52 +++++++ .../proxy/auth/test_team_grants.py | 129 ++++++++++++++++++ .../proxy/auth/test_user_api_key_auth.py | 116 ++++++++++++++++ .../test_team_endpoints.py | 44 ++++++ type-discipline-budget.json | 2 +- 13 files changed, 547 insertions(+), 47 deletions(-) create mode 100644 litellm/proxy/auth/team_grants.py create mode 100644 tests/test_litellm/proxy/auth/test_team_grants.py diff --git a/basedpyright-code-budget.json b/basedpyright-code-budget.json index 0b0a61192e6..8bdc251c684 100644 --- a/basedpyright-code-budget.json +++ b/basedpyright-code-budget.json @@ -105,7 +105,7 @@ "limit": 109 }, "reportUnknownMemberType": { - "limit": 38271 + "limit": 38269 }, "reportUnknownParameterType": { "limit": 19584 diff --git a/litellm/proxy/auth/auth_checks.py b/litellm/proxy/auth/auth_checks.py index dc693317de0..a71a1993064 100644 --- a/litellm/proxy/auth/auth_checks.py +++ b/litellm/proxy/auth/auth_checks.py @@ -475,6 +475,7 @@ def _is_model_cost_zero(model: str | list[str] | None, llm_router: Router | None _NO_MODEL_INFO: Final[Mapping[str, object]] = MappingProxyType({}) +_TEAM_GRANT_RELATIONS: Final[Mapping[str, object]] = MappingProxyType({"litellm_model_table": True}) def _has_ptu_flat_cost(model: str, llm_router: "Router") -> bool: @@ -2858,7 +2859,9 @@ class TeamNotFoundError(HTTPException): async def _get_team_db_check( team_id: str, prisma_client: PrismaClient, team_id_upsert: bool | None = None ) -> "_PrismaTeamRow | None": - response = await _team_table(TeamRepository(prisma_client)).find_unique(where={"team_id": team_id}) + response = await _team_table(TeamRepository(prisma_client)).find_unique( + where={"team_id": team_id}, include=_TEAM_GRANT_RELATIONS + ) if response is None and team_id_upsert: from litellm.proxy.management_endpoints.team_endpoints import new_team @@ -3158,7 +3161,9 @@ async def get_team_object_by_alias( # Query database by team_alias try: - teams: Final = await _team_table(TeamRepository(prisma_client)).find_many(where={"team_alias": team_alias}) + teams: Final = await _team_table(TeamRepository(prisma_client)).find_many( + where={"team_alias": team_alias}, include=_TEAM_GRANT_RELATIONS + ) if not teams: raise HTTPException( diff --git a/litellm/proxy/auth/handle_jwt.py b/litellm/proxy/auth/handle_jwt.py index 0795cee7409..69091ee8344 100644 --- a/litellm/proxy/auth/handle_jwt.py +++ b/litellm/proxy/auth/handle_jwt.py @@ -53,6 +53,7 @@ from litellm.proxy._types import ( ) from litellm.proxy.auth.auth_checks import can_team_access_model from litellm.proxy.auth.route_checks import RouteChecks +from litellm.proxy.auth.team_grants import team_model_aliases from litellm.proxy.common_utils.user_api_key_cache import ( UserApiKeyCache, get_management_object_ttl, @@ -1595,7 +1596,7 @@ class JWTAuthManager: model=requested_model, team_object=team_object, llm_router=llm_router, - team_model_aliases=None, + team_model_aliases=team_model_aliases(team_object), ) ): is_allowed = allowed_routes_check( @@ -2132,7 +2133,7 @@ class JWTAuthManager: model=requested_model, team_object=team_object, llm_router=llm_router, - team_model_aliases=None, + team_model_aliases=team_model_aliases(team_object), ) except ProxyException: continue diff --git a/litellm/proxy/auth/team_grants.py b/litellm/proxy/auth/team_grants.py new file mode 100644 index 00000000000..1196011dcdd --- /dev/null +++ b/litellm/proxy/auth/team_grants.py @@ -0,0 +1,122 @@ +"""Project a team row (plus the caller's membership in it) onto the ``team_*`` fields of ``UserAPIKeyAuth``. + +The virtual-key path gets these fields for free from the combined-view SQL join. Every other auth path +starts from a ``LiteLLM_TeamTable`` object instead and has to copy them over by hand, which is how JWT +callers kept losing grants (aliases, permissions, limits) one field at a time. Build the badge through +``team_grants`` and the two paths cannot drift. +""" + +from collections.abc import Mapping, Sequence +from types import MappingProxyType +from typing import Annotated, Final + +from pydantic import BaseModel, BeforeValidator, ConfigDict, TypeAdapter, ValidationError +from pydantic.main import IncEx +from typing_extensions import ReadOnly, TypedDict + +from litellm.proxy._types import ( + LiteLLM_ObjectPermissionTable, + LiteLLM_TeamMembership, + LiteLLM_TeamTable, + Member, +) + +_MODEL_ALIASES_ADAPTER: Final = TypeAdapter(dict[str, str]) +_JSON_COLUMNS: Final[Mapping[str, IncEx | bool]] = MappingProxyType( + {"metadata": True, "litellm_model_table": MappingProxyType({"model_aliases": True})} +) + + +def _decode_model_aliases(value: object) -> object: + """``LiteLLM_ModelTable.model_aliases`` is typed ``str | dict``; writers hand Prisma ``json.dumps(...)``, so take both.""" + if not isinstance(value, str): + return value + try: + return _MODEL_ALIASES_ADAPTER.validate_json(value) + except ValidationError: + return None + + +class TeamModelAliasTable(BaseModel): + model_config = ConfigDict(protected_namespaces=()) + + model_aliases: Annotated[Mapping[str, str] | None, BeforeValidator(_decode_model_aliases)] = None + + +class _TeamJsonColumns(BaseModel): + """The two loosely typed columns on ``LiteLLM_TeamTable``, re-read with the shape the badge needs.""" + + metadata: Mapping[str, object] | None = None + litellm_model_table: TeamModelAliasTable | None = None + + +class TeamGrants(TypedDict, total=False): + """Keyword arguments for ``UserAPIKeyAuth``. Empty when the caller has no team, so the model's own defaults apply.""" + + team_alias: ReadOnly[str | None] + team_tpm_limit: ReadOnly[int | None] + team_rpm_limit: ReadOnly[int | None] + team_max_budget: ReadOnly[float | None] + team_soft_budget: ReadOnly[float | None] + team_spend: ReadOnly[float | None] + team_models: ReadOnly[Sequence[str]] + team_blocked: ReadOnly[bool] + team_metadata: ReadOnly[Mapping[str, object] | None] + team_model_aliases: ReadOnly[Mapping[str, str] | None] + team_object_permission_id: ReadOnly[str | None] + team_object_permission: ReadOnly[LiteLLM_ObjectPermissionTable | None] + team_member: ReadOnly[Member | None] + team_member_spend: ReadOnly[float | None] + team_member_tpm_limit: ReadOnly[int | None] + team_member_rpm_limit: ReadOnly[int | None] + + +def _json_columns(team_object: LiteLLM_TeamTable) -> _TeamJsonColumns: + try: + return _TeamJsonColumns.model_validate(team_object.model_dump(include=_JSON_COLUMNS)) + except ValidationError: + return _TeamJsonColumns() + + +def team_model_aliases(team_object: LiteLLM_TeamTable | None) -> Mapping[str, str] | None: + if team_object is None: + return None + alias_table: Final = _json_columns(team_object).litellm_model_table + return alias_table.model_aliases if alias_table is not None else None + + +def team_grants( + team_object: LiteLLM_TeamTable | None, + team_membership: LiteLLM_TeamMembership | None, + user_id: str | None, +) -> TeamGrants: + if team_object is None: + return TeamGrants() + json_columns: Final = _json_columns(team_object) + return TeamGrants( + team_alias=team_object.team_alias, + team_tpm_limit=team_object.tpm_limit, + team_rpm_limit=team_object.rpm_limit, + team_max_budget=team_object.max_budget, + team_soft_budget=team_object.soft_budget, + team_spend=team_object.spend, + team_models=tuple(team_object.models), + team_blocked=team_object.blocked, + team_metadata=json_columns.metadata, + team_model_aliases=( + json_columns.litellm_model_table.model_aliases if json_columns.litellm_model_table is not None else None + ), + team_object_permission_id=team_object.object_permission_id, + team_object_permission=team_object.object_permission, + team_member=next( + (m for m in team_object.members_with_roles if user_id is not None and m.user_id == user_id), + None, + ), + team_member_spend=team_membership.spend if team_membership is not None else None, + team_member_tpm_limit=( + team_membership.safe_get_team_member_tpm_limit() if team_membership is not None else None + ), + team_member_rpm_limit=( + team_membership.safe_get_team_member_rpm_limit() if team_membership is not None else None + ), + ) diff --git a/litellm/proxy/auth/user_api_key_auth.py b/litellm/proxy/auth/user_api_key_auth.py index 93293db24c6..08dbe2508ec 100644 --- a/litellm/proxy/auth/user_api_key_auth.py +++ b/litellm/proxy/auth/user_api_key_auth.py @@ -82,6 +82,7 @@ from litellm.proxy.auth.oauth2_proxy_hook import handle_oauth2_proxy_request from litellm.proxy.auth.resolvers import CredentialRef, Principal from litellm.proxy.auth.resolvers.store import IdentityStore from litellm.proxy.auth.route_checks import RouteChecks +from litellm.proxy.auth.team_grants import team_grants from litellm.proxy.auth.trusted_proxy_utils import get_trusted_proxy_cidrs from litellm.proxy.common_utils.cache_coordinator import EventDrivenCacheCoordinator from litellm.proxy.common_utils.http_parsing_utils import ( @@ -1476,24 +1477,16 @@ async def _user_api_key_auth_builder( user_id=user_id, user_email=user_email, team_id=team_id, - team_alias=(team_object.team_alias if team_object is not None else None), - team_tpm_limit=(team_object.tpm_limit if team_object is not None else None), - team_rpm_limit=(team_object.rpm_limit if team_object is not None else None), - team_models=(team_object.models if team_object is not None else []), - team_metadata=(team_object.metadata if team_object is not None else None), org_id=org_id, end_user_id=end_user_id, parent_otel_span=parent_otel_span, jwt_claims=jwt_claims, + **team_grants(team_object=team_object, team_membership=team_membership, user_id=user_id), ) valid_token = UserAPIKeyAuth( api_key=None, team_id=team_id, - team_alias=(team_object.team_alias if team_object is not None else None), - team_tpm_limit=(team_object.tpm_limit if team_object is not None else None), - team_rpm_limit=(team_object.rpm_limit if team_object is not None else None), - team_models=(team_object.models if team_object is not None else []), user_role=( LitellmUserRoles(user_object.user_role) if user_object is not None and user_object.user_role is not None @@ -1507,17 +1500,8 @@ async def _user_api_key_auth_builder( user_tpm_limit=(user_object.tpm_limit if user_object is not None else None), user_rpm_limit=(user_object.rpm_limit if user_object is not None else None), user_model_max_budget=(user_object.model_max_budget if user_object is not None else None), - team_member_rpm_limit=( - team_membership.safe_get_team_member_rpm_limit() if team_membership is not None else None - ), - team_member_tpm_limit=( - team_membership.safe_get_team_member_tpm_limit() if team_membership is not None else None - ), - team_metadata=(team_object.metadata if team_object is not None else None), jwt_claims=jwt_claims, - ) - valid_token.team_object_permission = ( - team_object.object_permission if team_object is not None else None + **team_grants(team_object=team_object, team_membership=team_membership, user_id=user_id), ) # AUTO_REGISTER deferred from _resolve_jwt_to_virtual_key. diff --git a/litellm/proxy/management_endpoints/team_endpoints.py b/litellm/proxy/management_endpoints/team_endpoints.py index 2ec68a10f65..c050368b3fe 100644 --- a/litellm/proxy/management_endpoints/team_endpoints.py +++ b/litellm/proxy/management_endpoints/team_endpoints.py @@ -5601,7 +5601,7 @@ async def team_model_add( updated_team: Final = await _team_db(prisma_client).update( where={"team_id": data.team_id}, data={"updated_at": datetime.now(timezone.utc)}, - include={"object_permission": True}, + include={"litellm_model_table": True, "object_permission": True}, ) if updated_team is None: raise HTTPException( @@ -5688,7 +5688,7 @@ async def team_model_delete( updated_team: Final = await _team_db(prisma_client).update( where={"team_id": data.team_id}, data={"models": updated_models}, - include={"object_permission": True}, + include={"litellm_model_table": True, "object_permission": True}, ) if updated_team is None: raise