From 6cdb9959df49132e7402b723ea6aeb79fd01242f Mon Sep 17 00:00:00 2001 From: Mateo Wang <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 29 Oct 2026 08:33:36 -0700 Subject: [PATCH 1/7] feat(router): opt in to non-chat deployment filtering (#46257) * feat(proxy): reject chat routes on models declared non-chat via model_info.supported_endpoints A model whose every deployment sets a non-chat model_info.mode and a supported_endpoints list with no chat endpoint now answers /v1/chat/completions, /v1/messages and /v1/responses with a 400 before any provider call. Models without both fields keep today's behavior. * fix(proxy): let fallbacks answer and cover more chat routes in the non-chat model check The check now steps aside when a request, key, team or router fallback would answer, skips such models in the rate-limit fallback loop, ignores mode and supported_endpoints copied from the cost map by old Admin UI saves, treats a blank mode as unset, and also covers /v1/completions, the Responses WebSocket, Gemini generateContent and /v1beta/interactions * fix(proxy): only step aside for fallbacks that can answer chat, gate the Responses WebSocket after alias resolution, and compare declarations to the loaded catalog * fix(proxy): gate the Responses WebSocket with the router after aliases resolve, and ignore client-set internal fallback flags * fix(proxy): walk fallback groups breadth first and release reservations on a rejected WebSocket Each fallback group is looked up once however often the request repeats it, and a WebSocket closed by the endpoint gate runs the failure hook so rate-limit reservations are released * fix(proxy): walk fallback layers without recursion and gate the compact Responses route The fallback walk now folds one layer at a time, so the recursion check passes, and /v1/responses/compact on a model declared non-chat gets the same 400 as the other chat-shaped routes * fix(proxy): keep chat on rows whose cost map entry serves chat A row that carries the cost map key of an entry that serves chat counts as declaring nothing, so a catalog edit that adds a chat path never turns old Admin UI copies into 400s * fix(proxy): check chat sent to a deployment id or backend model name * fix(proxy): check a model the read-through loads from the database * fix(proxy): step aside for fallbacks the check cannot follow and wildcards that serve a backend name A fallback target with no deployments of its own, a client routing stamp, and a wildcard that does not match the requested backend name no longer turn a working request into a 400 * fix(proxy): ignore a client model_group that is not a string when building rate-limit headers A client metadata.model_group sent as an int or a list crashed the error handler with a 500 once the embedding-only check rejected the call before routing. The header helper now treats a non-string stamp as no model group. Drops the pattern clause in the backend-name lookup, since the group lookup already returns wildcard matches. Adds the third audit round's routing-stamp cells. * fix(proxy): keep a non-string client model_group out of spend rows * fix(proxy): validate standard fallbacks before the typed fallback-group lookup * feat(router): opt in to non-chat deployment filtering * fix(router): skip unresolved local rate-limit fallbacks * fix(router): ignore caller-provided endpoint selector values * fix(router): normalize declared chat endpoint prefixes * test(router): cover preflight error-value behavior * test(router): derive replay metadata defaults from model info * fix(router): preserve fallback aliases and check synchronous text calls * test(routing): rewrite the WebSocket model with a pre-call callback, since main no longer runs guardrails at socket admission * test(routing): keep the endpoint gate's owned proxies on the worker's database * fix(router): keep the endpoint gate on model_info patches that omit it and on same-group stream retries A partial PATCH of model_info reset a stored enforce_supported_endpoints true back to false, because the field defaulted to False and the update merges model_dump(exclude_none=True). The field now defaults to None so an omitted flag keeps the stored value The same-group retry of a stream dropped before content called the provider without the endpoint selection context, so it could land on an enforced embedding-only deployment of the same group. It now runs inside endpoint_selection like every other router attempt * fix(router): probe rate-limit fallbacks through get_routable_deployments so the router surface does not grow Main's router fence lets code outside the router use only listed Router members and router_utils modules. The proxy's fallback probe now calls the listed get_routable_deployments with the request's call type, and the gate's 400 moves to litellm.exceptions as EndpointNotSupportedByModelError. It stays separate from main's UnsupportedEndpointError, which the router retries on the group's other deployments, while the gate refuses the whole group * test(responses): cover the WebSocket error frame for a model that refuses the endpoint * fix(proxy): let the rate-limit fallback probe try a fallback this worker has not loaded yet The probe now skips a fallback only when the endpoint check removed every deployment it has, so a model added on another worker still reaches route_request and its database lookup * fix(proxy): keep key router settings off the Responses WebSocket, read unrooted or single-string endpoint lists, and add status to the refusal frame * fix(router): warn when enforce_supported_endpoints has no usable list and match chat endpoints regardless of case, trailing slash, or Gemini path * test(router): type the load-warning test's caplog and router * test(routing): match the endpoint gate's integration tests to base fallback behavior on unloaded fallbacks and the Responses WebSocket --------- Co-authored-by: Mateo --- litellm/exceptions.py | 11 + litellm/proxy/common_request_processing.py | 30 + litellm/proxy/common_utils/callback_utils.py | 5 +- .../proxy/response_api_endpoints/endpoints.py | 16 +- litellm/proxy/route_llm_request.py | 22 + .../spend_tracking/spend_tracking_utils.py | 8 +- litellm/router.py | 60 +- litellm/router_utils/endpoint_eligibility.py | 120 ++ litellm/types/router.py | 1 + .../llm_nonconversational.yaml | 1 + ..._embedding_only_supported_endpoints_e2e.py | 119 ++ tests/e2e/models.py | 3 +- .../test_non_chat_model_endpoint_gate.py | 1515 +++++++++++++++++ .../proxy/common_utils/test_callback_utils.py | 13 + .../test_model_management_endpoints.py | 27 + .../response_api_endpoints/test_endpoints.py | 16 + .../test_spend_tracking_utils.py | 22 + .../proxy/test_common_request_processing.py | 58 +- tests/unit/proxy/test_route_llm_request.py | 25 + .../router_utils/test_endpoint_eligibility.py | 139 ++ tests/unit/test_router/test_router.py | 113 ++ ui/litellm-dashboard/src/lib/http/schema.d.ts | 2 + 22 files changed, 2300 insertions(+), 26 deletions(-) create mode 100644 litellm/router_utils/endpoint_eligibility.py create mode 100644 tests/e2e/llm_translation/test_embedding_only_supported_endpoints_e2e.py create mode 100644 tests/integration/routing/test_non_chat_model_endpoint_gate.py create mode 100644 tests/unit/router_utils/test_endpoint_eligibility.py diff --git a/litellm/exceptions.py b/litellm/exceptions.py index 61e3a8fb5dc..28211b3dab2 100644 --- a/litellm/exceptions.py +++ b/litellm/exceptions.py @@ -323,6 +323,17 @@ class UnsupportedEndpointError(BadRequestError): pass +class EndpointNotSupportedByModelError(BadRequestError): + def __init__(self, message: str, model: str, llm_provider: str = "") -> None: + super().__init__( # pyright: ignore[reportUnknownMemberType] # inherited public exception constructor is untyped + message=message, model=model, llm_provider=llm_provider + ) + self.message = message + self.type = "invalid_request_error" + self.param = "model" + self.code = "400" + + class UnprocessableEntityError(openai.UnprocessableEntityError): def __init__( self, diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py index fc29a6e3a08..4884e75af4e 100644 --- a/litellm/proxy/common_request_processing.py +++ b/litellm/proxy/common_request_processing.py @@ -129,6 +129,7 @@ from litellm.proxy.guardrails.auto_router_compression import arm_pre_call as _ar from litellm.proxy.hooks.parallel_request_limiter_v3 import get_or_create_request_stash from litellm.proxy.native_compaction import with_proxy_compaction_executor from litellm.proxy.route_llm_request import ( + endpoint_unsupported_by_model, route_request, ) from litellm.proxy.telemetry.request_context import bind_call_id as bind_telemetry_call_id @@ -871,6 +872,17 @@ _DECODED_JSON: Final = TypeAdapter(object) _ROUTER_SETTINGS_ADAPTER: Final[TypeAdapter[Mapping[str, object] | None]] = TypeAdapter(Mapping[str, object] | None) +def _fallback_probe_data( + pristine: object, fallback_model: str, key_aliases: object, group_alias: object +) -> Mapping[str, object]: + return { + **_REQUEST_FIELDS.validate_python(pristine), + "model": resolve_served_model( + fallback_model, _DECODED_JSON.validate_python(litellm.model_alias_map), key_aliases, group_alias + ), + } + + def _request_model(data: Mapping[str, object]) -> str | list[str] | None: try: return _REQUEST_MODEL.validate_python(data.get("model"), strict=True) @@ -2676,6 +2688,17 @@ class ProxyBaseLLMRequestProcessing: original_model, fallback_models, ) + raw_router_settings: Final = _DECODED_JSON.validate_python( + _REQUEST_FIELDS.validate_python(rate_limited_data).get("router_settings_override") + ) + group_alias: Final = _DECODED_JSON.validate_python( + _REQUEST_FIELDS.validate_python(raw_router_settings).get("model_group_alias") + if isinstance(raw_router_settings, Mapping) + else None + ) + key_aliases: Final[object] = cast( # cast-ok: auth aliases are dynamically typed at runtime + object, user_api_key_dict.aliases + ) try: for fallback_model in fallback_models: if fallback_model == original_model: @@ -2691,6 +2714,13 @@ class ProxyBaseLLMRequestProcessing: self.data = independent_snapshot(pristine) self.data["model"] = fallback_model try: + if endpoint_unsupported_by_model( + route_type, + _fallback_probe_data(pristine, fallback_model, key_aliases, group_alias), + llm_router, + user_api_key_dict.team_id, + ): + continue return await self.common_processing_pre_call_logic( request=request, general_settings=general_settings, diff --git a/litellm/proxy/common_utils/callback_utils.py b/litellm/proxy/common_utils/callback_utils.py index f889748c504..d1787f3ff8b 100644 --- a/litellm/proxy/common_utils/callback_utils.py +++ b/litellm/proxy/common_utils/callback_utils.py @@ -436,10 +436,7 @@ def get_model_group_from_litellm_kwargs(kwargs: dict) -> str | None: def get_model_group_from_request_data(data: dict) -> str | None: _metadata: Final = data.get("metadata", None) or {} _model_group: Final = _metadata.get("model_group", None) - if _model_group is not None: - return _model_group - - return None + return _model_group if isinstance(_model_group, str) else None def get_remaining_tokens_and_requests_from_request_data(data: dict) -> dict[str, str]: diff --git a/litellm/proxy/response_api_endpoints/endpoints.py b/litellm/proxy/response_api_endpoints/endpoints.py index 9e642a02d49..772bb3e53d2 100644 --- a/litellm/proxy/response_api_endpoints/endpoints.py +++ b/litellm/proxy/response_api_endpoints/endpoints.py @@ -34,6 +34,7 @@ from litellm._logging import verbose_proxy_logger from litellm.anthropic_interface.exceptions.exception_mapping_utils import AnthropicExceptionMapping from litellm.caching.dual_cache import DualCache from litellm.constants import EMPTY_MAPPING, MANAGED_BACKGROUND_RESPONSE_METADATA_KEY, TERMINAL_RESPONSE_STATUSES +from litellm.exceptions import EndpointNotSupportedByModelError from litellm.integrations.custom_guardrail import ModifyResponseException from litellm.litellm_core_utils.core_helpers import as_str_mapping, get_metadata_variable_name_from_kwargs from litellm.litellm_core_utils.internal_call_metadata import is_background_response, requests_background @@ -2086,6 +2087,14 @@ def _routing_hints_from_first_ws_frame(first_message: str) -> Mapping[str, objec def _responses_ws_failure_frame(failure: Exception) -> str: + if isinstance(failure, EndpointNotSupportedByModelError): + return json.dumps( + { + "type": "error", + "status": 400, + "error": {"type": "invalid_request_error", "message": failure.message.split("\n\n", 1)[0]}, + } + ) proxy_exception: Final = handle_exception_on_proxy(failure) status: Final = int(proxy_exception.code) if str(proxy_exception.code).isdigit() else 500 error_type: Final = ( @@ -2444,7 +2453,12 @@ async def _responses_websocket_session( with contextlib.suppress(Exception): await websocket.send_text(_responses_ws_failure_frame(e)) with contextlib.suppress(Exception): - await websocket.close(code=1011, reason="Internal server error") + await websocket.close( + code=1008 if isinstance(e, EndpointNotSupportedByModelError) else 1011, + reason="Pre-call error" + if isinstance(e, EndpointNotSupportedByModelError) + else "Internal server error", + ) return crash_row return bill_socket_failure(outcome) if isinstance(outcome, Exception) else None diff --git a/litellm/proxy/route_llm_request.py b/litellm/proxy/route_llm_request.py index db1951ec9e2..fea4f342b91 100644 --- a/litellm/proxy/route_llm_request.py +++ b/litellm/proxy/route_llm_request.py @@ -431,6 +431,28 @@ def raise_if_required_body_param_missing( ) +def endpoint_unsupported_by_model( + route_type: str, + data: Mapping[str, object], + llm_router: LitellmRouter | None, + team_id: str | None, +) -> bool: + model: Final = data.get("model") + if not isinstance(model, str) or llm_router is None: + return False + metadata: Final = data.get("metadata") + request_kwargs: Final = { + **data, + "metadata": { + **(metadata if isinstance(metadata, Mapping) else {}), + "user_api_key_team_id": team_id, + }, + } + return bool(llm_router.get_routable_deployments(model, request_kwargs)) and not llm_router.get_routable_deployments( + model, request_kwargs, call_type=route_type + ) + + class MockTestingParamsDisabledError(HTTPException): def __init__(self, params: tuple[str, ...]): super().__init__( diff --git a/litellm/proxy/spend_tracking/spend_tracking_utils.py b/litellm/proxy/spend_tracking/spend_tracking_utils.py index 833ee71ae9c..df9ed611a47 100644 --- a/litellm/proxy/spend_tracking/spend_tracking_utils.py +++ b/litellm/proxy/spend_tracking/spend_tracking_utils.py @@ -464,7 +464,11 @@ def _sl_attribution_fallback( ) -> str: if standard_logging_payload is None: return "" - return standard_logging_payload.get(field) or "" + return _string_or_empty(standard_logging_payload.get(field)) + + +def _string_or_empty(value: object) -> str: + return value if isinstance(value, str) else "" def _deployment_provider(deployment: DeploymentTypedDict) -> str | None: @@ -682,7 +686,7 @@ def get_logging_payload( _model_id: Final = metadata.get("model_info", {}).get("id", "") or _sl_attribution_fallback( standard_logging_payload, "model_id" ) - _model_group: Final = metadata.get("model_group", "") or _sl_attribution_fallback( + _model_group: Final = _string_or_empty(metadata.get("model_group")) or _sl_attribution_fallback( standard_logging_payload, "model_group" ) _api_base: Final = litellm_params.get("api_base", "") or _sl_attribution_fallback( diff --git a/litellm/router.py b/litellm/router.py index ff31f48c49e..31d679f616e 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -287,6 +287,11 @@ from litellm.router_utils.embedding_batch_size_check import ( embedding_deployment_selection, log_rejection_before_provider_call, ) +from litellm.router_utils.endpoint_eligibility import ( + eligible_deployments, + endpoint_selection, + warn_on_unenforced_endpoint_flag, +) from litellm.router_utils.fallback_event_handlers import ( MID_STREAM_FALLBACK_CONTROLS_KEY, AttemptedFallbackTargets, @@ -1449,7 +1454,7 @@ class Router: self.assistants_config = assistants_config self.search_tools = search_tools or [] self.guardrail_list = guardrail_list or [] - self.deployment_names: list = [] # names of models under litellm_params. ex. azure/chatgpt-v-2 + self.deployment_names: list[str] = [] # names of models under litellm_params. ex. azure/chatgpt-v-2 self.deployment_latency_map = {} ### CACHING ### cache_type: Literal["local", "redis", "redis-semantic", "s3", "disk"] = "local" # default to an in-memory cache @@ -6054,12 +6059,13 @@ class Router: kwargs.setdefault("metadata", {}).update({"model_group": model}) # pick the one that is available (lowest TPM/RPM) - deployment: Final = self.get_available_deployment( - model=model, - messages=messages, - specific_deployment=kwargs.pop("specific_deployment", None), - request_kwargs=kwargs, - ) + with endpoint_selection("text_completion", model=model): + deployment: Final = self.get_available_deployment( + model=model, + messages=messages, + specific_deployment=kwargs.pop("specific_deployment", None), + request_kwargs=kwargs, + ) data: Final = deployment["litellm_params"].copy() for k, v in self.default_litellm_params.items(): @@ -6915,6 +6921,7 @@ class Router: """ model_group: Final = cast(str, initial_kwargs.get("model")) # cast-ok: kwargs always carries the model group retry_kwargs: Final = mid_stream_retry_kwargs(initial_kwargs) + handler: Final = retry_kwargs.get("original_generic_function") healthy_deployments, all_deployments = await self._async_get_healthy_deployments( model=model_group, parent_otel_span=get_parent_otel_span_from_kwargs(retry_kwargs), @@ -6945,7 +6952,10 @@ class Router: ) carry_superseded_attempt_usage(retry_kwargs) try: - response = await attempt_function(**retry_kwargs) + with endpoint_selection( + getattr(handler, "__name__", None), getattr(handler, "__module__", None), model=model_group + ): + response = await attempt_function(**retry_kwargs) except Exception as retry_error: # noqa: BLE001 # the fallback chain judges the failed retry wrapped = _stream_fallback_error_for_raised(retry_error, model_group, False, e.llm_provider) if wrapped is None: @@ -9315,9 +9325,18 @@ class Router: Handler for making a call to the .completion()/.embeddings()/etc. functions. """ model_group: Final = kwargs.get("model") - response = original_function(*args, **kwargs) - if coroutine_checker.is_async_callable(response) or inspect.isawaitable(response): - response = await response + generic_function: Final = _OPAQUE_VALUE_ADAPTER.validate_python(kwargs.get("original_generic_function")) + endpoint_function: Final = ( + generic_function if callable(generic_function) else _OPAQUE_VALUE_ADAPTER.validate_python(original_function) + ) + with endpoint_selection( + getattr(endpoint_function, "__name__", None), + getattr(endpoint_function, "__module__", None), + model=_OPAQUE_VALUE_ADAPTER.validate_python(model_group), + ): + response = original_function(*args, **kwargs) + if coroutine_checker.is_async_callable(response) or inspect.isawaitable(response): + response = await response await self.increment_deployment_usage_for_response(response=response, request_kwargs=kwargs) ## PROCESS RESPONSE HEADERS response = await self.set_response_headers(response=response, model_group=model_group, request_kwargs=kwargs) @@ -10638,6 +10657,7 @@ class Router: declared_params: Final = _LITELLM_PARAMS_ADAPTER.validate_python(_litellm_params) warn_on_provider_credential_mismatch(model_name=_model_name, litellm_params=declared_params) warn_on_ignored_model_capability_flags(model_name=_model_name, litellm_params=declared_params) + warn_on_unenforced_endpoint_flag(model_name=_model_name, model_info=_model_info) deployment = without_model_capability_flags( Deployment( **deployment_info, @@ -11441,6 +11461,9 @@ class Router: declared_params: Final = deployment.litellm_params.model_dump(exclude_none=True) warn_on_provider_credential_mismatch(model_name=deployment.model_name, litellm_params=declared_params) warn_on_ignored_model_capability_flags(model_name=deployment.model_name, litellm_params=declared_params) + warn_on_unenforced_endpoint_flag( + model_name=deployment.model_name, model_info=deployment.model_info.model_dump() + ) served: Final = with_admin_api_key_secret_reference(without_model_capability_flags(deployment)) # add to model list @@ -14797,9 +14820,15 @@ class Router: """ return [m for m in self.model_list if m["litellm_params"]["model"] == model] - def get_routable_deployments(self, model: str) -> tuple[DeploymentTypedDict, ...]: + def get_routable_deployments( + self, + model: str, + request_kwargs: Mapping[str, object] = MappingProxyType({}), + call_type: str | None = None, + ) -> tuple[DeploymentTypedDict, ...]: try: - _, candidates = self._common_checks_available_deployment(model=model, request_kwargs={}) # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType] # its signature leaves input and the result untyped + with endpoint_selection(call_type, model=model): + _, candidates = self._common_checks_available_deployment(model=model, request_kwargs=request_kwargs) # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType] # its signature leaves input and the result untyped except (litellm.BadRequestError, ValueError): return () if isinstance(candidates, dict): @@ -15075,7 +15104,10 @@ class Router: model=model, llm_provider="", ) - return selectable + eligible: Final = eligible_deployments(model, selectable) + if isinstance(eligible, tuple): + return list(eligible) + raise eligible def _filter_reserved_deployments( self, diff --git a/litellm/router_utils/endpoint_eligibility.py b/litellm/router_utils/endpoint_eligibility.py new file mode 100644 index 00000000000..4a87e582c97 --- /dev/null +++ b/litellm/router_utils/endpoint_eligibility.py @@ -0,0 +1,120 @@ +from collections.abc import Generator, Mapping, Sequence +from contextlib import contextmanager +from contextvars import ContextVar +from types import MappingProxyType +from typing import Final, TypeVar + +from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError, field_validator + +from litellm._logging import verbose_router_logger +from litellm.exceptions import EndpointNotSupportedByModelError + +CHAT_ENDPOINT_BY_CALL: Final[Mapping[str, str]] = MappingProxyType( + { + "completion": "/v1/chat/completions", + "acompletion": "/v1/chat/completions", + "_completion": "/v1/chat/completions", + "_acompletion": "/v1/chat/completions", + "text_completion": "/v1/completions", + "atext_completion": "/v1/completions", + "_atext_completion": "/v1/completions", + "anthropic_messages": "/v1/messages", + "aresponses": "/v1/responses", + "agenerate_content": "generateContent", + "agenerate_content_stream": "streamGenerateContent", + "acreate_interaction": "/v1beta/interactions", + "litellm.interactions.main.acreate": "/v1beta/interactions", + "litellm.interactions.main.create": "/v1beta/interactions", + "_aresponses_websocket": "/v1/responses", + "acompact_responses": "/v1/responses/compact", + } +) + + +def _endpoint_path(endpoint: str) -> str: + rooted: Final = "/" + endpoint.strip().strip("/").casefold() + path: Final = rooted.removeprefix("/openai").removeprefix("/v1beta").removeprefix("/v1") + return "/" + path.rsplit(":", 1)[-1] if ":" in path else path + + +CHAT_ENDPOINTS: Final = frozenset(_endpoint_path(endpoint) for endpoint in CHAT_ENDPOINT_BY_CALL.values()) +_Deployment: Final = TypeVar("_Deployment", bound=Mapping[str, object]) +_ENFORCE_FLAG: Final = TypeAdapter(bool | None) +_REQUEST_SELECTION: Final[ContextVar[tuple[str | None, str | None]]] = ContextVar( + "router_request_endpoint", default=(None, None) +) + + +class _EndpointDeclaration(BaseModel): + model_config = ConfigDict(frozen=True) + + enforce_supported_endpoints: bool | None = None + supported_endpoints: tuple[str, ...] | None = None + + @field_validator("supported_endpoints", mode="before") + @classmethod + def _one_endpoint_as_a_list(cls, value: object) -> object: + return (value,) if isinstance(value, str) else value + + +@contextmanager +def endpoint_selection(call_type: object, module: object = None, model: object = None) -> Generator[None]: + qualified_call: Final = f"{module}.{call_type}" if module == "litellm.interactions.main" else call_type + endpoint: Final = CHAT_ENDPOINT_BY_CALL.get(qualified_call) if isinstance(qualified_call, str) else None + token: Final = _REQUEST_SELECTION.set((endpoint, model if isinstance(model, str) else None)) + try: + yield + finally: + _REQUEST_SELECTION.reset(token) + + +def _enforced_endpoints(model_info: object) -> tuple[str, ...] | None: + try: + declaration: Final = _EndpointDeclaration.model_validate(model_info) + except ValidationError: + return None + return declaration.supported_endpoints if declaration.enforce_supported_endpoints else None + + +def warn_on_unenforced_endpoint_flag(model_name: str, model_info: Mapping[str, object]) -> str | None: + try: + flagged: Final = _ENFORCE_FLAG.validate_python(model_info.get("enforce_supported_endpoints")) + except ValidationError: + return None + if not flagged or _enforced_endpoints(model_info) is not None: + return None + warning: Final = ( + f"Deployment '{model_name}' sets enforce_supported_endpoints without a supported_endpoints list of " + 'strings, so litellm enforces nothing for it. Set supported_endpoints, e.g. ["/v1/embeddings"].' + ) + verbose_router_logger.warning(warning) + return warning + + +def non_chat_endpoints(deployment: Mapping[str, object]) -> frozenset[str] | None: + declared: Final = _enforced_endpoints(deployment.get("model_info")) + if declared is None: + return None + endpoints: Final = frozenset(declared) + normalized: Final = frozenset(_endpoint_path(endpoint) for endpoint in endpoints) + return endpoints if normalized.isdisjoint(CHAT_ENDPOINTS) else None + + +def eligible_deployments( + model: str, deployments: Sequence[_Deployment] +) -> tuple[_Deployment, ...] | EndpointNotSupportedByModelError: + endpoint, requested_model = _REQUEST_SELECTION.get() + if endpoint is None or not deployments: + return tuple(deployments) + eligible: Final = tuple(deployment for deployment in deployments if non_chat_endpoints(deployment) is None) + if eligible: + return eligible + requested: Final = requested_model or model + supported: Final = ", ".join( + sorted(frozenset[str]().union(*(non_chat_endpoints(deployment) or frozenset() for deployment in deployments))) + ) + return EndpointNotSupportedByModelError( + message=f"{endpoint} is not supported by model={requested}. Supported endpoints: {supported or 'none'}", + model=model, + llm_provider="", + ) diff --git a/litellm/types/router.py b/litellm/types/router.py index db8be9e91a3..56dcad4051a 100644 --- a/litellm/types/router.py +++ b/litellm/types/router.py @@ -270,6 +270,7 @@ def _whole_float_as_int(value: object) -> object: class ModelInfo(MirroredPricingParams): + enforce_supported_endpoints: bool | None = None id: str | None # Allow id to be optional on input, but it will always be present as a str in the model instance db_model: bool = False # used for proxy - to separate models which are stored in the db vs. config. updated_at: datetime.datetime | None = None diff --git a/tests/e2e/coverage_registry/llm_nonconversational.yaml b/tests/e2e/coverage_registry/llm_nonconversational.yaml index 330ddf88fc5..367135eae26 100644 --- a/tests/e2e/coverage_registry/llm_nonconversational.yaml +++ b/tests/e2e/coverage_registry/llm_nonconversational.yaml @@ -2,6 +2,7 @@ - {id: llm.completions.openai.basic.nonstream.works, module: llm, tier: P0, subject_endpoint: completions, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "test_completions_endpoint_e2e.py", rationale: "Legacy text /completions endpoint, second-highest production request volume"} - {id: llm.embeddings.openai.basic.nonstream.works, module: llm, tier: P0, subject_endpoint: embeddings, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "test_embeddings_endpoint_e2e.py:23", rationale: "Core endpoint, live vector response"} - {id: llm.embeddings.openai.input_validation.nonstream.works, module: llm, tier: P1, subject_endpoint: embeddings, route: openai, capability: input_validation, streaming: nonstream, assertions: [works], source: "vendor strategy ยง9.3 / LIT-4778", rationale: "Missing model/input on /embeddings return client errors"} +- {id: llm.embeddings.openai.supported_endpoints_gate.nonstream.works, module: llm, tier: P1, subject_endpoint: embeddings, route: openai, capability: input_validation, streaming: nonstream, assertions: [works], fail_before_fix: proven, source: "test_embedding_only_supported_endpoints_e2e.py", rationale: "Explicit endpoint restrictions preserve real embeddings, refuse chat with a gateway 400, and route chat through a live fallback; zero-provider-call contracts live in integration/routing/test_non_chat_model_endpoint_gate.py"} - {id: llm.embeddings.openai.basic.nonstream.cost_logged, module: llm, tier: P0, subject_endpoint: embeddings, route: openai, capability: basic, streaming: nonstream, assertions: [cost_logged], source: "SPEND_TRACKING_COVERAGE_MATRIX.md:34", rationale: "Cost tracking on embeddings"} - {id: llm.embeddings.openai.passthrough.nonstream.cost_logged, module: llm, tier: P0, subject_endpoint: embeddings, route: openai, capability: basic, streaming: nonstream, assertions: [cost_logged], source: "test_passthrough_e2e.py", rationale: "POST /openai_passthrough/v1/embeddings is costed; the route wrote no spend row at all, so budgets never saw traffic OpenAI was billing for (GitHub issue #36646)"} - {id: llm.embeddings.azure_openai.basic.nonstream.works, module: llm, tier: P0, subject_endpoint: embeddings, route: azure_openai, capability: basic, streaming: nonstream, assertions: [works], source: "llms/azure/azure.py", rationale: "Azure embeddings via translation"} diff --git a/tests/e2e/llm_translation/test_embedding_only_supported_endpoints_e2e.py b/tests/e2e/llm_translation/test_embedding_only_supported_endpoints_e2e.py new file mode 100644 index 00000000000..c45adfeeb84 --- /dev/null +++ b/tests/e2e/llm_translation/test_embedding_only_supported_endpoints_e2e.py @@ -0,0 +1,119 @@ +"""Live provider checks for embeddings, the chat refusal, and chat fallback from an explicitly restricted model.""" + +from __future__ import annotations + +import dataclasses +from typing import Final + +import pytest +from e2e_config import unique_marker +from e2e_metadata import Domain, Mode, Provider, Route, Subject, meta +from lifecycle import ResourceManager +from models import LiteLLMParamsBody, ModelInfoBody, ModelMode, ModelNewBody +from openai import BadRequestError, OpenAI +from proxy_client import ProxyClient +from sdk_clients import NO_PROXY_CACHE, SdkClients + +pytestmark = pytest.mark.e2e + +EMBEDDING_BACKEND: Final = "openai/text-embedding-3-small" +CHAT_FALLBACK_BACKEND: Final = "openai/gpt-5.4-nano" +EMBEDDINGS_ONLY: Final = ("/v1/embeddings",) +PROMPT: Final = "Say this is a test!" + + +def _subject(route: Route) -> Subject: + return Subject( + domain=Domain.LLM_TRANSLATION, + route=route, + providers=(Provider.OPENAI,), + models=(EMBEDDING_BACKEND,), + mode=Mode.NONSTREAM, + ) + + +def _register( + proxy: ProxyClient, + resources: ResourceManager, + prefix: str, + supported_endpoints: tuple[str, ...] | None, + backend: str = EMBEDDING_BACKEND, + mode: ModelMode = "embedding", +) -> str: + model: Final = f"{prefix}-{unique_marker()}" + model_id: Final = proxy.register_model( + ModelNewBody( + model_name=model, + litellm_params=LiteLLMParamsBody(model=backend, api_key="os.environ/OPENAI_API_KEY"), + model_info=ModelInfoBody( + enforce_supported_endpoints=supported_endpoints is not None, + mode=mode, + supported_endpoints=None if supported_endpoints is None else list(supported_endpoints), + ), + ) + ) + resources.defer(lambda: proxy.delete_model(model_id)) + return model + + +def _openai_at(sdk: SdkClients, api_prefix: str, key: str) -> OpenAI: + return dataclasses.replace(sdk, base_url=f"{sdk.base_url}{api_prefix}").openai(key) + + +class TestEmbeddingOnlySupportedEndpoints: + @pytest.mark.covers("llm.embeddings.openai.supported_endpoints_gate.nonstream.works") + @meta(_subject(Route.EMBEDDINGS)) + @pytest.mark.parametrize( + ("prefix", "supported_endpoints"), + ( + pytest.param("e2e-embed-only", EMBEDDINGS_ONLY, id="embed-only"), + pytest.param("e2e-embed-control", None, id="control"), + ), + ) + def test_embeddings_return_a_vector( + self, + proxy: ProxyClient, + resources: ResourceManager, + sdk: SdkClients, + prefix: str, + supported_endpoints: tuple[str, ...] | None, + ) -> None: + model: Final = _register(proxy, resources, prefix, supported_endpoints) + embeddings: Final = sdk.openai(resources.key()).embeddings.create( + model=model, input=PROMPT, extra_body=NO_PROXY_CACHE + ) + assert embeddings.data, f"/embeddings returned no data for {model}: {embeddings!r}" + assert any(component != 0.0 for component in embeddings.data[0].embedding), ( + f"/embeddings returned an empty or all-zero vector for {model}" + ) + + @pytest.mark.covers("llm.embeddings.openai.supported_endpoints_gate.nonstream.works") + @meta(dataclasses.replace(_subject(Route.CHAT_COMPLETIONS), models=(EMBEDDING_BACKEND, CHAT_FALLBACK_BACKEND))) + def test_chat_completions_with_a_fallback_are_answered_by_the_fallback( + self, proxy: ProxyClient, resources: ResourceManager, sdk: SdkClients + ) -> None: + model: Final = _register(proxy, resources, "e2e-embed-only", EMBEDDINGS_ONLY) + fallback: Final = _register(proxy, resources, "e2e-chat-fallback", None, CHAT_FALLBACK_BACKEND, "chat") + completion: Final = _openai_at(sdk, "/v1", resources.key()).chat.completions.create( + model=model, + messages=[{"role": "user", "content": PROMPT}], + extra_body={**NO_PROXY_CACHE, "fallbacks": [fallback]}, + ) + assert completion.choices and completion.choices[0].message.content, ( + f"the fallback {fallback} should answer a chat call to {model}, got {completion!r}" + ) + + @pytest.mark.covers("llm.embeddings.openai.supported_endpoints_gate.nonstream.works") + @meta(_subject(Route.CHAT_COMPLETIONS)) + def test_chat_completions_without_a_fallback_are_refused_by_the_gateway( + self, proxy: ProxyClient, resources: ResourceManager, sdk: SdkClients + ) -> None: + model: Final = _register(proxy, resources, "e2e-embed-only", EMBEDDINGS_ONLY) + with pytest.raises(BadRequestError) as refused: + _ = _openai_at(sdk, "/v1", resources.key()).chat.completions.create( + model=model, messages=[{"role": "user", "content": PROMPT}], extra_body=NO_PROXY_CACHE + ) + assert ( + f"/v1/chat/completions is not supported by model={model}. Supported endpoints: /v1/embeddings" + in refused.value.message + ), f"chat on {model} should be refused by the gateway before any provider call, got {refused.value!r}" diff --git a/tests/e2e/models.py b/tests/e2e/models.py index afddc7e25d6..504adec5401 100644 --- a/tests/e2e/models.py +++ b/tests/e2e/models.py @@ -1416,7 +1416,7 @@ class LiteLLMParamsBody(BaseModel): order: int | None = None -ModelMode = Literal["chat", "batch", "realtime", "image_generation"] +ModelMode = Literal["chat", "embedding", "batch", "realtime", "image_generation"] class TokenCountModelBody(BaseModel, frozen=True): @@ -1425,6 +1425,7 @@ class TokenCountModelBody(BaseModel, frozen=True): class ModelInfoBody(BaseModel): + enforce_supported_endpoints: bool | None = None # id is left unset so the proxy assigns a unique model_id per deployment. # Pinning it to the model_name made re-registrations of a fixed-name model # (e.g. the batch suite's openai-batch) collide on the model_id unique diff --git a/tests/integration/routing/test_non_chat_model_endpoint_gate.py b/tests/integration/routing/test_non_chat_model_endpoint_gate.py new file mode 100644 index 00000000000..465f5801692 --- /dev/null +++ b/tests/integration/routing/test_non_chat_model_endpoint_gate.py @@ -0,0 +1,1515 @@ +import asyncio +import contextlib +import json +import uuid +from collections.abc import Iterator, Mapping +from dataclasses import dataclass +from pathlib import Path +from types import MappingProxyType +from typing import Final + +import httpx +import openai +import pytest +import websockets +import yaml +from integration._support.client import ( + JSON_OBJECT, + Gateway, + Scenario, + eventually, + gateway_from_environment, + held, + list_value, + object_value, + string_value, +) +from integration._support import responses_ws_peer +from integration._support.database import read_rows +from integration._support.openai_wire import answering_model_discovery, chat_reply, responses_reply +from integration._support.process import OwnedState, graceful_stop_seconds, owned_proxy_process +from integration._support.tls import write_self_signed_cert +from integration._support.wire import Reply, Request, Wire, wire_server +from pydantic import JsonValue, TypeAdapter +from websockets.exceptions import ConnectionClosed + +_PROVIDER_KEY: Final = "integration-provider-key" +_EMBEDDING_MODEL: Final = "text-embedding-3-small" +_CHAT_MODEL: Final = "gpt-4o-mini" +_CHAT_BACKEND_PREFIX: Final = "gate-chat-backend-" +_SERVED_TEXT: Final = "served by the chat upstream" +_VECTOR: Final = (0.25, 0.5, 0.75) +_COUNTED_TOKENS: Final = 17 +_CLOSE: Final = MappingProxyType({"connection": "close"}) +_RELOAD_SECONDS: Final = 5 +_EMBED_ONLY: Final[Mapping[str, JsonValue]] = MappingProxyType( + {"mode": "embedding", "supported_endpoints": ["/v1/embeddings"], "enforce_supported_endpoints": True} +) + + +def _embedding_reply() -> Reply: + return Reply( + body=json.dumps( + { + "object": "list", + "data": [{"object": "embedding", "index": 0, "embedding": list(_VECTOR)}], + "model": _EMBEDDING_MODEL, + "usage": {"prompt_tokens": 3, "total_tokens": 3}, + } + ).encode() + ) + + +def _not_a_chat_model() -> Reply: + return Reply( + status=404, + body=json.dumps( + {"error": {"message": "scripted: not a chat model", "type": "invalid_request_error", "code": None}} + ).encode(), + ) + + +def _text_completion_reply() -> Reply: + return Reply( + body=json.dumps( + { + "id": f"cmpl-{uuid.uuid4().hex}", + "object": "text_completion", + "created": 1, + "model": _CHAT_MODEL, + "choices": [{"index": 0, "text": _SERVED_TEXT, "finish_reason": "stop", "logprobs": None}], + "usage": {"prompt_tokens": 3, "completion_tokens": 3, "total_tokens": 6}, + } + ).encode() + ) + + +def _serves_chat(model: JsonValue) -> bool: + return model == _CHAT_MODEL or (isinstance(model, str) and model.startswith(_CHAT_BACKEND_PREFIX)) + + +def _respond(request: Request) -> Reply: + path: Final = request.target.split("?")[0] + body: Final = JSON_OBJECT.validate_json(request.body or b"{}") + streamed: Final = body.get("stream") is True + if path == "/v1/embeddings": + return _embedding_reply() + if path == "/v1/responses/input_tokens": + return Reply(body=json.dumps({"object": "response.input_tokens", "input_tokens": _COUNTED_TOKENS}).encode()) + if not _serves_chat(body.get("model")): + return _not_a_chat_model() + if path == "/v1/chat/completions": + return chat_reply(f"chatcmpl-{uuid.uuid4().hex}", _CHAT_MODEL, _SERVED_TEXT, stream=streamed) + if path in ("/v1/responses", "/v1/responses/compact"): + return responses_reply(f"resp_{uuid.uuid4().hex}", _CHAT_MODEL, _SERVED_TEXT, stream=streamed) + if path == "/v1/completions": + return _text_completion_reply() + return _not_a_chat_model() + + +class _Upstream: + def __init__(self, wire: Wire) -> None: + self.wire = wire + self.seen: tuple[Request, ...] = () + + def calls(self, marker: str) -> tuple[Request, ...]: + self.seen = (*self.seen, *self.wire.drain()) + return tuple(request for request in self.seen if marker.encode() in request.body or marker in request.target) + + def targets(self, marker: str) -> tuple[str, ...]: + return tuple(request.target for request in self.calls(marker)) + + +@dataclass(frozen=True, slots=True) +class _Deployment: + litellm_model: str + declarations: tuple[Mapping[str, JsonValue], ...] + + +def _single(declaration: Mapping[str, JsonValue], litellm_model: str = f"openai/{_CHAT_MODEL}") -> _Deployment: + return _Deployment(litellm_model, (declaration,)) + + +_SHARED: Final[Mapping[str, _Deployment]] = MappingProxyType( + { + "embedding_only": _single(_EMBED_ONLY, f"openai/{_EMBEDDING_MODEL}"), + "embedding_only_second": _single(_EMBED_ONLY, f"openai/{_EMBEDDING_MODEL}"), + "image_generation": _single( + {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/images/generations"]} + ), + "chat_target": _single({"mode": "chat"}), + "chat_rate_limited": _single({"mode": "chat"}), + "mode_not_a_string": _single( + {"mode": 7, "supported_endpoints": ["/v1/embeddings"]}, "openai/gate-mode-not-a-string" + ), + } +) + + +@dataclass(frozen=True, slots=True) +class _Rig: + gateway: Gateway + upstream: _Upstream + names: Mapping[str, str] + api_base: str + + +def _create(gateway: Gateway, name: str, litellm_params: Mapping[str, JsonValue], info: Mapping[str, JsonValue]) -> str: + created: Final = gateway.post( + "/model/new", + { + "model_name": name, + "litellm_params": {"api_key": _PROVIDER_KEY, **litellm_params}, + "model_info": dict(info), + }, + ) + return string_value(object_value(created["model_info"])["id"]) + + +def _deploy(scenario: Scenario, api_base: str, label: str, deployment: _Deployment) -> tuple[str, tuple[str, ...]]: + name: Final = f"gate-{label}-{uuid.uuid4().hex[:12]}" + identities: Final = tuple( + _create(scenario.gateway, name, {"model": deployment.litellm_model, "api_base": api_base}, declaration) + for declaration in deployment.declarations + ) + tuple(scenario.cleanups.callback(scenario.delete_model, identity) for identity in identities) + return name, identities + + +def _listed_ids(gateway: Gateway) -> frozenset[str]: + listed: Final = gateway.request("GET", "/model/info", headers=_CLOSE) + assert listed.status_code == 200, listed.text + return frozenset( + string_value(object_value(object_value(entry)["model_info"])["id"]) + for entry in list_value(JSON_OBJECT.validate_json(listed.content)["data"]) + ) + + +def _served_by_every_worker(gateway: Gateway, identities: frozenset[str]) -> None: + held( + lambda: identities <= _listed_ids(gateway), + lambda served: served, + holding=_RELOAD_SECONDS * 2, + seconds=_RELOAD_SECONDS * 12, + ) + + +@pytest.fixture(scope="module") +def rig() -> Iterator[_Rig]: + with ( + wire_server(answering_model_discovery(_respond)) as wire, + gateway_from_environment() as gateway, + gateway.scenario() as scenario, + ): + api_base: Final = wire.url + "/v1" + deployed: Final = {label: _deploy(scenario, api_base, label, spec) for label, spec in _SHARED.items()} + _served_by_every_worker( + gateway, frozenset(identity for _name, identities in deployed.values() for identity in identities) + ) + yield _Rig( + gateway, + _Upstream(wire), + MappingProxyType({label: name for label, (name, _ids) in deployed.items()}), + api_base, + ) + + +def _marker() -> str: + return f"gate-{uuid.uuid4().hex}" + + +def _openai_error(endpoint: str, model: str, supported: str) -> dict[str, JsonValue]: + return { + "error": { + "message": f"{endpoint} is not supported by model={model}. Supported