HTTPException( diff --git a/litellm/proxy/management_endpoints/ui_sso.py b/litellm/proxy/management_endpoints/ui_sso.py index 3e6434a5afd..c60888e298f 100644 --- a/litellm/proxy/management_endpoints/ui_sso.py +++ b/litellm/proxy/management_endpoints/ui_sso.py @@ -22,7 +22,6 @@ from html import escape from types import MappingProxyType from typing import ( TYPE_CHECKING, - Annotated, Any, Final, Literal, @@ -42,7 +41,7 @@ if TYPE_CHECKING: import jwt from fastapi import APIRouter, Depends, Header, HTTPException, Request, Response, status from fastapi.responses import RedirectResponse -from pydantic import BaseModel, BeforeValidator, ConfigDict, TypeAdapter, ValidationError +from pydantic import BaseModel, TypeAdapter, ValidationError import litellm from litellm._logging import verbose_proxy_logger @@ -95,6 +94,7 @@ from litellm.proxy.auth.auth_utils import ( ) from litellm.proxy.auth.handle_jwt import JWTHandler from litellm.proxy.auth.ip_address_utils import IPAddressUtils +from litellm.proxy.auth.team_grants import TeamModelAliasTable from litellm.proxy.auth.user_api_key_auth import user_api_key_auth from litellm.proxy.common_utils.admin_ui_utils import ( admin_ui_disabled, @@ -209,31 +209,14 @@ def _team_detail_db(repo: TeamRepository) -> "TableActions[_TeamDetailRow]": return repo.table -_MODEL_ALIASES_ADAPTER: Final = TypeAdapter(dict[str, str]) _SSO_TOKEN_CLAIMS_ADAPTER: Final = TypeAdapter(Mapping[str, object]) -def _decode_model_aliases(value: object) -> object: - """``/team/new`` stores team model aliases as a JSON-encoded string in the Json column.""" - if not isinstance(value, str): - return value - try: - return _MODEL_ALIASES_ADAPTER.validate_json(value) - except ValidationError: - return None - - -class _TeamModelAliasTable(BaseModel): - model_config = ConfigDict(protected_namespaces=()) - - model_aliases: Annotated[Mapping[str, str] | None, BeforeValidator(_decode_model_aliases)] = None - - class _TeamRowGrants(BaseModel): team_id: str team_alias: str | None = None models: tuple[str, ...] = () - litellm_model_table: _TeamModelAliasTable | None = None + litellm_model_table: TeamModelAliasTable | None = None class CliSsoTeamDetail(BaseModel): diff --git a/tests/test_litellm/proxy/auth/test_auth_checks.py b/tests/test_litellm/proxy/auth/test_auth_checks.py index 2284a05b2e9..5bfef2b6445 100644 --- a/tests/test_litellm/proxy/auth/test_auth_checks.py +++ b/tests/test_litellm/proxy/auth/test_auth_checks.py @@ -2374,6 +2374,44 @@ def _mock_prisma_for_team_lookup(find_unique): return mock_prisma_client +_TEAM_ALIAS_TABLE_ROW = {"id": 1, "model_aliases": '{"fast": "gpt-4o"}', "created_by": "admin", "updated_by": "admin"} + + +def _prisma_team_row(include): + """Mimics Prisma: the `litellm_model_table` relation rides on the row only when the query `include`s it.""" + columns = {"team_id": "team-aliases", "team_alias": "aliases", "models": ["gpt-4o"]} + row = ( + {**columns, "litellm_model_table": _TEAM_ALIAS_TABLE_ROW} + if (include or {}).get("litellm_model_table") + else columns + ) + return SimpleNamespace(dict=lambda: row, model_dump=lambda: row) + + +@pytest.mark.asyncio +async def test_get_team_object_loads_model_aliases_relation(): + """LIT-5858: the auth path read teams without `include`ing `litellm_model_table`, so every JWT + team came back with `model_aliases=None` and alias requests 403'd.""" + from litellm.proxy.auth.auth_checks import get_team_object + from litellm.proxy.auth.team_grants import team_model_aliases + + async def find_unique(where, include=None): + return _prisma_team_row(include) + + mock_cache = MagicMock() + mock_cache.async_get_cache = AsyncMock(return_value=None) + mock_cache.async_set_cache = AsyncMock() + + team = await get_team_object( + team_id="team-aliases", + prisma_client=_mock_prisma_for_team_lookup(AsyncMock(side_effect=find_unique)), + user_api_key_cache=mock_cache, + check_db_only=True, + ) + + assert team_model_aliases(team) == {"fast": "gpt-4o"} + + @pytest.mark.asyncio async def test_get_team_object_distinguishes_absent_team_from_unreadable_row(): """A deleted team and a database that would not answer both surface as a 404, @@ -6195,6 +6233,32 @@ async def test_get_team_object_by_alias_db_fetch_returns_cached_obj(): assert result.models == ["gpt-4"] +@pytest.mark.asyncio +async def test_get_team_object_by_alias_loads_model_aliases_relation(): + """LIT-5858: same regression as `test_get_team_object_loads_model_aliases_relation`, for the + `team_alias_jwt_field` lookup.""" + from litellm.proxy.auth.auth_checks import get_team_object_by_alias + from litellm.proxy.auth.team_grants import team_model_aliases + + async def find_many(where, include=None): + return [_prisma_team_row(include)] + + mock_prisma_client = MagicMock() + mock_prisma_client.db.litellm_teamtable.find_many = AsyncMock(side_effect=find_many) + + mock_cache = MagicMock() + mock_cache.async_get_cache = AsyncMock(return_value=None) + mock_cache.async_set_cache = AsyncMock() + + team = await get_team_object_by_alias( + team_alias="aliases", + prisma_client=mock_prisma_client, + user_api_key_cache=mock_cache, + ) + + assert team_model_aliases(team) == {"fast": "gpt-4o"} + + @pytest.mark.asyncio async def test_get_org_object_by_alias_db_fetch_returns_validated_org(): from litellm.proxy._types import LiteLLM_OrganizationTable diff --git a/tests/test_litellm/proxy/auth/test_handle_jwt.py b/tests/test_litellm/proxy/auth/test_handle_jwt.py index 99a0a4c0a8b..94226b5404d 100644 --- a/tests/test_litellm/proxy/auth/test_handle_jwt.py +++ b/tests/test_litellm/proxy/auth/test_handle_jwt.py @@ -13,6 +13,7 @@ from litellm.proxy._types import ( DEFAULT_JWKS_STALE_TTL, JWTLiteLLMRoleMap, LiteLLM_JWTAuth, + LiteLLM_ModelTable, LiteLLM_TeamMembership, LiteLLM_TeamTable, LiteLLM_UserTable, @@ -1255,6 +1256,57 @@ async def test_find_team_with_model_access_model_group(monkeypatch): assert team_obj.team_id == "team-1" +@pytest.mark.asyncio +@pytest.mark.parametrize( + "model_aliases", + ['{"fast": "gpt-4o"}', {"fast": "gpt-4o"}], + ids=["json-string", "dict"], +) +async def test_find_team_with_model_access_resolves_team_model_alias(monkeypatch, model_aliases): + """LIT-5858: a JWT team that grants `gpt-4o` under the alias `fast` must resolve a request + for `fast`. The JWT path used to pass `team_model_aliases=None`, so every alias request 403'd.""" + import sys + import types + + from litellm.caching import DualCache + from litellm.proxy.utils import ProxyLogging + from litellm.router import Router + + router = Router(model_list=[{"model_name": "gpt-4o", "litellm_params": {"model": "gpt-4o"}}]) + proxy_server_module = types.ModuleType("proxy_server") + proxy_server_module.llm_router = router + monkeypatch.setitem(sys.modules, "litellm.proxy.proxy_server", proxy_server_module) + + team = LiteLLM_TeamTable( + team_id="team-aliases", + models=["gpt-4o"], + litellm_model_table=LiteLLM_ModelTable(model_aliases=model_aliases, created_by="admin", updated_by="admin"), + ) + + async def mock_get_team_object(*args, **kwargs): + return team + + monkeypatch.setattr("litellm.proxy.auth.handle_jwt.get_team_object", mock_get_team_object) + + jwt_handler = JWTHandler() + jwt_handler.litellm_jwtauth = LiteLLM_JWTAuth() + user_api_key_cache = DualCache() + + team_id, team_obj = await JWTAuthManager.find_team_with_model_access( + team_ids={"team-aliases"}, + requested_model="fast", + route="/chat/completions", + jwt_handler=jwt_handler, + prisma_client=None, + user_api_key_cache=user_api_key_cache, + parent_otel_span=None, + proxy_logging_obj=ProxyLogging(user_api_key_cache=user_api_key_cache), + ) + + assert team_id == "team-aliases" + assert team_obj is team + + @pytest.mark.asyncio async def test_find_team_with_model_access_v1_messages_default_routes(monkeypatch): """Regression for #31189: a single-team JWT that grants the requested model diff --git a/tests/test_litellm/proxy/auth/test_team_grants.py b/tests/test_litellm/proxy/auth/test_team_grants.py new file mode 100644 index 00000000000..447fc1c93a1 --- /dev/null +++ b/tests/test_litellm/proxy/auth/test_team_grants.py @@ -0,0 +1,129 @@ +import pytest + +from litellm.proxy._types import ( + LiteLLM_BudgetTable, + LiteLLM_ObjectPermissionTable, + LiteLLM_TeamMembership, + LiteLLM_TeamTable, + LiteLLM_VerificationTokenView, + Member, + UserAPIKeyAuth, +) +from litellm.models.team import LiteLLM_ModelTable +from litellm.proxy.auth.team_grants import team_grants, team_model_aliases + +TEAM_ID = "team-grants" +USER_ID = "user-in-team" +ALIASES = {"fast": "gpt-4o-mini", "smart": "gpt-4o"} + + +def _alias_table(model_aliases) -> LiteLLM_ModelTable: + return LiteLLM_ModelTable(model_aliases=model_aliases, created_by="admin", updated_by="admin") + + +def _full_team(model_aliases=ALIASES) -> LiteLLM_TeamTable: + return LiteLLM_TeamTable( + team_id=TEAM_ID, + team_alias="grants-team", + tpm_limit=1000, + rpm_limit=10, + max_budget=50.0, + soft_budget=25.0, + spend=12.5, + models=["gpt-4o", "gpt-4o-mini"], + blocked=True, + metadata={"tier": "gold"}, + litellm_model_table=_alias_table(model_aliases), + object_permission_id="op-1", + object_permission=LiteLLM_ObjectPermissionTable(object_permission_id="op-1", mcp_servers=["mcp-a"]), + members_with_roles=[ + Member(user_id="someone-else", role="user"), + Member(user_id=USER_ID, role="admin"), + ], + ) + + +def _membership() -> LiteLLM_TeamMembership: + return LiteLLM_TeamMembership( + user_id=USER_ID, + team_id=TEAM_ID, + spend=3.25, + litellm_budget_table=LiteLLM_BudgetTable(tpm_limit=500, rpm_limit=5), + ) + + +def test_team_grants_cover_every_team_field_the_key_path_gets(): + """Class guard for LIT-5858 and its siblings: every ``team_*`` column the combined-view SQL hands the + virtual-key path must come out of the projection too, with the team's actual value, so adding a column + to ``LiteLLM_VerificationTokenView`` without teaching ``team_grants`` fails here instead of in prod.""" + team = _full_team() + grants = team_grants(team_object=team, team_membership=_membership(), user_id=USER_ID) + token = UserAPIKeyAuth(team_id=TEAM_ID, **grants) + + view_team_fields = {name for name in LiteLLM_VerificationTokenView.model_fields if name.startswith("team_")} + assert view_team_fields - {"team_id"} <= set(grants) + assert all(grants[name] is not None for name in view_team_fields - {"team_id"}) + + assert token.team_alias == "grants-team" + assert token.team_tpm_limit == 1000 + assert token.team_rpm_limit == 10 + assert token.team_max_budget == 50.0 + assert token.team_soft_budget == 25.0 + assert token.team_spend == 12.5 + assert token.team_models == ["gpt-4o", "gpt-4o-mini"] + assert token.team_blocked is True + assert token.team_metadata == {"tier": "gold"} + assert token.team_model_aliases == ALIASES + assert token.team_object_permission_id == "op-1" + assert token.team_object_permission is not None + assert token.team_object_permission.mcp_servers == ["mcp-a"] + assert token.team_member == Member(user_id=USER_ID, role="admin") + assert token.team_member_spend == 3.25 + assert token.team_member_tpm_limit == 500 + assert token.team_member_rpm_limit == 5 + + +def test_team_grants_without_team_leave_token_defaults(): + token = UserAPIKeyAuth(**team_grants(team_object=None, team_membership=None, user_id=USER_ID)) + assert token == UserAPIKeyAuth() + + +@pytest.mark.parametrize( + "stored_aliases", + [ALIASES, '{"fast": "gpt-4o-mini", "smart": "gpt-4o"}'], + ids=["json-object", "json-string-as-written-by-team-new"], +) +def test_team_model_aliases_decode_both_storage_shapes(stored_aliases): + team = _full_team(model_aliases=stored_aliases) + assert team_model_aliases(team) == ALIASES + assert team_grants(team_object=team, team_membership=None, user_id=None)["team_model_aliases"] == ALIASES + + +@pytest.mark.parametrize("stored_aliases", [None, "not json", '["a", "b"]', {"fast": 3}], ids=str) +def test_team_model_aliases_treat_unusable_column_as_no_aliases(stored_aliases): + team = _full_team(model_aliases=stored_aliases) + assert team_model_aliases(team) is None + assert team_grants(team_object=team, team_membership=None, user_id=None)["team_model_aliases"] is None + + +def test_team_model_aliases_none_without_relation_loaded(): + team = _full_team() + team.litellm_model_table = None + assert team_model_aliases(team) is None + assert team_model_aliases(None) is None + + +def test_team_member_is_the_callers_row_only(): + team = _full_team() + assert team_grants(team_object=team, team_membership=None, user_id="someone-else")["team_member"] == Member( + user_id="someone-else", role="user" + ) + assert team_grants(team_object=team, team_membership=None, user_id="stranger")["team_member"] is None + assert team_grants(team_object=team, team_membership=None, user_id=None)["team_member"] is None + + +def test_membership_limits_absent_without_membership_row(): + grants = team_grants(team_object=_full_team(), team_membership=None, user_id=USER_ID) + assert grants["team_member_spend"] is None + assert grants["team_member_tpm_limit"] is None + assert grants["team_member_rpm_limit"] is None diff --git a/tests/test_litellm/proxy/auth/test_user_api_key_auth.py b/tests/test_litellm/proxy/auth/test_user_api_key_auth.py index d44f96d95bf..78ffbb0db23 100644 --- a/tests/test_litellm/proxy/auth/test_user_api_key_auth.py +++ b/tests/test_litellm/proxy/auth/test_user_api_key_auth.py @@ -6872,3 +6872,119 @@ class TestLitellmReceivedAtStamping: assert result == earlier assert request.state.litellm_received_at == earlier + + +@pytest.mark.asyncio +@pytest.mark.parametrize("is_proxy_admin", [False, True], ids=["standard-return", "proxy-admin-return"]) +async def test_jwt_builder_returns_every_team_grant_the_key_path_gets(is_proxy_admin): + """LIT-5858: the team-based JWT path hand-built ``UserAPIKeyAuth`` from a short list of team fields, so the + team's model aliases (and on the admin return, its object permission) never reached the token and alias + requests 403'd. Both returns now go through ``team_grants``; pin the fields that used to be dropped.""" + import litellm.proxy.proxy_server as _proxy_server_mod + from fastapi import Request + from starlette.datastructures import URL + + from litellm.models.team import LiteLLM_ModelTable + from litellm.proxy._types import ( + LiteLLM_ObjectPermissionTable, + LiteLLM_TeamMembership, + LiteLLM_TeamTable, + Member, + ) + + class _AcceptEveryJwt(JWTHandler): + def is_jwt(self, token: str) -> bool: + return True + + jwt_handler = _AcceptEveryJwt() + jwt_handler.litellm_jwtauth = LiteLLM_JWTAuth() + + team = LiteLLM_TeamTable( + team_id="team-jwt-aliases", + team_alias="jwt-aliases", + models=["gpt-4o"], + max_budget=40.0, + spend=4.0, + blocked=False, + metadata={"tier": "gold"}, + litellm_model_table=LiteLLM_ModelTable( + model_aliases='{"fast": "gpt-4o"}', created_by="admin", updated_by="admin" + ), + object_permission_id="op-jwt", + object_permission=LiteLLM_ObjectPermissionTable(object_permission_id="op-jwt", mcp_servers=["mcp-a"]), + members_with_roles=[Member(user_id="jwt-user", role="admin")], + ) + membership = LiteLLM_TeamMembership(user_id="jwt-user", team_id="team-jwt-aliases", spend=1.5) + builder_result = { + "is_proxy_admin": is_proxy_admin, + "team_object": team, + "user_object": None, + "end_user_object": None, + "org_object": None, + "token": "jwt", + "team_id": "team-jwt-aliases", + "user_id": "jwt-user", + "user_email": "jwt-user@example.com", + "end_user_id": None, + "org_id": None, + "team_membership": membership, + "jwt_claims": {"sub": "jwt-user"}, + } + + mock_proxy_logging_obj = MagicMock() + mock_proxy_logging_obj.internal_usage_cache = MagicMock() + mock_proxy_logging_obj.internal_usage_cache.dual_cache = AsyncMock() + mock_proxy_logging_obj.post_call_failure_hook = AsyncMock(return_value=None) + attrs = { + "prisma_client": MagicMock(), + "user_api_key_cache": DualCache(), + "proxy_logging_obj": mock_proxy_logging_obj, + "master_key": "sk-master-key", + "general_settings": {"enable_jwt_auth": True}, + "llm_model_list": [], + "llm_router": None, + "open_telemetry_logger": None, + "model_max_budget_limiter": MagicMock(), + "user_custom_auth": None, + "jwt_handler": jwt_handler, + "premium_user": True, + "litellm_proxy_admin_name": "admin", + } + originals = {a: getattr(_proxy_server_mod, a, None) for a in attrs} + try: + for k, v in attrs.items(): + setattr(_proxy_server_mod, k, v) + request = Request(scope={"type": "http", "headers": [], "method": "POST"}) + request._url = URL(url="/chat/completions") + with patch( # test-quality-ok: auth_builder is the claim-resolution seam; the regression is how its result is projected onto the token + "litellm.proxy.auth.user_api_key_auth.JWTAuthManager.auth_builder", + new_callable=AsyncMock, + return_value=builder_result, + ): + token = await _user_api_key_auth_builder( + request=request, + api_key="Bearer header.payload.signature", + azure_api_key_header="", + anthropic_api_key_header=None, + google_ai_studio_api_key_header=None, + azure_apim_header=None, + request_data={}, + ) + finally: + for k, v in originals.items(): + setattr(_proxy_server_mod, k, v) + + assert token.team_id == "team-jwt-aliases" + assert token.user_role == (LitellmUserRoles.PROXY_ADMIN if is_proxy_admin else LitellmUserRoles.INTERNAL_USER) + assert token.team_model_aliases == {"fast": "gpt-4o"} + assert token.team_object_permission is not None + assert token.team_object_permission.mcp_servers == ["mcp-a"] + assert token.team_object_permission_id == "op-jwt" + assert token.team_alias == "jwt-aliases" + assert token.team_models == ["gpt-4o"] + assert token.team_max_budget == 40.0 + assert token.team_spend == 4.0 + assert token.team_metadata == {"tier": "gold"} + assert token.team_member == Member(user_id="jwt-user", role="admin") + assert token.team_member_spend == 1.5 + assert token.jwt_claims == {"sub": "jwt-user"} diff --git a/tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py index 051e6bed4fd..2f6561046b1 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py +++ b/tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py @@ -2205,6 +2205,50 @@ async def test_team_model_add_delete_refresh_team_cache(endpoint_name): assert update_call_kwargs.get("include", {}).get("object_permission") is True +@pytest.mark.asyncio +@pytest.mark.parametrize("endpoint_name", ["team_model_add", "team_model_delete"]) +async def test_team_model_add_delete_keep_model_aliases_in_team_cache(endpoint_name, monkeypatch): + """LIT-5858: Prisma only returns `litellm_model_table` when the `update` asks for it, so the refreshed + cache entry lost the team's model aliases and JWT alias requests 403'd until the next DB read.""" + from litellm.proxy._types import TeamModelAddRequest, TeamModelDeleteRequest + from litellm.proxy.auth.team_grants import team_model_aliases + from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache + from litellm.proxy.management_endpoints.team_endpoints import team_model_add, team_model_delete + + columns = {"team_id": "team-1234", "models": ["gpt-4o", "openai/*"]} + alias_table = {"id": 1, "model_aliases": '{"fast": "gpt-4o"}', "created_by": "admin", "updated_by": "admin"} + + async def update(where, data, include=None): + row = {**columns, "litellm_model_table": alias_table} if (include or {}).get("litellm_model_table") else columns + return SimpleNamespace(team_id="team-1234", model_dump=lambda: row) + + prisma_client = MagicMock() + prisma_client.db.litellm_teamtable.find_unique = AsyncMock(return_value=SimpleNamespace(model_dump=lambda: columns)) + prisma_client.db.litellm_teamtable.update = AsyncMock(side_effect=update) + prisma_client.db.execute_raw = AsyncMock(return_value=None) + cache = UserApiKeyCache() + monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", prisma_client) + monkeypatch.setattr("litellm.proxy.proxy_server.user_api_key_cache", cache) + monkeypatch.setattr("litellm.proxy.proxy_server.proxy_logging_obj", None) + + admin = UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN, user_id="admin") + if endpoint_name == "team_model_add": + await team_model_add( + data=TeamModelAddRequest(team_id="team-1234", models=["team-byok-1"]), + http_request=MagicMock(), + user_api_key_dict=admin, + ) + else: + await team_model_delete( + data=TeamModelDeleteRequest(team_id="team-1234", models=["openai/*"]), + http_request=MagicMock(), + user_api_key_dict=admin, + ) + + cached_team = await cache.async_get_cache(key="team_id:team-1234", model_type=LiteLLM_TeamTableCachedObj) + assert team_model_aliases(cached_team) == {"fast": "gpt-4o"} + + @pytest.mark.asyncio @pytest.mark.parametrize( "endpoint_name", diff --git a/type-discipline-budget.json b/type-discipline-budget.json index e7186dfe186..7db8ed501f8 100644 --- a/type-discipline-budget.json +++ b/type-discipline-budget.json @@ -3,7 +3,7 @@ "limit": 22180 }, "LIT002": { - "limit": 26729 + "limit": 26727 }, "LIT003": { "limit": 261 From 0086b62b4575fb5bb69653c2ad3d196822bd3f13 Mon Sep 17 00:00:00 2001 From: kerry Date: Wed, 9 Sep 2026 02:05:08 +0000 Subject: [PATCH 2/8] fix(cost-map): keep first fetch blocking, run retries in background Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/__init__.py | 4 +- .../litellm_core_utils/get_model_cost_map.py | 132 +++++++++- litellm/proxy/proxy_server.py | 23 +- .../test_get_model_cost_map.py | 235 +++++++++++++----- 4 files changed, 296 insertions(+), 98 deletions(-) diff --git a/litellm/__init__.py b/litellm/__init__.py index 42c0ea881fd..4e3754399da 100644 --- a/litellm/__init__.py +++ b/litellm/__init__.py @@ -541,7 +541,7 @@ _key_management_system: Optional["KeyManagementSystem"] = None #### PII MASKING #### output_parse_pii: bool = False ############################################# -from litellm.litellm_core_utils.get_model_cost_map import get_model_cost_map +from litellm.litellm_core_utils.get_model_cost_map import get_model_cost_map, mark_litellm_import_complete model_cost = get_model_cost_map(url=model_cost_map_url) cost_discount_config: Dict[str, float] = {} # Provider-specific cost discounts {"vertex_ai": 0.05} = 5% discount @@ -2397,3 +2397,5 @@ def __getattr__(name: str) -> Any: # ALL_LITELLM_RESPONSE_TYPES is lazy-loaded via __getattr__ to avoid loading utils at import time + +mark_litellm_import_complete() diff --git a/litellm/litellm_core_utils/get_model_cost_map.py b/litellm/litellm_core_utils/get_model_cost_map.py index 9cba5db8ab7..4118cd3420e 100644 --- a/litellm/litellm_core_utils/get_model_cost_map.py +++ b/litellm/litellm_core_utils/get_model_cost_map.py @@ -12,6 +12,7 @@ import asyncio import json import os import random +import threading import time from collections.abc import Awaitable, Callable from dataclasses import dataclass @@ -159,6 +160,15 @@ class GetModelCostMap: RETRYABLE_FETCH_STATUS_CODES: Final = frozenset({429, 500, 502, 503, 504}) MODEL_COST_MAP_FETCH_MAX_ATTEMPTS: Final = 3 MODEL_COST_MAP_FETCH_MAX_WAIT_SECONDS: Final = 30.0 +_litellm_import_complete = threading.Event() + + +def mark_litellm_import_complete() -> None: + _litellm_import_complete.set() + + +def _start_daemon_thread(fn: Callable[[], None]) -> None: + threading.Thread(target=fn, name="litellm-model-cost-map-retry", daemon=True).start() @dataclass(frozen=True, slots=True) @@ -297,9 +307,15 @@ def _fetch_remote_model_cost_map_with_retry_sync( sleep: Callable[[float], None], rng: random.Random, client: _SyncGetClient, + starting_attempt: int = 1, + initial_outcome: _FetchAttemptRetryable | None = None, ) -> ModelCostMapReloadResult: - for attempt in range(1, max_attempts + 1): - outcome = _attempt_fetch_sync(client=client, url=url, timeout=timeout) + for attempt in range(starting_attempt, max_attempts + 1): + outcome = ( + initial_outcome + if initial_outcome is not None and attempt == starting_attempt + else _attempt_fetch_sync(client=client, url=url, timeout=timeout) + ) if not isinstance(outcome, _FetchAttemptRetryable): return outcome wait_seconds = _next_retry_wait(outcome=outcome, attempt=attempt, max_attempts=max_attempts, rng=rng) @@ -464,6 +480,70 @@ def _finalize_model_cost_map(model_cost: dict) -> dict: return _expand_model_aliases(model_cost) +def adopt_model_cost_map( + new_model_cost_map: dict, # mutable-ok: public API preserves the mutable cost-map contract +) -> int: + import litellm + from litellm import utils + + litellm.model_cost = new_model_cost_map + utils._invalidate_model_cost_lowercase_map() # pyright: ignore[reportPrivateUsage] # required cache invalidation + litellm.add_known_models(model_cost_map=new_model_cost_map) + fetched_model_count: Final = len(new_model_cost_map) if new_model_cost_map else 0 + utils.reapply_runtime_model_cost_registrations() + return fetched_model_count + + +def _continue_remote_fetch_in_background( + url: str, + timeout: int, + max_attempts: int, + sleep: Callable[[float], None], + rng: random.Random, + client: _SyncGetClient, + first_outcome: _FetchAttemptRetryable, + apply: Callable[[dict], object], # mutable-ok: injected callback receives the mutable cost-map dict +) -> None: + try: + result: Final = _fetch_remote_model_cost_map_with_retry_sync( + url=url, + timeout=timeout, + max_attempts=max_attempts, + sleep=sleep, + rng=rng, + client=client, + initial_outcome=first_outcome, + ) + if isinstance(result, ModelCostMapReloadUnavailable): + verbose_logger.warning( + "LiteLLM: Failed to fetch remote model cost map from %s after %d attempts; keeping local backup", + url, + max_attempts, + ) + return + backup_model_count: Final = GetModelCostMap._get_backup_model_count() # pyright: ignore[reportPrivateUsage] # integrity cache + if not GetModelCostMap.validate_model_cost_map( + fetched_map=result.model_cost_map, + backup_model_count=backup_model_count, + ): + verbose_logger.warning( + "LiteLLM: Fetched model cost map failed integrity check. Keeping local backup. url=%s", + url, + ) + _cost_map_source_info.fallback_reason = "Remote data failed integrity validation" + return + finalized_map: Final = _finalize_model_cost_map(result.model_cost_map) + _litellm_import_complete.wait() + apply(finalized_map) + _cost_map_source_info.source = "remote" + _cost_map_source_info.url = url + _cost_map_source_info.is_env_forced = False + _cost_map_source_info.fallback_reason = None + _cost_map_source_info.loaded_at = datetime.now(timezone.utc) + except Exception as e: + verbose_logger.warning("LiteLLM: Background model cost map retry failed: %s", e) + + def get_model_cost_map( url: str, timeout: int = 5, @@ -471,14 +551,19 @@ def get_model_cost_map( sleep: Callable[[float], None] = time.sleep, rng: random.Random | None = None, client: "_SyncGetClient | None" = None, + start_background: Callable[[Callable[[], None]], None] = _start_daemon_thread, + apply: Callable[ # mutable-ok: injected callback receives the mutable cost-map dict + [dict], + object, + ] = adopt_model_cost_map, ) -> dict: """ Public entry point — returns the model cost map dict. 1. If ``LITELLM_LOCAL_MODEL_COST_MAP`` is set, uses the local backup only. - 2. Otherwise fetches from ``url``, retrying transient HTTP errors - (429/5xx/transport) with Retry-After-aware backoff, validates - integrity, and falls back to the local backup on any failure. + 2. Otherwise fetches from ``url``, validates the first response, and falls + back to the local backup while retrying transient HTTP errors in the + background. Only the backup model count is cached (a single int) for validation. The full backup dict is only parsed when it must be *returned* as a @@ -497,14 +582,35 @@ def get_model_cost_map( _cost_map_source_info.url = url _cost_map_source_info.is_env_forced = False - result: Final = _fetch_remote_model_cost_map_with_retry_sync( - url=url, - timeout=timeout, - max_attempts=max_attempts, - sleep=sleep, - rng=rng if rng is not None else random.Random(), - client=client if client is not None else httpx, - ) + fetch_client: Final = client if client is not None else httpx + fetch_rng: Final = rng if rng is not None else random.Random() + first_outcome: Final = _attempt_fetch_sync(client=fetch_client, url=url, timeout=timeout) + if isinstance(first_outcome, _FetchAttemptRetryable): + verbose_logger.warning( + "LiteLLM: model cost map fetch attempt 1/%d failed: %s; " + "using local backup while retrying in the background", + max_attempts, + first_outcome.reason, + ) + _cost_map_source_info.source = "local" + _cost_map_source_info.fallback_reason = f"Remote fetch failed: {first_outcome.reason}" + local_map: Final = _finalize_model_cost_map(GetModelCostMap.load_local_model_cost_map()) + if max_attempts > 1: + start_background( + lambda: _continue_remote_fetch_in_background( + url=url, + timeout=timeout, + max_attempts=max_attempts, + sleep=sleep, + rng=fetch_rng, + client=fetch_client, + first_outcome=first_outcome, + apply=apply, + ) + ) + return local_map + + result: Final = first_outcome if isinstance(result, ModelCostMapReloadUnavailable): verbose_logger.warning( "LiteLLM: Failed to fetch remote model cost map from %s: %s. Falling back to local backup.", diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 1cd08fe27c0..efe38376353 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -132,11 +132,7 @@ from litellm.types.utils import ( TextCompletionResponse, TokenCountResponse, ) -from litellm.utils import ( - _invalidate_model_cost_lowercase_map, - load_credentials_from_list, - reapply_runtime_model_cost_registrations, -) +from litellm.utils import load_credentials_from_list if TYPE_CHECKING: from aiohttp import ClientSession @@ -4411,20 +4407,9 @@ def resolve_classifier_plugin( def _swap_in_model_cost_map(new_model_cost_map: dict) -> int: - """Adopt a freshly fetched cost map into this process's litellm state, return the model count""" - litellm.model_cost = new_model_cost_map - # Invalidate case-insensitive lookup map since model_cost was replaced - _invalidate_model_cost_lowercase_map() - # Repopulate provider model sets (e.g. litellm.anthropic_models) so that - # wildcard patterns like "anthropic/*" include any newly added models. - litellm.add_known_models(model_cost_map=new_model_cost_map) - # Counted before the re-apply below, which writes into this same dict, so the - # number reported describes the fetched price data alone. - fetched_model_count: Final = len(new_model_cost_map) if new_model_cost_map else 0 - # The swap discards everything registered at runtime (deployment model_info, - # register_model overrides), so put it back on top of the fresh catalog. - reapply_runtime_model_cost_registrations() - return fetched_model_count + from litellm.litellm_core_utils.get_model_cost_map import adopt_model_cost_map + + return adopt_model_cost_map(new_model_cost_map) def should_load_db_object(object_type: str | SupportedDBObjectType) -> bool: diff --git a/tests/test_litellm/litellm_core_utils/test_get_model_cost_map.py b/tests/test_litellm/litellm_core_utils/test_get_model_cost_map.py index 18185126775..dce1a431f13 100644 --- a/tests/test_litellm/litellm_core_utils/test_get_model_cost_map.py +++ b/tests/test_litellm/litellm_core_utils/test_get_model_cost_map.py @@ -20,21 +20,18 @@ from litellm.litellm_core_utils.get_model_cost_map import ( GetModelCostMap, _count_model_entries, _finalize_model_cost_map, + adopt_model_cost_map, ) def _load_root_cost_map() -> dict: - path = os.path.join( - os.path.dirname(__file__), "../../../model_prices_and_context_window.json" - ) + path = os.path.join(os.path.dirname(__file__), "../../../model_prices_and_context_window.json") with open(path) as f: return json.load(f) def _make_models(n: int) -> dict: - return { - f"model-{i}": {"litellm_provider": "openai", "mode": "chat"} for i in range(n) - } + return {f"model-{i}": {"litellm_provider": "openai", "mode": "chat"} for i in range(n)} def test_count_model_entries_excludes_reserved_keys(): @@ -117,9 +114,7 @@ def test_finalize_pops_key_and_installs_rules(): def test_finalize_with_no_block_clears_rules(): previous = list(get_fallback_generalization_rules()) try: - set_fallback_generalizations( - [{"name": "stale", "pattern": r"^x", "model_info": {"a": 1}}] - ) + set_fallback_generalizations([{"name": "stale", "pattern": r"^x", "model_info": {"a": 1}}]) _finalize_model_cost_map(_make_models(2)) assert match_capability_generalizations("x-1") is None finally: @@ -306,9 +301,7 @@ def test_get_model_cost_map_stamps_loaded_at(monkeypatch): from litellm.litellm_core_utils import get_model_cost_map as module monkeypatch.setattr(module._cost_map_source_info, "loaded_at", None) - client, _calls = _mock_client( - [httpx.Response(200, content=_real_map_bytes())], client_cls=httpx.Client - ) + client, _calls = _mock_client([httpx.Response(200, content=_real_map_bytes())], client_cls=httpx.Client) before = datetime.now(timezone.utc) module.get_model_cost_map(url="https://example.invalid/cost_map.json", client=client) @@ -317,6 +310,7 @@ def test_get_model_cost_map_stamps_loaded_at(monkeypatch): assert loaded_at is not None assert before <= loaded_at <= datetime.now(timezone.utc) + # --------------------------------------------------------------------------- # refetch_model_cost_map: retry/backoff behavior for runtime reloads # --------------------------------------------------------------------------- @@ -382,9 +376,7 @@ async def test_refetch_retries_429_honoring_retry_after(): ] ) sleeper = _SleepRecorder() - result = await refetch_model_cost_map( - url=_URL, sleep=sleeper, rng=random.Random(0), client=client - ) + result = await refetch_model_cost_map(url=_URL, sleep=sleeper, rng=random.Random(0), client=client) assert isinstance(result, ModelCostMapReloaded) assert len(result.model_cost_map) > 100 assert calls["count"] == 3 @@ -396,9 +388,7 @@ async def test_refetch_gives_up_after_max_attempts_with_exponential_backoff(): """All 429 without Retry-After: exponential backoff waits, then a failure value.""" client, calls = _mock_client([httpx.Response(429)]) sleeper = _SleepRecorder() - result = await refetch_model_cost_map( - url=_URL, sleep=sleeper, rng=random.Random(0), client=client - ) + result = await refetch_model_cost_map(url=_URL, sleep=sleeper, rng=random.Random(0), client=client) assert isinstance(result, ModelCostMapReloadUnavailable) assert "429" in result.reason assert "after 3 attempts" in result.reason @@ -418,9 +408,7 @@ async def test_refetch_caps_retry_after_wait(): ] ) sleeper = _SleepRecorder() - result = await refetch_model_cost_map( - url=_URL, sleep=sleeper, rng=random.Random(0), client=client - ) + result = await refetch_model_cost_map(url=_URL, sleep=sleeper, rng=random.Random(0), client=client) assert isinstance(result, ModelCostMapReloaded) assert sleeper.waits == [30.0] @@ -435,9 +423,7 @@ async def test_refetch_retries_transport_errors(): ] ) sleeper = _SleepRecorder() - result = await refetch_model_cost_map( - url=_URL, sleep=sleeper, rng=random.Random(0), client=client - ) + result = await refetch_model_cost_map(url=_URL, sleep=sleeper, rng=random.Random(0), client=client) assert isinstance(result, ModelCostMapReloaded) assert calls["count"] == 2 assert len(sleeper.waits) == 1 @@ -448,9 +434,7 @@ async def test_refetch_non_retryable_status_fails_immediately(): """A 404 is permanent: one attempt, no sleeps, failure value.""" client, calls = _mock_client([httpx.Response(404)]) sleeper = _SleepRecorder() - result = await refetch_model_cost_map( - url=_URL, sleep=sleeper, rng=random.Random(0), client=client - ) + result = await refetch_model_cost_map(url=_URL, sleep=sleeper, rng=random.Random(0), client=client) assert isinstance(result, ModelCostMapReloadUnavailable) assert "404" in result.reason assert calls["count"] == 1 @@ -461,9 +445,7 @@ async def test_refetch_non_retryable_status_fails_immediately(): async def test_refetch_invalid_json_fails_immediately(): client, calls = _mock_client([httpx.Response(200, content=b"not json")]) sleeper = _SleepRecorder() - result = await refetch_model_cost_map( - url=_URL, sleep=sleeper, rng=random.Random(0), client=client - ) + result = await refetch_model_cost_map(url=_URL, sleep=sleeper, rng=random.Random(0), client=client) assert isinstance(result, ModelCostMapReloadUnavailable) assert "invalid JSON" in result.reason assert calls["count"] == 1 @@ -475,9 +457,7 @@ async def test_refetch_shrunk_map_fails_integrity_not_swapped_in(): """A drastically shrunk upstream file is rejected instead of being adopted.""" tiny = json.dumps(_make_models(60)).encode() client, _calls = _mock_client([httpx.Response(200, content=tiny)]) - result = await refetch_model_cost_map( - url=_URL, sleep=_SleepRecorder(), rng=random.Random(0), client=client - ) + result = await refetch_model_cost_map(url=_URL, sleep=_SleepRecorder(), rng=random.Random(0), client=client) assert isinstance(result, ModelCostMapReloadUnavailable) assert "integrity validation" in result.reason @@ -520,62 +500,187 @@ class _SyncSleepRecorder: self.waits.append(seconds) -def test_boot_load_retries_transient_failures_instead_of_falling_back(): - """A refused connection then a 503 at pod boot used to pin the process to the bundled - backup for its lifetime; both are transient and must be retried before giving up.""" +class _BackgroundRecorder: + def __init__(self): + self.callbacks = [] + + def __call__(self, callback): + self.callbacks.append(callback) + + +def test_boot_load_returns_local_map_and_schedules_transient_retry(): + client, calls = _mock_client( + [ + httpx.ConnectError("connection refused"), + ], + client_cls=httpx.Client, + ) + sleeper = _SyncSleepRecorder() + background = _BackgroundRecorder() + + cost_map = get_model_cost_map( + url=_URL, + sleep=sleeper, + rng=random.Random(0), + client=client, + start_background=background, + ) + assert calls["count"] == 1 + assert sleeper.waits == [] + assert len(background.callbacks) == 1 + assert len(cost_map) > 100 + source = get_model_cost_map_source_info() + assert source["source"] == "local" + assert source["fallback_reason"] is not None + + +def test_background_retry_adopts_valid_remote_map(): client, calls = _mock_client( [ httpx.ConnectError("connection refused"), - httpx.Response(503), httpx.Response(200, content=_real_map_bytes()), ], client_cls=httpx.Client, ) sleeper = _SyncSleepRecorder() + background = _BackgroundRecorder() + applied = [] - cost_map = get_model_cost_map(url=_URL, sleep=sleeper, rng=random.Random(0), client=client) + get_model_cost_map( + url=_URL, + sleep=sleeper, + rng=random.Random(0), + client=client, + start_background=background, + apply=applied.append, + ) + assert calls["count"] == 1 + assert sleeper.waits == [] + assert len(background.callbacks) == 1 + + background.callbacks[0]() + + assert calls["count"] == 2 + assert len(sleeper.waits) == 1 + assert 2.0 <= sleeper.waits[0] < 3.0 + assert len(applied) == 1 + assert applied[0].keys() >= _load_root_cost_map().keys() - {"sample_spec", FALLBACK_GENERALIZATIONS_KEY} + source = get_model_cost_map_source_info() + assert source["source"] == "remote" + assert source["fallback_reason"] is None + + +def test_background_retry_keeps_local_map_after_remaining_failures(): + client, calls = _mock_client( + [httpx.ConnectError("connection refused"), httpx.Response(503)], + client_cls=httpx.Client, + ) + sleeper = _SyncSleepRecorder() + background = _BackgroundRecorder() + applied = [] + + get_model_cost_map( + url=_URL, + sleep=sleeper, + rng=random.Random(0), + client=client, + start_background=background, + apply=applied.append, + ) + background.callbacks[0]() assert calls["count"] == 3 assert len(sleeper.waits) == 2 assert 2.0 <= sleeper.waits[0] < 3.0 assert 4.0 <= sleeper.waits[1] < 5.0 + assert applied == [] + assert get_model_cost_map_source_info()["source"] == "local" + + +def test_boot_load_does_not_schedule_non_retryable_failure(): + client, calls = _mock_client([httpx.Response(404)], client_cls=httpx.Client) + sleeper = _SyncSleepRecorder() + background = _BackgroundRecorder() + + get_model_cost_map( + url=_URL, + sleep=sleeper, + rng=random.Random(0), + client=client, + start_background=background, + ) + + assert calls["count"] == 1 + assert sleeper.waits == [] + assert background.callbacks == [] source = get_model_cost_map_source_info() - assert source["source"] == "remote" - assert source["fallback_reason"] is None + assert source["source"] == "local" + assert source["fallback_reason"] is not None + + +def test_boot_load_success_does_not_schedule_background_retry(): + client, calls = _mock_client([httpx.Response(200, content=_real_map_bytes())], client_cls=httpx.Client) + sleeper = _SyncSleepRecorder() + background = _BackgroundRecorder() + + cost_map = get_model_cost_map( + url=_URL, + sleep=sleeper, + rng=random.Random(0), + client=client, + start_background=background, + ) + assert calls["count"] == 1 + assert sleeper.waits == [] + assert background.callbacks == [] + assert get_model_cost_map_source_info()["source"] == "remote" assert cost_map.keys() >= _load_root_cost_map().keys() - {"sample_spec", FALLBACK_GENERALIZATIONS_KEY} -def test_boot_load_honors_retry_after_then_falls_back_after_max_attempts(): - """An outage longer than the retry budget still ends on the bundled backup, and the - recorded fallback reason says how many attempts were spent so operators can tell.""" - client, calls = _mock_client( - [httpx.Response(429, headers={"Retry-After": "7"})], client_cls=httpx.Client +def test_boot_load_with_one_attempt_does_not_schedule_background_retry(): + client, calls = _mock_client([httpx.ConnectError("connection refused")], client_cls=httpx.Client) + sleeper = _SyncSleepRecorder() + background = _BackgroundRecorder() + + get_model_cost_map( + url=_URL, + max_attempts=1, + sleep=sleeper, + rng=random.Random(0), + client=client, + start_background=background, ) - sleeper = _SyncSleepRecorder() - cost_map = get_model_cost_map(url=_URL, sleep=sleeper, rng=random.Random(0), client=client) - - assert calls["count"] == 3 - assert sleeper.waits == [7.0, 7.0] - source = get_model_cost_map_source_info() - assert source["source"] == "local" - assert "after 3 attempts" in source["fallback_reason"] - assert len(cost_map) > 100 - - -def test_boot_load_does_not_retry_permanent_failures(): - """A 404 or a malformed URL cannot heal by waiting: one attempt, no sleeps, backup.""" - client, calls = _mock_client([httpx.Response(404)], client_cls=httpx.Client) - sleeper = _SyncSleepRecorder() - - get_model_cost_map(url=_URL, sleep=sleeper, rng=random.Random(0), client=client) assert calls["count"] == 1 assert sleeper.waits == [] + assert background.callbacks == [] assert get_model_cost_map_source_info()["source"] == "local" - get_model_cost_map(url="not a url", sleep=sleeper, rng=random.Random(0)) - assert sleeper.waits == [] - assert get_model_cost_map_source_info()["source"] == "local" + +def test_adopt_model_cost_map_replays_runtime_registration_and_provider_models(): + import litellm + from litellm import utils as litellm_utils + + original_model_cost = litellm.model_cost + original_registry = dict(litellm_utils._runtime_registered_model_cost) + original_anthropic_models = set(litellm.anthropic_models) + try: + litellm.register_model( + model_cost={"custom/deployment-model": {"litellm_provider": "custom", "max_input_tokens": 4321}} + ) + + models_count = adopt_model_cost_map({"anthropic/new-model": {"litellm_provider": "anthropic", "mode": "chat"}}) + + assert models_count == 1 + assert "anthropic/new-model" in litellm.anthropic_models + assert litellm.model_cost["custom/deployment-model"]["max_input_tokens"] == 4321 + finally: + litellm.model_cost = original_model_cost # test-quality-ok: restore the module state changed by adoption + litellm_utils._runtime_registered_model_cost.clear() + litellm_utils._runtime_registered_model_cost.update(original_registry) + litellm.anthropic_models.clear() + litellm.anthropic_models.update(original_anthropic_models) + litellm_utils._invalidate_model_cost_lowercase_map() def test_boot_load_respects_local_env_override(monkeypatch): From aa0a9ab3ea4dff53a22cbc60fbc0195c4ab6098c Mon Sep 17 00:00:00 2001 From: kerry Date: Wed, 9 Sep 2026 02:29:56 +0000 Subject: [PATCH 3/8] refactor(cost-map): drop initial_outcome flag from retry loop Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../litellm_core_utils/get_model_cost_map.py | 50 ++++++++++--------- 1 file changed, 26 insertions(+), 24 deletions(-) diff --git a/litellm/litellm_core_utils/get_model_cost_map.py b/litellm/litellm_core_utils/get_model_cost_map.py index 90ff696f926..d9b5a492539 100644 --- a/litellm/litellm_core_utils/get_model_cost_map.py +++ b/litellm/litellm_core_utils/get_model_cost_map.py @@ -324,19 +324,14 @@ async def _fetch_remote_model_cost_map_with_retry( def _fetch_remote_model_cost_map_with_retry_sync( url: str, timeout: int, - max_attempts: int, + attempts: range, sleep: Callable[[float], None], rng: random.Random, client: _SyncGetClient, - starting_attempt: int = 1, - initial_outcome: _FetchAttemptRetryable | None = None, ) -> ModelCostMapReloadResult: - for attempt in range(starting_attempt, max_attempts + 1): - outcome = ( - initial_outcome - if initial_outcome is not None and attempt == starting_attempt - else _attempt_fetch_sync(client=client, url=url, timeout=timeout) - ) + max_attempts: Final = attempts.stop - 1 + for attempt in attempts: + outcome = _attempt_fetch_sync(client=client, url=url, timeout=timeout) if not isinstance(outcome, _FetchAttemptRetryable): return outcome wait_seconds = _next_retry_wait(outcome=outcome, attempt=attempt, max_attempts=max_attempts, rng=rng) @@ -557,18 +552,18 @@ def _continue_remote_fetch_in_background( sleep: Callable[[float], None], rng: random.Random, client: _SyncGetClient, - first_outcome: _FetchAttemptRetryable, + first_wait: float, apply: Callable[[dict], object], # mutable-ok: injected callback receives the mutable cost-map dict ) -> None: try: + sleep(first_wait) result: Final = _fetch_remote_model_cost_map_with_retry_sync( url=url, timeout=timeout, - max_attempts=max_attempts, + attempts=range(2, max_attempts + 1), sleep=sleep, rng=rng, client=client, - initial_outcome=first_outcome, ) if isinstance(result, ModelCostMapReloadUnavailable): verbose_logger.warning( @@ -652,19 +647,26 @@ def get_model_cost_map( local_map: Final = _finalize_loaded_model_cost_map( GetModelCostMap.load_local_model_cost_map_with_revision() ).model_cost_map - if max_attempts > 1: - start_background( - lambda: _continue_remote_fetch_in_background( - url=url, - timeout=timeout, - max_attempts=max_attempts, - sleep=sleep, - rng=fetch_rng, - client=fetch_client, - first_outcome=first_outcome, - apply=apply, - ) + first_wait: Final = _next_retry_wait( + outcome=first_outcome, + attempt=1, + max_attempts=max_attempts, + rng=fetch_rng, + ) + if isinstance(first_wait, ModelCostMapReloadUnavailable): + return local_map + start_background( + lambda: _continue_remote_fetch_in_background( + url=url, + timeout=timeout, + max_attempts=max_attempts, + sleep=sleep, + rng=fetch_rng, + client=fetch_client, + first_wait=first_wait, + apply=apply, ) + ) return local_map result: Final = first_outcome From 536a85b42967fd0c9c6b49d2df9298232034427a Mon Sep 17 00:00:00 2001 From: kerry Date: Wed, 9 Sep 2026 02:48:17 +0000 Subject: [PATCH 4/8] fix(cost-map): keep register_model url fetch to a single attempt Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/utils.py | 2 +- tests/test_litellm/test_utils.py | 23 +++++++++++++++++++++++ 2 files changed, 24 insertions(+), 1 deletion(-) diff --git a/litellm/utils.py b/litellm/utils.py index 36b48d3b8d8..8df28870544 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -3079,7 +3079,7 @@ def register_model( # Convert stringified numbers to appropriate numeric types loaded_model_cost = model_cost elif isinstance(model_cost, str): - loaded_model_cost = litellm.get_model_cost_map(url=model_cost) + loaded_model_cost = litellm.get_model_cost_map(url=model_cost, max_attempts=1) if persist_across_reloads: _registrations: Final[Mapping[str, Mapping[str, object]]] = loaded_model_cost diff --git a/tests/test_litellm/test_utils.py b/tests/test_litellm/test_utils.py index fee5e3a2e4c..e90372141bd 100644 --- a/tests/test_litellm/test_utils.py +++ b/tests/test_litellm/test_utils.py @@ -2,10 +2,12 @@ import asyncio import json import logging import os +import threading from datetime import datetime, timedelta, timezone from typing import Final from unittest.mock import AsyncMock, MagicMock, patch +import httpx import pytest import respx from jsonschema import validate @@ -2382,6 +2384,27 @@ def test_register_model_with_scientific_notation(): _invalidate_model_cost_lowercase_map() +@respx.mock +def test_register_model_url_fetch_uses_single_attempt(monkeypatch): + monkeypatch.setattr(litellm, "model_cost", dict(litellm.model_cost)) + before = dict(litellm.model_cost) + threads_before = {thread.name for thread in threading.enumerate()} + route = respx.get("https://example.invalid/custom_pricing.json").mock( + return_value=httpx.Response(503) + ) + + litellm.register_model(model_cost="https://example.invalid/custom_pricing.json") + + threads_after = {thread.name for thread in threading.enumerate()} + assert route.call_count == 1 + assert not (threads_after - threads_before) & {"litellm-model-cost-map-retry"} + assert not any( + thread.name == "litellm-model-cost-map-retry" and thread.is_alive() + for thread in threading.enumerate() + ) + assert litellm.model_cost.keys() >= before.keys() + + def test_register_model_openrouter_without_slash(): """ Test that register_model handles openrouter models without '/' in the name. From 4179f086e0196107dfc83da2a4e0bb1d15568edd Mon Sep 17 00:00:00 2001 From: kerry Date: Wed, 9 Sep 2026 02:58:51 +0000 Subject: [PATCH 5/8] refactor(cost-map): share local-fallback and remote-accept paths Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../litellm_core_utils/get_model_cost_map.py | 111 ++++++++---------- 1 file changed, 47 insertions(+), 64 deletions(-) diff --git a/litellm/litellm_core_utils/get_model_cost_map.py b/litellm/litellm_core_utils/get_model_cost_map.py index d9b5a492539..67d06a0758e 100644 --- a/litellm/litellm_core_utils/get_model_cost_map.py +++ b/litellm/litellm_core_utils/get_model_cost_map.py @@ -531,6 +531,32 @@ def _finalize_loaded_model_cost_map(loaded: ModelCostMapReloaded) -> ModelCostMa return replace(loaded, model_cost_map=_finalize_model_cost_map(loaded.model_cost_map)) +def _use_local_backup(reason: str | None) -> dict: # mutable-ok: returns the mutable model-cost map contract + _cost_map_source_info.source = "local" + _cost_map_source_info.fallback_reason = reason + return _finalize_loaded_model_cost_map(GetModelCostMap.load_local_model_cost_map_with_revision()).model_cost_map + + +def _accept_remote( + result: ModelCostMapReloaded, url: str +) -> dict | None: # mutable-ok: returns the mutable model-cost map contract + if not GetModelCostMap.validate_model_cost_map( + fetched_map=result.model_cost_map, + backup_model_count=GetModelCostMap._get_backup_model_count(), # pyright: ignore[reportPrivateUsage] # integrity cache + ): + verbose_logger.warning( + "LiteLLM: Fetched model cost map failed integrity check. Using local backup instead. url=%s", + url, + ) + return None + finalized: Final = _finalize_loaded_model_cost_map(result).model_cost_map + _cost_map_source_info.source = "remote" + _cost_map_source_info.url = url + _cost_map_source_info.is_env_forced = False + _cost_map_source_info.fallback_reason = None + return finalized + + def adopt_model_cost_map( new_model_cost_map: dict, # mutable-ok: public API preserves the mutable cost-map contract ) -> int: @@ -545,17 +571,20 @@ def adopt_model_cost_map( return fetched_model_count -def _continue_remote_fetch_in_background( +def _retry_remote_fetch_in_background( url: str, timeout: int, max_attempts: int, sleep: Callable[[float], None], rng: random.Random, client: _SyncGetClient, - first_wait: float, + first_outcome: _FetchAttemptRetryable, apply: Callable[[dict], object], # mutable-ok: injected callback receives the mutable cost-map dict ) -> None: try: + first_wait: Final = _next_retry_wait(outcome=first_outcome, attempt=1, max_attempts=max_attempts, rng=rng) + if isinstance(first_wait, ModelCostMapReloadUnavailable): + return sleep(first_wait) result: Final = _fetch_remote_model_cost_map_with_retry_sync( url=url, @@ -572,23 +601,12 @@ def _continue_remote_fetch_in_background( max_attempts, ) return - backup_model_count: Final = GetModelCostMap._get_backup_model_count() # pyright: ignore[reportPrivateUsage] # integrity cache - if not GetModelCostMap.validate_model_cost_map( - fetched_map=result.model_cost_map, - backup_model_count=backup_model_count, - ): - verbose_logger.warning( - "LiteLLM: Fetched model cost map failed integrity check. Keeping local backup. url=%s", - url, - ) + _litellm_import_complete.wait() + accepted: Final = _accept_remote(result, url) + if accepted is None: _cost_map_source_info.fallback_reason = "Remote data failed integrity validation" return - _litellm_import_complete.wait() - apply(_finalize_loaded_model_cost_map(result).model_cost_map) - _cost_map_source_info.source = "remote" - _cost_map_source_info.url = url - _cost_map_source_info.is_env_forced = False - _cost_map_source_info.fallback_reason = None + apply(accepted) _cost_map_source_info.loaded_at = datetime.now(timezone.utc) except Exception as e: verbose_logger.warning("LiteLLM: Background model cost map retry failed: %s", e) @@ -623,77 +641,42 @@ def get_model_cost_map( # Note: can't use get_secret_bool here — this runs during litellm.__init__ # before litellm._key_management_settings is set. if os.getenv("LITELLM_LOCAL_MODEL_COST_MAP", "").lower() == "true": - _cost_map_source_info.source = "local" _cost_map_source_info.url = None _cost_map_source_info.is_env_forced = True - _cost_map_source_info.fallback_reason = None - return _finalize_loaded_model_cost_map(GetModelCostMap.load_local_model_cost_map_with_revision()).model_cost_map + return _use_local_backup(None) _cost_map_source_info.url = url _cost_map_source_info.is_env_forced = False fetch_client: Final = client if client is not None else httpx fetch_rng: Final = rng if rng is not None else random.Random() - first_outcome: Final = _attempt_fetch_sync(client=fetch_client, url=url, timeout=timeout) - if isinstance(first_outcome, _FetchAttemptRetryable): + outcome: Final = _attempt_fetch_sync(client=fetch_client, url=url, timeout=timeout) + if isinstance(outcome, ModelCostMapReloaded): + accepted: Final = _accept_remote(outcome, url) + return accepted if accepted is not None else _use_local_backup("Remote data failed integrity validation") + if isinstance(outcome, _FetchAttemptRetryable) and max_attempts > 1: verbose_logger.warning( "LiteLLM: model cost map fetch attempt 1/%d failed: %s; " "using local backup while retrying in the background", max_attempts, - first_outcome.reason, + outcome.reason, ) - _cost_map_source_info.source = "local" - _cost_map_source_info.fallback_reason = f"Remote fetch failed: {first_outcome.reason}" - local_map: Final = _finalize_loaded_model_cost_map( - GetModelCostMap.load_local_model_cost_map_with_revision() - ).model_cost_map - first_wait: Final = _next_retry_wait( - outcome=first_outcome, - attempt=1, - max_attempts=max_attempts, - rng=fetch_rng, - ) - if isinstance(first_wait, ModelCostMapReloadUnavailable): - return local_map start_background( - lambda: _continue_remote_fetch_in_background( + lambda: _retry_remote_fetch_in_background( url=url, timeout=timeout, max_attempts=max_attempts, sleep=sleep, rng=fetch_rng, client=fetch_client, - first_wait=first_wait, + first_outcome=outcome, apply=apply, ) ) - return local_map - - result: Final = first_outcome - if isinstance(result, ModelCostMapReloadUnavailable): + else: verbose_logger.warning( "LiteLLM: Failed to fetch remote model cost map from %s: %s. Falling back to local backup.", url, - result.reason, + outcome.reason, ) - _cost_map_source_info.source = "local" - _cost_map_source_info.fallback_reason = f"Remote fetch failed: {result.reason}" - return _finalize_loaded_model_cost_map(GetModelCostMap.load_local_model_cost_map_with_revision()).model_cost_map - content: Final = result.model_cost_map - - # Validate using cached count (cheap int comparison, no file I/O) - if not GetModelCostMap.validate_model_cost_map( - fetched_map=content, - backup_model_count=GetModelCostMap._get_backup_model_count(), - ): - verbose_logger.warning( - "LiteLLM: Fetched model cost map failed integrity check. Using local backup instead. url=%s", - url, - ) - _cost_map_source_info.source = "local" - _cost_map_source_info.fallback_reason = "Remote data failed integrity validation" - return _finalize_loaded_model_cost_map(GetModelCostMap.load_local_model_cost_map_with_revision()).model_cost_map - - _cost_map_source_info.source = "remote" - _cost_map_source_info.fallback_reason = None - return _finalize_loaded_model_cost_map(result).model_cost_map + return _use_local_backup(f"Remote fetch failed: {outcome.reason}") From 9a721abf0d098f40caa6ba67d343e99de74f7fc2 Mon Sep 17 00:00:00 2001 From: kerry Date: Wed, 9 Sep 2026 03:50:49 +0000 Subject: [PATCH 6/8] test(cost-map): clear LITELLM_LOCAL_MODEL_COST_MAP in register_model url test Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/test_litellm/test_utils.py | 1 + 1 file changed, 1 insertion(+) diff --git a/tests/test_litellm/test_utils.py b/tests/test_litellm/test_utils.py index e90372141bd..e42608c9904 100644 --- a/tests/test_litellm/test_utils.py +++ b/tests/test_litellm/test_utils.py @@ -2386,6 +2386,7 @@ def test_register_model_with_scientific_notation(): @respx.mock def test_register_model_url_fetch_uses_single_attempt(monkeypatch): + monkeypatch.delenv("LITELLM_LOCAL_MODEL_COST_MAP", raising=False) monkeypatch.setattr(litellm, "model_cost", dict(litellm.model_cost)) before = dict(litellm.model_cost) threads_before = {thread.name for thread in threading.enumerate()} From 884f90c72715b0cfe0d5e2ea09f3744cb8376a78 Mon Sep 17 00:00:00 2001 From: kerry Date: Wed, 9 Sep 2026 05:15:53 +0000 Subject: [PATCH 7/8] refactor(cost-map): inline background retry, trim tests Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../litellm_core_utils/get_model_cost_map.py | 119 ++++----- .../test_get_model_cost_map.py | 243 +++++++----------- 2 files changed, 139 insertions(+), 223 deletions(-) diff --git a/litellm/litellm_core_utils/get_model_cost_map.py b/litellm/litellm_core_utils/get_model_cost_map.py index 67d06a0758e..f81ddbfee2e 100644 --- a/litellm/litellm_core_utils/get_model_cost_map.py +++ b/litellm/litellm_core_utils/get_model_cost_map.py @@ -184,10 +184,6 @@ def mark_litellm_import_complete() -> None: _litellm_import_complete.set() -def _start_daemon_thread(fn: Callable[[], None]) -> None: - threading.Thread(target=fn, name="litellm-model-cost-map-retry", daemon=True).start() - - @dataclass(frozen=True, slots=True) class ModelCostMapReloaded: model_cost_map: dict # mutable-ok: adopted as litellm.model_cost, whose consumer contract is a plain mutable dict @@ -531,32 +527,6 @@ def _finalize_loaded_model_cost_map(loaded: ModelCostMapReloaded) -> ModelCostMa return replace(loaded, model_cost_map=_finalize_model_cost_map(loaded.model_cost_map)) -def _use_local_backup(reason: str | None) -> dict: # mutable-ok: returns the mutable model-cost map contract - _cost_map_source_info.source = "local" - _cost_map_source_info.fallback_reason = reason - return _finalize_loaded_model_cost_map(GetModelCostMap.load_local_model_cost_map_with_revision()).model_cost_map - - -def _accept_remote( - result: ModelCostMapReloaded, url: str -) -> dict | None: # mutable-ok: returns the mutable model-cost map contract - if not GetModelCostMap.validate_model_cost_map( - fetched_map=result.model_cost_map, - backup_model_count=GetModelCostMap._get_backup_model_count(), # pyright: ignore[reportPrivateUsage] # integrity cache - ): - verbose_logger.warning( - "LiteLLM: Fetched model cost map failed integrity check. Using local backup instead. url=%s", - url, - ) - return None - finalized: Final = _finalize_loaded_model_cost_map(result).model_cost_map - _cost_map_source_info.source = "remote" - _cost_map_source_info.url = url - _cost_map_source_info.is_env_forced = False - _cost_map_source_info.fallback_reason = None - return finalized - - def adopt_model_cost_map( new_model_cost_map: dict, # mutable-ok: public API preserves the mutable cost-map contract ) -> int: @@ -579,7 +549,6 @@ def _retry_remote_fetch_in_background( rng: random.Random, client: _SyncGetClient, first_outcome: _FetchAttemptRetryable, - apply: Callable[[dict], object], # mutable-ok: injected callback receives the mutable cost-map dict ) -> None: try: first_wait: Final = _next_retry_wait(outcome=first_outcome, attempt=1, max_attempts=max_attempts, rng=rng) @@ -602,12 +571,20 @@ def _retry_remote_fetch_in_background( ) return _litellm_import_complete.wait() - accepted: Final = _accept_remote(result, url) - if accepted is None: - _cost_map_source_info.fallback_reason = "Remote data failed integrity validation" + if not GetModelCostMap.validate_model_cost_map( + fetched_map=result.model_cost_map, + backup_model_count=GetModelCostMap._get_backup_model_count(), # pyright: ignore[reportPrivateUsage] # integrity cache + ): + verbose_logger.warning( + "LiteLLM: Fetched model cost map failed integrity check. Using local backup instead. url=%s", + url, + ) return - apply(accepted) + finalized: Final = _finalize_loaded_model_cost_map(result).model_cost_map + _cost_map_source_info.source = "remote" + _cost_map_source_info.fallback_reason = None _cost_map_source_info.loaded_at = datetime.now(timezone.utc) + adopt_model_cost_map(finalized) except Exception as e: verbose_logger.warning("LiteLLM: Background model cost map retry failed: %s", e) @@ -619,19 +596,12 @@ def get_model_cost_map( sleep: Callable[[float], None] = time.sleep, rng: random.Random | None = None, client: "_SyncGetClient | None" = None, - start_background: Callable[[Callable[[], None]], None] = _start_daemon_thread, - apply: Callable[ # mutable-ok: injected callback receives the mutable cost-map dict - [dict], - object, - ] = adopt_model_cost_map, ) -> dict: """ Public entry point — returns the model cost map dict. 1. If ``LITELLM_LOCAL_MODEL_COST_MAP`` is set, uses the local backup only. - 2. Otherwise fetches from ``url``, validates the first response, and falls - back to the local backup while retrying transient HTTP errors in the - background. + 2. Otherwise fetches from ``url``, retrying transient errors in a background thread. Only the backup model count is cached (a single int) for validation. The full backup dict is only parsed when it must be *returned* as a @@ -641,9 +611,11 @@ def get_model_cost_map( # Note: can't use get_secret_bool here — this runs during litellm.__init__ # before litellm._key_management_settings is set. if os.getenv("LITELLM_LOCAL_MODEL_COST_MAP", "").lower() == "true": + _cost_map_source_info.source = "local" _cost_map_source_info.url = None _cost_map_source_info.is_env_forced = True - return _use_local_backup(None) + _cost_map_source_info.fallback_reason = None + return _finalize_loaded_model_cost_map(GetModelCostMap.load_local_model_cost_map_with_revision()).model_cost_map _cost_map_source_info.url = url _cost_map_source_info.is_env_forced = False @@ -651,32 +623,45 @@ def get_model_cost_map( fetch_client: Final = client if client is not None else httpx fetch_rng: Final = rng if rng is not None else random.Random() outcome: Final = _attempt_fetch_sync(client=fetch_client, url=url, timeout=timeout) - if isinstance(outcome, ModelCostMapReloaded): - accepted: Final = _accept_remote(outcome, url) - return accepted if accepted is not None else _use_local_backup("Remote data failed integrity validation") if isinstance(outcome, _FetchAttemptRetryable) and max_attempts > 1: - verbose_logger.warning( - "LiteLLM: model cost map fetch attempt 1/%d failed: %s; " - "using local backup while retrying in the background", - max_attempts, - outcome.reason, - ) - start_background( - lambda: _retry_remote_fetch_in_background( - url=url, - timeout=timeout, - max_attempts=max_attempts, - sleep=sleep, - rng=fetch_rng, - client=fetch_client, - first_outcome=outcome, - apply=apply, - ) - ) - else: + threading.Thread( + target=_retry_remote_fetch_in_background, + kwargs={ # mutable-ok: threading requires a mutable keyword-arguments mapping + "url": url, + "timeout": timeout, + "max_attempts": max_attempts, + "sleep": sleep, + "rng": fetch_rng, + "client": fetch_client, + "first_outcome": outcome, + }, + name="litellm-model-cost-map-retry", + daemon=True, + ).start() + if not isinstance(outcome, ModelCostMapReloaded): verbose_logger.warning( "LiteLLM: Failed to fetch remote model cost map from %s: %s. Falling back to local backup.", url, outcome.reason, ) - return _use_local_backup(f"Remote fetch failed: {outcome.reason}") + _cost_map_source_info.source = "local" + _cost_map_source_info.fallback_reason = f"Remote fetch failed: {outcome.reason}" + return _finalize_loaded_model_cost_map(GetModelCostMap.load_local_model_cost_map_with_revision()).model_cost_map + content: Final = outcome.model_cost_map + + # Validate using cached count (cheap int comparison, no file I/O) + if not GetModelCostMap.validate_model_cost_map( + fetched_map=content, + backup_model_count=GetModelCostMap._get_backup_model_count(), + ): + verbose_logger.warning( + "LiteLLM: Fetched model cost map failed integrity check. Using local backup instead. url=%s", + url, + ) + _cost_map_source_info.source = "local" + _cost_map_source_info.fallback_reason = "Remote data failed integrity validation" + return _finalize_loaded_model_cost_map(GetModelCostMap.load_local_model_cost_map_with_revision()).model_cost_map + + _cost_map_source_info.source = "remote" + _cost_map_source_info.fallback_reason = None + return _finalize_loaded_model_cost_map(outcome).model_cost_map diff --git a/tests/test_litellm/litellm_core_utils/test_get_model_cost_map.py b/tests/test_litellm/litellm_core_utils/test_get_model_cost_map.py index e39658a8f37..266c2ca1465 100644 --- a/tests/test_litellm/litellm_core_utils/test_get_model_cost_map.py +++ b/tests/test_litellm/litellm_core_utils/test_get_model_cost_map.py @@ -6,6 +6,7 @@ count actual model entries, not reserved meta keys) and the extraction of the import json import os +import threading import pytest @@ -20,7 +21,6 @@ from litellm.litellm_core_utils.get_model_cost_map import ( GetModelCostMap, _count_model_entries, _finalize_model_cost_map, - adopt_model_cost_map, get_model_cost_map_provenance, git_blob_id, ) @@ -566,194 +566,125 @@ from litellm.litellm_core_utils.get_model_cost_map import ( class _SyncSleepRecorder: """Injected in place of time.sleep so the boot path's waits are asserted without delay.""" - def __init__(self): + def __init__(self, block=False): self.waits = [] + self.block = block + self.started = threading.Event() + self.release = threading.Event() def __call__(self, seconds: float) -> None: + if self.block: + self.started.set() + self.release.wait(timeout=10) self.waits.append(seconds) -class _BackgroundRecorder: - def __init__(self): - self.callbacks = [] - - def __call__(self, callback): - self.callbacks.append(callback) +def _retry_threads(): + return [thread for thread in threading.enumerate() if thread.name == "litellm-model-cost-map-retry"] -def test_boot_load_returns_local_map_and_schedules_transient_retry(): - client, calls = _mock_client( - [ - httpx.ConnectError("connection refused"), - ], - client_cls=httpx.Client, - ) - sleeper = _SyncSleepRecorder() - background = _BackgroundRecorder() - - cost_map = get_model_cost_map( - url=_URL, - sleep=sleeper, - rng=random.Random(0), - client=client, - start_background=background, - ) - assert calls["count"] == 1 - assert sleeper.waits == [] - assert len(background.callbacks) == 1 - assert len(cost_map) > 100 - source = get_model_cost_map_source_info() - assert source["source"] == "local" - assert source["fallback_reason"] is not None - - -def test_background_retry_adopts_valid_remote_map(): - client, calls = _mock_client( - [ - httpx.ConnectError("connection refused"), - httpx.Response(200, content=_real_map_bytes()), - ], - client_cls=httpx.Client, - ) - sleeper = _SyncSleepRecorder() - background = _BackgroundRecorder() - applied = [] - - get_model_cost_map( - url=_URL, - sleep=sleeper, - rng=random.Random(0), - client=client, - start_background=background, - apply=applied.append, - ) - assert calls["count"] == 1 - assert sleeper.waits == [] - assert len(background.callbacks) == 1 - - background.callbacks[0]() - - assert calls["count"] == 2 - assert len(sleeper.waits) == 1 - assert 2.0 <= sleeper.waits[0] < 3.0 - assert len(applied) == 1 - assert applied[0].keys() >= _load_root_cost_map().keys() - {"sample_spec", FALLBACK_GENERALIZATIONS_KEY} - source = get_model_cost_map_source_info() - assert source["source"] == "remote" - assert source["fallback_reason"] is None - - -def test_background_retry_keeps_local_map_after_remaining_failures(): - client, calls = _mock_client( - [httpx.ConnectError("connection refused"), httpx.Response(503)], - client_cls=httpx.Client, - ) - sleeper = _SyncSleepRecorder() - background = _BackgroundRecorder() - applied = [] - - get_model_cost_map( - url=_URL, - sleep=sleeper, - rng=random.Random(0), - client=client, - start_background=background, - apply=applied.append, - ) - background.callbacks[0]() - - assert calls["count"] == 3 - assert len(sleeper.waits) == 2 - assert 2.0 <= sleeper.waits[0] < 3.0 - assert 4.0 <= sleeper.waits[1] < 5.0 - assert applied == [] - assert get_model_cost_map_source_info()["source"] == "local" - - -def test_boot_load_does_not_schedule_non_retryable_failure(): - client, calls = _mock_client([httpx.Response(404)], client_cls=httpx.Client) - sleeper = _SyncSleepRecorder() - background = _BackgroundRecorder() - - get_model_cost_map( - url=_URL, - sleep=sleeper, - rng=random.Random(0), - client=client, - start_background=background, - ) - - assert calls["count"] == 1 - assert sleeper.waits == [] - assert background.callbacks == [] - source = get_model_cost_map_source_info() - assert source["source"] == "local" - assert source["fallback_reason"] is not None - - -def test_boot_load_success_does_not_schedule_background_retry(): +def test_boot_load_success_does_not_start_background_retry(): client, calls = _mock_client([httpx.Response(200, content=_real_map_bytes())], client_cls=httpx.Client) sleeper = _SyncSleepRecorder() - background = _BackgroundRecorder() cost_map = get_model_cost_map( url=_URL, sleep=sleeper, rng=random.Random(0), client=client, - start_background=background, ) assert calls["count"] == 1 assert sleeper.waits == [] - assert background.callbacks == [] - assert get_model_cost_map_source_info()["source"] == "remote" + assert _retry_threads() == [] assert cost_map.keys() >= _load_root_cost_map().keys() - {"sample_spec", FALLBACK_GENERALIZATIONS_KEY} + assert get_model_cost_map_source_info()["source"] == "remote" -def test_boot_load_with_one_attempt_does_not_schedule_background_retry(): - client, calls = _mock_client([httpx.ConnectError("connection refused")], client_cls=httpx.Client) - sleeper = _SyncSleepRecorder() - background = _BackgroundRecorder() +def test_boot_load_transient_failure_returns_local_then_background_retry_adopts_remote(monkeypatch): + import litellm + from litellm import utils as litellm_utils + from litellm.litellm_core_utils import get_model_cost_map as module - get_model_cost_map( + original_model_cost = litellm.model_cost + monkeypatch.setattr(litellm, "model_cost", dict(original_model_cost)) + for name, provider_models in tuple(vars(litellm).items()): + if name.endswith("_models") and isinstance(provider_models, set): + monkeypatch.setattr(litellm, name, set(provider_models)) + monkeypatch.setattr(litellm, "models_by_provider", dict(litellm.models_by_provider)) + monkeypatch.setattr( + litellm_utils, + "_runtime_registered_model_cost", + dict(litellm_utils._runtime_registered_model_cost), + ) + source_info = module._cost_map_source_info + for name in ("source", "url", "is_env_forced", "fallback_reason", "loaded_at", "source_revision", "etag"): + monkeypatch.setattr(source_info, name, getattr(source_info, name)) + + remote_map = _load_root_cost_map() + remote_map["claude-remote-only-test"] = {"litellm_provider": "anthropic", "mode": "chat"} + client, calls = _mock_client( + [ + httpx.ConnectError("connection refused"), + httpx.Response(200, content=json.dumps(remote_map).encode()), + ], + client_cls=httpx.Client, + ) + sleeper = _SyncSleepRecorder(block=True) + litellm.register_model({"my-runtime-model": {"litellm_provider": "custom", "max_input_tokens": 4321}}) + + cost_map = get_model_cost_map( url=_URL, - max_attempts=1, + max_attempts=3, sleep=sleeper, rng=random.Random(0), client=client, - start_background=background, ) assert calls["count"] == 1 assert sleeper.waits == [] - assert background.callbacks == [] - assert get_model_cost_map_source_info()["source"] == "local" - - -def test_adopt_model_cost_map_replays_runtime_registration_and_provider_models(): - import litellm - from litellm import utils as litellm_utils - - original_model_cost = litellm.model_cost - original_registry = dict(litellm_utils._runtime_registered_model_cost) - original_anthropic_models = set(litellm.anthropic_models) + assert sleeper.started.wait(timeout=10) + threads = _retry_threads() try: - litellm.register_model( - model_cost={"custom/deployment-model": {"litellm_provider": "custom", "max_input_tokens": 4321}} - ) - - models_count = adopt_model_cost_map({"anthropic/new-model": {"litellm_provider": "anthropic", "mode": "chat"}}) - - assert models_count == 1 - assert "anthropic/new-model" in litellm.anthropic_models - assert litellm.model_cost["custom/deployment-model"]["max_input_tokens"] == 4321 + assert len(threads) == 1 + assert "claude-remote-only-test" not in cost_map + assert cost_map.keys() == _finalize_model_cost_map(GetModelCostMap.load_local_model_cost_map()).keys() + source = get_model_cost_map_source_info() + assert source["source"] == "local" + assert source["fallback_reason"].startswith("Remote fetch failed:") + sleeper.release.set() + for thread in threads: + thread.join(timeout=10) + assert all(not thread.is_alive() for thread in threads) + assert sleeper.waits and 2.0 <= sleeper.waits[0] < 3.0 + assert calls["count"] == 2 + assert "claude-remote-only-test" in litellm.model_cost + assert "claude-remote-only-test" in litellm.anthropic_models + assert "my-runtime-model" in litellm.model_cost + source = get_model_cost_map_source_info() + assert source["source"] == "remote" + assert source["fallback_reason"] is None finally: - litellm.model_cost = original_model_cost # test-quality-ok: restore the module state changed by adoption - litellm_utils._runtime_registered_model_cost.clear() - litellm_utils._runtime_registered_model_cost.update(original_registry) - litellm.anthropic_models.clear() - litellm.anthropic_models.update(original_anthropic_models) - litellm_utils._invalidate_model_cost_lowercase_map() + sleeper.release.set() + for thread in _retry_threads(): + thread.join(timeout=10) + + +def test_boot_load_does_not_retry_non_retryable_failure(): + client, calls = _mock_client([httpx.Response(404)], client_cls=httpx.Client) + sleeper = _SyncSleepRecorder() + + get_model_cost_map( + url=_URL, + sleep=sleeper, + rng=random.Random(0), + client=client, + ) + assert calls["count"] == 1 + assert sleeper.waits == [] + assert _retry_threads() == [] + source = get_model_cost_map_source_info() + assert source["source"] == "local" + assert source["fallback_reason"] is not None def test_boot_load_respects_local_env_override(monkeypatch): From 1183b2abc6645a92f9a852daa7843662b6fa6f20 Mon Sep 17 00:00:00 2001 From: yujonglee Date: Wed, 9 Sep 2026 09:46:37 -0700 Subject: [PATCH 8/8] fix(integrations): pass original request object to post-call guardrail hooks (#40414) --- litellm/integrations/custom_guardrail.py | 30 ++++--- .../integrations/test_custom_guardrail.py | 90 ++++++++++++++++++- 2 files changed, 106 insertions(+), 14 deletions(-) diff --git a/litellm/integrations/custom_guardrail.py b/litellm/integrations/custom_guardrail.py index 37d6a7e793d..77bf4820a1a 100644 --- a/litellm/integrations/custom_guardrail.py +++ b/litellm/integrations/custom_guardrail.py @@ -850,20 +850,24 @@ class CustomGuardrail(CustomLogger): if self.should_run_guardrail(data=request_data, event_type=GuardrailEventHooks.post_call) is not True: return None - # CHECK IF GUARDRAIL REJECTS THE REQUEST target: Final = self._deployment_hook_target() - hook_request_data: Final = {**request_data, "guardrail_to_apply": self} if target is not self else request_data - result: Final = await target.async_post_call_success_hook( - user_api_key_dict=UserAPIKeyAuth( - user_id=request_data.get("user_api_key_user_id"), - team_id=request_data.get("user_api_key_team_id"), - end_user_id=request_data.get("user_api_key_end_user_id"), - api_key=request_data.get("user_api_key_hash"), - request_route=request_data.get("user_api_key_request_route"), - ), - data=hook_request_data, - response=response, - ) + try: + if target is not self: + request_data["guardrail_to_apply"] = self # rebind-ok: dispatch consumes this key + result: Final = await target.async_post_call_success_hook( + user_api_key_dict=UserAPIKeyAuth( + user_id=request_data.get("user_api_key_user_id"), + team_id=request_data.get("user_api_key_team_id"), + end_user_id=request_data.get("user_api_key_end_user_id"), + api_key=request_data.get("user_api_key_hash"), + request_route=request_data.get("user_api_key_request_route"), + ), + data=request_data, + response=response, + ) + finally: + if target is not self: + request_data.pop("guardrail_to_apply", None) if not self._is_valid_response_type(result): return None diff --git a/tests/test_litellm/integrations/test_custom_guardrail.py b/tests/test_litellm/integrations/test_custom_guardrail.py index cd8d609cf71..ddc8439a83a 100644 --- a/tests/test_litellm/integrations/test_custom_guardrail.py +++ b/tests/test_litellm/integrations/test_custom_guardrail.py @@ -1842,12 +1842,14 @@ class _ApplyStyleGuardrail(CustomGuardrail): self.block = block self.apply_called = False self.seen_texts = None + self.seen_request_data = None async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None): from fastapi import HTTPException self.apply_called = True self.seen_texts = inputs.get("texts") + self.seen_request_data = request_data if self.block: raise HTTPException(status_code=400, detail={"error": "Violated moderation policy"}) return inputs @@ -2646,6 +2648,91 @@ class TestCustomGuardrailPostCallSuccessDeploymentHook: which starved every later callback in litellm.callbacks (notably the lazily-appended VectorStorePreCallHook that attaches provider_specific_fields["search_results"]).""" + @pytest.mark.asyncio + async def test_apply_guardrail_retains_request_identity(self) -> None: + from litellm.types.guardrails import GuardrailEventHooks + from litellm.types.utils import Choices, Message, ModelResponse + + guardrail: Final = _ApplyStyleGuardrail(block=False) + guardrail.event_hook = GuardrailEventHooks.post_call + request_data: Final = {"guardrails": ["apply-style-guardrail"]} + response: Final = ModelResponse(choices=[Choices(message=Message(content="review me"))]) + + await guardrail.async_post_call_success_deployment_hook( + request_data=request_data, response=response, call_type=CallTypes.acompletion + ) + + assert guardrail.seen_request_data is request_data + assert guardrail.seen_texts == ["review me"] + assert "guardrail_to_apply" not in request_data + + @pytest.mark.asyncio + @pytest.mark.parametrize("call_type", (None, CallTypes.acompletion)) + async def test_apply_guardrail_masks_response_and_records_metadata(self, call_type: CallTypes | None) -> None: + from litellm.proxy.guardrails.guardrail_hooks.litellm_content_filter.content_filter import ( + ContentFilterGuardrail, + ) + from litellm.types.guardrails import BlockedWord, ContentFilterAction, GuardrailEventHooks + from litellm.types.utils import Choices, Message, ModelResponse + + guardrail: Final = ContentFilterGuardrail( + guardrail_name="response-filter", + event_hook=GuardrailEventHooks.post_call, + blocked_words=[BlockedWord(keyword="secret", action=ContentFilterAction.MASK)], + ) + request_data: Final = {"guardrails": ["response-filter"]} + response: Final = ModelResponse(choices=[Choices(message=Message(content="a secret"))]) + + result: Final = await guardrail.async_post_call_success_deployment_hook( + request_data=request_data, response=response, call_type=call_type + ) + + assert isinstance(result, ModelResponse) + assert result.choices[0].message.content == f"a {guardrail.keyword_redaction_tag}" + entries: Final = _guardrail_entries(request_data) + assert len(entries) == 1 + assert entries[0]["guardrail_name"] == "response-filter" + assert entries[0]["guardrail_mode"] == "post_call" + assert "guardrail_to_apply" not in request_data + + @pytest.mark.asyncio + @pytest.mark.parametrize("error_type", (None, RuntimeError, asyncio.CancelledError)) + async def test_dispatch_cleans_up_request_on_every_exit(self, error_type: type[BaseException] | None) -> None: + from contextlib import nullcontext + + from litellm.integrations.custom_logger import CustomLogger + from litellm.types.utils import LLMResponseTypes, ModelResponse + + error: Final = error_type("dispatch interrupted") if error_type is not None else None + + class Dispatch(CustomLogger): + request_data: dict[str, object] | None = None + + async def async_post_call_success_hook( + self, data: dict[str, object], user_api_key_dict: UserAPIKeyAuth, response: LLMResponseTypes + ) -> LLMResponseTypes: + self.request_data = data + if error is not None: + raise error + return response + + dispatch: Final = Dispatch() + + class Guardrail(_ApplyStyleGuardrail): + def _deployment_hook_target(self) -> CustomLogger: + return dispatch + + guardrail: Final = Guardrail(block=False) + guardrail.event_hook = GuardrailEventHooks.post_call + request_data: Final = {"guardrails": ["apply-style-guardrail"]} + with pytest.raises(error_type) if error_type is not None else nullcontext(): + await guardrail.async_post_call_success_deployment_hook( + request_data=request_data, response=ModelResponse(), call_type=CallTypes.acompletion + ) + + assert dispatch.request_data is request_data + assert "guardrail_to_apply" not in request_data + @pytest.mark.asyncio async def test_returns_none_when_request_has_no_guardrails(self): from litellm.types.utils import ModelResponse @@ -2740,4 +2827,5 @@ class TestCustomGuardrailPostCallSuccessDeploymentHook: assert result is response assert response.choices[0].message.content == "filtered response" - assert request_data == {"guardrails": ["test-guardrail"]} + assert "guardrail_to_apply" not in request_data + assert len(_guardrail_entries(request_data)) == 1