endpoints: {supported}", + "type": "invalid_request_error", + "param": "model", + "code": "400", + } + } + + +def _anthropic_error(model: str, supported: str) -> dict[str, JsonValue]: + return { + "type": "error", + "error": { + "type": "invalid_request_error", + "message": f"/v1/messages is not supported by model={model}. Supported endpoints: {supported}", + }, + } + + +def _chat_body(model: str, marker: str, *, stream: bool = False) -> dict[str, JsonValue]: + return {"model": model, "messages": [{"role": "user", "content": marker}], **({"stream": True} if stream else {})} + + +def _failure_row(call_id: str) -> dict[str, JsonValue]: + rows: Final = eventually( + lambda: read_rows('SELECT request_id, status FROM "LiteLLM_SpendLogs" WHERE request_id = %s', (call_id,)), + lambda found: len(found) >= 1, + seconds=70, + ) + assert len(rows) == 1, rows + return rows[0] + + +def _assert_error(response: httpx.Response, expected: dict[str, JsonValue]) -> None: + assert response.status_code == 400, response.text + body: Final = JSON_OBJECT.validate_json(response.content) + error: Final = object_value(body["error"]) + message: Final = string_value(error["message"]).split("\n\n", 1)[0] + assert {**body, "error": {**error, "message": message}} == expected, response.text + + +def _assert_rejected_before( + upstream: _Upstream, response: httpx.Response, marker: str, expected: dict[str, JsonValue] +) -> None: + _assert_error(response, expected) + assert upstream.calls(marker) == (), upstream.targets(marker) + assert _failure_row(response.headers["x-litellm-call-id"])["status"] == "failure" + + +def _assert_rejected(rig: _Rig, response: httpx.Response, marker: str, expected: dict[str, JsonValue]) -> None: + _assert_rejected_before(rig.upstream, response, marker, expected) + + +@dataclass(frozen=True, slots=True) +class _Route: + path: str + endpoint: str + anthropic_shape: bool + stream: bool + model_in_path: bool + shape: str + + +_ROUTES: Final[Mapping[str, _Route]] = MappingProxyType( + { + "interactions": _Route("/v1beta/interactions", "/v1beta/interactions", False, False, False, "interactions"), + "chat_v1": _Route("/v1/chat/completions", "/v1/chat/completions", False, False, False, "chat"), + "chat_v1_stream": _Route("/v1/chat/completions", "/v1/chat/completions", False, True, False, "chat"), + "chat_bare": _Route("/chat/completions", "/v1/chat/completions", False, False, False, "chat"), + "chat_engines": _Route("/engines/{model}/chat/completions", "/v1/chat/completions", False, False, True, "chat"), + "chat_azure_deployments": _Route( + "/openai/deployments/{model}/chat/completions", "/v1/chat/completions", False, False, True, "chat" + ), + "messages": _Route("/v1/messages", "/v1/messages", True, False, False, "messages"), + "messages_stream": _Route("/v1/messages", "/v1/messages", True, True, False, "messages"), + "responses_v1": _Route("/v1/responses", "/v1/responses", False, False, False, "responses"), + "responses_v1_stream": _Route("/v1/responses", "/v1/responses", False, True, False, "responses"), + "responses_bare": _Route("/responses", "/v1/responses", False, False, False, "responses"), + "cursor_responses_body": _Route("/cursor/chat/completions", "/v1/responses", False, False, False, "cursor"), + "cursor_chat_body": _Route("/cursor/chat/completions", "/v1/chat/completions", False, False, False, "chat"), + } +) + + +def _route_body(route: _Route, model: str, marker: str) -> dict[str, JsonValue]: + shaped: Final[Mapping[str, dict[str, JsonValue]]] = { + "chat": {"messages": [{"role": "user", "content": marker}]}, + "messages": {"max_tokens": 16, "messages": [{"role": "user", "content": marker}]}, + "responses": {"input": marker}, + "interactions": {"input": marker}, + "cursor": {"input": [{"role": "user", "content": marker}]}, + } + return { + **({} if route.model_in_path else {"model": model}), + **shaped[route.shape], + **({"stream": True} if route.stream else {}), + } + + +def _send(gateway: Gateway, route: _Route, model: str, marker: str) -> httpx.Response: + path: Final = route.path.format(model=model) + body: Final = _route_body(route, model, marker) + if not route.stream: + return gateway.request("POST", path, body) + with gateway.client.stream("POST", path, json=body, headers={"Authorization": f"Bearer {gateway.key}"}) as streamed: + streamed.read() + return streamed + + +@pytest.mark.parametrize("route", tuple(_ROUTES)) +def test_chat_shaped_route_rejects_an_embedding_only_model_before_the_upstream(rig: _Rig, route: str) -> None: + spec: Final = _ROUTES[route] + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = _send(rig.gateway, spec, model, marker) + expected: Final = ( + _anthropic_error(model, "/v1/embeddings") + if spec.anthropic_shape + else _openai_error(spec.endpoint, model, "/v1/embeddings") + ) + _assert_rejected(rig, response, marker, expected) + + +def _sdk_embedding(rig: _Rig, model: str, marker: str) -> tuple[tuple[float, ...], str]: + async def call() -> tuple[tuple[float, ...], str]: + async with openai.AsyncOpenAI( + base_url=f"{rig.gateway.client.base_url}/v1", api_key=rig.gateway.key, max_retries=0 + ) as sdk: + raw: Final = await sdk.embeddings.with_raw_response.create( + model=model, input=marker, encoding_format="float" + ) + parsed: Final = raw.parse() + return tuple(parsed.data[0].embedding), raw.headers["x-litellm-call-id"] + + return asyncio.run(call()) + + +def _httpx_embedding(rig: _Rig, model: str, marker: str) -> tuple[tuple[float, ...], str]: + response: Final = rig.gateway.request( + "POST", "/v1/embeddings", {"model": model, "input": marker, "encoding_format": "float"} + ) + assert response.status_code == 200, response.text + data: Final = list_value(JSON_OBJECT.validate_json(response.content)["data"]) + vector: Final = list_value(object_value(data[0])["embedding"]) + return tuple(float(str(value)) for value in vector), response.headers["x-litellm-call-id"] + + +@pytest.mark.parametrize("client", ("httpx_sync", "openai_async_sdk")) +def test_embeddings_still_reach_the_embedding_only_model(rig: _Rig, client: str) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + vector, call_id = ( + _httpx_embedding(rig, model, marker) if client == "httpx_sync" else _sdk_embedding(rig, model, marker) + ) + assert vector == _VECTOR + assert rig.upstream.targets(marker) == ("/v1/embeddings",) + row: Final = eventually( + lambda: read_rows( + 'SELECT status, call_type FROM "LiteLLM_SpendLogs" WHERE request_id = %s OR litellm_call_id = %s', + (call_id, call_id), + ), + lambda found: len(found) >= 1, + seconds=70, + ) + assert [(entry["status"], entry["call_type"]) for entry in row] == [("success", "aembedding")], row + + +def _assert_served(rig: _Rig, response: httpx.Response, marker: str, target: str) -> None: + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert rig.upstream.targets(marker) == (target,) + + +def test_key_alias_to_an_embedding_only_model_is_rejected_under_the_resolved_name(rig: _Rig) -> None: + target: Final = rig.names["embedding_only"] + alias: Final = f"gate-alias-{uuid.uuid4().hex[:12]}" + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(aliases={alias: target}) + response: Final = rig.gateway.request("POST", "/v1/chat/completions", _chat_body(alias, marker), key=key) + _assert_rejected(rig, response, marker, _openai_error("/v1/chat/completions", target, "/v1/embeddings")) + + +async def _websocket_first_frame_and_close( + gateway: Gateway, key: str, model: str, marker: str +) -> tuple[dict[str, JsonValue], int | None, str | None]: + url: Final = str(gateway.client.base_url).rstrip("/").replace("http://", "ws://", 1) + "/v1/responses" + async with websockets.connect(url, additional_headers={"Authorization": f"Bearer {key}"}, open_timeout=10) as ws: + await ws.send(json.dumps({"type": "response.create", "model": model, "input": marker})) + first: Final = JSON_OBJECT.validate_json(await asyncio.wait_for(ws.recv(), 30)) + with contextlib.suppress(ConnectionClosed): + await asyncio.wait_for(ws.recv(), 30) + return first, ws.close_code, ws.close_reason + + +def _websocket_rejection(model: str) -> tuple[dict[str, JsonValue], int | None, str | None]: + return ( + { + "type": "error", + "status": 400, + "error": { + "type": "invalid_request_error", + "message": f"/v1/responses is not supported by model={model}. Supported endpoints: /v1/embeddings", + }, + }, + 1008, + "Pre-call error", + ) + + +@pytest.mark.parametrize("aliased", [False, True]) +def test_responses_websocket_closes_on_an_embedding_only_model_before_the_upstream(rig: _Rig, aliased: bool) -> None: + target: Final = rig.names["embedding_only"] + alias: Final = f"gate-ws-alias-{uuid.uuid4().hex[:12]}" + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(aliases={alias: target}) + closed: Final = asyncio.run( + _websocket_first_frame_and_close(rig.gateway, key, alias if aliased else target, marker) + ) + assert closed == _websocket_rejection(target) + assert rig.upstream.calls(marker) == () + + +_RESERVING_KEYS: Final[Mapping[str, Mapping[str, JsonValue]]] = MappingProxyType( + { + "max_parallel_requests": {"max_parallel_requests": 1}, + "tpm_limit": {"tpm_limit": 100, "default_estimated_output_tokens": 80}, + } +) + + +@pytest.mark.parametrize("limit", tuple(_RESERVING_KEYS)) +def test_responses_websocket_rejections_leave_the_key_reservation_free(rig: _Rig, limit: str) -> None: + target: Final = rig.names["embedding_only"] + markers: Final = tuple(_marker() for _ in range(3)) + served_marker: Final = _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(**_RESERVING_KEYS[limit]) + closed: Final = tuple( + asyncio.run(_websocket_first_frame_and_close(rig.gateway, key, target, marker)) for marker in markers + ) + assert closed == (_websocket_rejection(target),) * 3 + served: Final = rig.gateway.request( + "POST", "/v1/chat/completions", _chat_body(rig.names["chat_target"], served_marker), key=key + ) + _assert_served(rig, served, served_marker, "/v1/chat/completions") + assert [rig.upstream.calls(marker) for marker in markers] == [()] * 3 + + +def test_team_scoped_embedding_only_model_is_rejected_under_its_public_name(rig: _Rig) -> None: + public: Final = f"gate-team-{uuid.uuid4().hex[:12]}" + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + team: Final = scenario.team() + identity: Final = _create( + rig.gateway, + public, + {"model": f"openai/{_EMBEDDING_MODEL}", "api_base": rig.api_base}, + {**_EMBED_ONLY, "team_id": team}, + ) + scenario.cleanups.callback(scenario.delete_model, identity) + key: Final = scenario.key(team_id=team) + _served_by_every_worker(rig.gateway, frozenset({identity})) + embedded: Final = rig.gateway.request( + "POST", "/v1/embeddings", {"model": public, "input": marker, "encoding_format": "float"}, key=key + ) + assert embedded.status_code == 200, embedded.text + assert rig.upstream.targets(marker) == ("/v1/embeddings",) + rejected_marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", "/v1/chat/completions", _chat_body(public, rejected_marker), key=key + ) + _assert_rejected( + rig, response, rejected_marker, _openai_error("/v1/chat/completions", public, "/v1/embeddings") + ) + + +def test_wildcard_embedding_only_deployment_rejects_chat_for_any_matched_model(rig: _Rig) -> None: + prefix: Final = f"gate-wild-{uuid.uuid4().hex[:12]}" + requested: Final = f"{prefix}/{_EMBEDDING_MODEL}" + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + identity: Final = _create( + rig.gateway, f"{prefix}/*", {"model": "openai/*", "api_base": rig.api_base}, _EMBED_ONLY + ) + scenario.cleanups.callback(scenario.delete_model, identity) + _served_by_every_worker(rig.gateway, frozenset({identity})) + response: Final = rig.gateway.request("POST", "/v1/chat/completions", _chat_body(requested, marker)) + _assert_rejected(rig, response, marker, _openai_error("/v1/chat/completions", requested, "/v1/embeddings")) + + +def test_key_without_access_gets_the_access_error_and_learns_nothing_about_endpoints(rig: _Rig) -> None: + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(models=[rig.names["chat_target"]]) + response: Final = rig.gateway.request( + "POST", "/v1/chat/completions", _chat_body(rig.names["embedding_only"], marker), key=key + ) + assert response.status_code == 403, response.text + assert object_value(JSON_OBJECT.validate_json(response.content)["error"])["type"] == "key_model_access_denied", ( + response.text + ) + assert "Supported endpoints" not in response.text, response.text + assert rig.upstream.calls(marker) == () + + +def test_unknown_model_keeps_the_invalid_model_error(rig: _Rig) -> None: + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", "/v1/chat/completions", _chat_body(f"gate-missing-{uuid.uuid4().hex}", marker) + ) + assert response.status_code == 400, response.text + assert "Invalid model name" in response.text and "Supported endpoints" not in response.text, response.text + assert rig.upstream.calls(marker) == () + + +def test_request_fallbacks_answer_for_an_embedding_only_primary(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", "/v1/chat/completions", {**_chat_body(model, marker), "fallbacks": [rig.names["chat_target"]]} + ) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert rig.upstream.targets(marker) == ("/v1/chat/completions",) + + +def test_an_empty_request_fallback_list_keeps_the_rejection(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", "/v1/chat/completions", {**_chat_body(model, marker), "fallbacks": []} + ) + _assert_rejected(rig, response, marker, _openai_error("/v1/chat/completions", model, "/v1/embeddings")) + + +def test_fallbacks_that_cannot_answer_chat_keep_the_rejection(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + fallbacks: Final = [rig.names["image_generation"], model] + response: Final = rig.gateway.request( + "POST", "/v1/chat/completions", {**_chat_body(model, marker), "fallbacks": fallbacks} + ) + _assert_rejected(rig, response, marker, _openai_error("/v1/chat/completions", model, "/v1/embeddings")) + + +def test_request_fallbacks_do_not_rescue_an_unknown_primary(rig: _Rig) -> None: + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", + "/v1/chat/completions", + {**_chat_body(f"gate-missing-{uuid.uuid4().hex}", marker), "fallbacks": [rig.names["chat_target"]]}, + ) + assert response.status_code == 400, response.text + assert "Invalid model name" in response.text, response.text + assert rig.upstream.calls(marker) == () + + +def _chat(rig: _Rig, model: str, marker: str, *, key: str | None = None, **extra: JsonValue) -> httpx.Response: + return rig.gateway.request("POST", "/v1/chat/completions", {**_chat_body(model, marker), **extra}, key=key) + + +def _assert_rejected_chat(rig: _Rig, response: httpx.Response, marker: str, model: str) -> None: + _assert_rejected(rig, response, marker, _openai_error("/v1/chat/completions", model, "/v1/embeddings")) + + +def test_a_fallback_chain_that_reaches_chat_two_hops_away_answers(rig: _Rig) -> None: + first, second, chat = rig.names["embedding_only"], rig.names["embedding_only_second"], rig.names["chat_target"] + marker: Final = _marker() + response: Final = _chat(rig, first, marker, fallbacks=[{first: [second]}, {second: [chat]}]) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert rig.upstream.targets(marker) == ("/v1/chat/completions",) + + +def test_a_fallback_chain_that_loops_between_embedding_only_groups_keeps_the_rejection(rig: _Rig) -> None: + first, second = rig.names["embedding_only"], rig.names["embedding_only_second"] + marker: Final = _marker() + response: Final = _chat(rig, first, marker, fallbacks=[{first: [second]}, {second: [first]}]) + _assert_rejected_chat(rig, response, marker, first) + + +def test_a_specific_chain_that_loops_beats_a_catch_all_that_could_answer(rig: _Rig) -> None: + first, second = rig.names["embedding_only"], rig.names["embedding_only_second"] + marker: Final = _marker() + response: Final = _chat( + rig, first, marker, fallbacks=[{"*": [rig.names["chat_target"]]}, {first: [second]}, {second: [first]}] + ) + _assert_rejected_chat(rig, response, marker, first) + + +def test_disable_fallbacks_keeps_the_rejection_over_request_fallbacks(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = _chat(rig, model, marker, fallbacks=[rig.names["chat_target"]], disable_fallbacks=True) + _assert_rejected_chat(rig, response, marker, model) + + +def _assert_fallback_answered( + upstream: _Upstream, response: httpx.Response, marker: str, target: str = "/v1/chat/completions" +) -> None: + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert upstream.targets(marker) == (target,), response.text + + +def _through_a_group_with_no_deployments(rig: _Rig) -> Mapping[str, JsonValue]: + ghost: Final = f"gate-ghost-{uuid.uuid4().hex}" + return {"fallbacks": [{rig.names["embedding_only"]: [ghost]}, {ghost: [rig.names["chat_target"]]}]} + + +def test_request_fallbacks_naming_a_chat_backend_name_answer_for_an_embedding_only_primary(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + backend: Final = f"openai/{_CHAT_BACKEND_PREFIX}{uuid.uuid4().hex[:12]}" + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + identity: Final = _create( + rig.gateway, f"gate-backend-{uuid.uuid4().hex[:12]}", {"model": backend, "api_base": rig.api_base}, {} + ) + scenario.cleanups.callback(scenario.delete_model, identity) + _served_by_every_worker(rig.gateway, frozenset({identity})) + response: Final = _chat(rig, model, marker, fallbacks=[{model: [backend]}]) + _assert_fallback_answered(rig.upstream, response, marker) + + +@pytest.mark.parametrize("level", ("key", "team")) +def test_router_settings_fallbacks_through_a_group_with_no_deployments_answer(rig: _Rig, level: str) -> None: + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + settings: Final = {"router_settings": dict(_through_a_group_with_no_deployments(rig))} + key: Final = scenario.key(**settings) if level == "key" else scenario.key(team_id=scenario.team(**settings)) + response: Final = _chat(rig, rig.names["embedding_only"], marker, key=key) + _assert_fallback_answered(rig.upstream, response, marker) + + +def test_team_router_settings_fallbacks_answer_for_an_embedding_only_primary(rig: _Rig) -> None: + model, chat = rig.names["embedding_only"], rig.names["chat_target"] + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + team: Final = scenario.team(router_settings={"fallbacks": [{model: [chat]}]}) + response: Final = _chat(rig, model, marker, key=scenario.key(team_id=team)) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert rig.upstream.targets(marker) == ("/v1/chat/completions",) + + +def test_key_router_settings_fallbacks_answer_for_an_embedding_only_primary(rig: _Rig) -> None: + model, chat = rig.names["embedding_only"], rig.names["chat_target"] + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(router_settings={"fallbacks": [{model: [chat]}]}) + response: Final = _chat(rig, model, marker, key=key) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert rig.upstream.targets(marker) == ("/v1/chat/completions",) + + +def test_key_router_settings_that_cannot_answer_override_team_fallbacks_that_could(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + team: Final = scenario.team(router_settings={"fallbacks": [{model: [rig.names["chat_target"]]}]}) + key: Final = scenario.key( + team_id=team, router_settings={"fallbacks": [{model: [rig.names["embedding_only_second"]]}]} + ) + response: Final = _chat(rig, model, marker, key=key) + _assert_rejected_chat(rig, response, marker, model) + + +def test_an_empty_request_fallback_list_overrides_key_router_settings_fallbacks(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(router_settings={"fallbacks": [{model: [rig.names["chat_target"]]}]}) + response: Final = _chat(rig, model, marker, key=key, fallbacks=[]) + _assert_rejected_chat(rig, response, marker, model) + + +def test_local_rate_limit_fallback_skips_an_embedding_only_hop(rig: _Rig) -> None: + limited, chat, model = rig.names["chat_rate_limited"], rig.names["chat_target"], rig.names["embedding_only"] + first, second = _marker(), _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key( + model_rpm_limit={limited: 1}, router_settings={"fallbacks": [{limited: [model, chat]}]} + ) + served_first: Final = _chat(rig, limited, first, key=key) + served_second: Final = _chat(rig, limited, second, key=key) + assert (served_first.status_code, served_first.headers.get("x-litellm-model-group")) == (200, limited), ( + served_first.text + ) + assert (served_second.status_code, served_second.headers.get("x-litellm-model-group")) == (200, chat), ( + served_second.text + ) + assert (rig.upstream.targets(first), rig.upstream.targets(second)) == ( + ("/v1/chat/completions",), + ("/v1/chat/completions",), + ) + + +_COMPACT_PATHS: Final = ("/v1/responses/compact", "/responses/compact", "/openai/v1/responses/compact") + + +def _compact_body(model: str, marker: str) -> dict[str, JsonValue]: + return {"model": model, "input": [{"role": "user", "content": marker}]} + + +@pytest.mark.parametrize("path", _COMPACT_PATHS) +def test_responses_compact_rejects_an_embedding_only_model_before_the_upstream(rig: _Rig, path: str) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = rig.gateway.request("POST", path, _compact_body(model, marker)) + _assert_rejected(rig, response, marker, _openai_error("/v1/responses/compact", model, "/v1/embeddings")) + + +def test_responses_compact_keeps_reaching_the_upstream_for_a_chat_model(rig: _Rig) -> None: + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", "/v1/responses/compact", _compact_body(rig.names["chat_target"], marker) + ) + _assert_served(rig, response, marker, "/v1/responses/compact") + + +def test_count_tokens_keeps_counting_for_an_embedding_only_model(rig: _Rig) -> None: + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", + "/v1/messages/count_tokens", + {"model": rig.names["embedding_only"], "messages": [{"role": "user", "content": marker}]}, + ) + assert response.status_code == 200, response.text + assert JSON_OBJECT.validate_json(response.content)["input_tokens"] == _COUNTED_TOKENS, response.text + assert rig.upstream.targets(marker) == ("/v1/responses/input_tokens",) + + +def test_text_completions_reject_an_embedding_only_model_before_the_upstream(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = rig.gateway.request("POST", "/v1/completions", {"model": model, "prompt": marker}) + _assert_rejected(rig, response, marker, _openai_error("/v1/completions", model, "/v1/embeddings")) + + +@pytest.mark.parametrize("method", ("generateContent", "streamGenerateContent")) +def test_gemini_routes_reject_an_embedding_only_model_before_the_upstream(rig: _Rig, method: str) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", f"/v1beta/models/{model}:{method}", {"contents": [{"role": "user", "parts": [{"text": marker}]}]} + ) + assert response.status_code == 400, response.text + assert f"{method} is not supported by model={model}. Supported endpoints: /v1/embeddings" in response.text + assert rig.upstream.calls(marker) == (), rig.upstream.targets(marker) + + +def test_repeated_rejections_log_one_failure_row_per_call(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + responses: Final = tuple( + rig.gateway.request("POST", "/v1/chat/completions", _chat_body(model, marker)) for _ in range(3) + ) + expected: Final = _openai_error("/v1/chat/completions", model, "/v1/embeddings") + for response in responses: + _assert_error(response, expected) + call_ids: Final = tuple(item.headers["x-litellm-call-id"] for item in responses) + assert len(set(call_ids)) == 3, call_ids + assert [_failure_row(call_id)["status"] for call_id in call_ids] == ["failure"] * 3 + assert rig.upstream.calls(marker) == () + + +def _closed_chat(rig: _Rig, model: str, marker: str) -> httpx.Response: + return rig.gateway.request("POST", "/v1/chat/completions", _chat_body(model, marker), headers=_CLOSE) + + +def test_live_declaration_update_turns_the_gate_on_and_off_on_every_worker(rig: _Rig) -> None: + chat_capable: Final[dict[str, JsonValue]] = { + "mode": "chat", + "supported_endpoints": ["/v1/chat/completions", "/v1/embeddings"], + } + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + name, identities = _deploy(scenario, rig.api_base, "flip", _single(chat_capable)) + _served_by_every_worker(rig.gateway, frozenset(identities)) + served: Final = _closed_chat(rig, name, marker) + _assert_served(rig, served, marker, "/v1/chat/completions") + patched: Final = rig.gateway.request( + "PATCH", f"/model/{identities[0]}/update", {"model_info": dict(_EMBED_ONLY)} + ) + assert patched.status_code == 200, patched.text + rejected: Final = held( + lambda: _closed_chat(rig, name, _marker()), + lambda response: response.status_code == 400, + holding=_RELOAD_SECONDS * 2, + seconds=_RELOAD_SECONDS * 12, + ) + _assert_error(rejected, _openai_error("/v1/chat/completions", name, "/v1/embeddings")) + restored: Final = rig.gateway.request("PATCH", f"/model/{identities[0]}/update", {"model_info": chat_capable}) + assert restored.status_code == 200, restored.text + held( + lambda: _closed_chat(rig, name, _marker()), + lambda response: response.status_code == 200, + holding=_RELOAD_SECONDS * 2, + seconds=_RELOAD_SECONDS * 12, + ) + + +_TESTS_ROOT: Final = Path(__file__).resolve().parents[2] +_CONFIG: Final = TypeAdapter(dict[str, JsonValue]) + + +def _owned_embedding(owned: Gateway, model: str, marker: str) -> httpx.Response: + return owned.request("POST", "/v1/embeddings", {"model": model, "input": marker, "encoding_format": "float"}) + + +@dataclass(frozen=True, slots=True) +class _Routed: + gateway: Gateway + upstream: _Upstream + names: Mapping[str, str] + + +def _routed_entry(name: str, api_base: str, litellm_model: str, info: Mapping[str, JsonValue]) -> dict[str, JsonValue]: + return { + "model_name": name, + "litellm_params": {"model": litellm_model, "api_base": api_base, "api_key": _PROVIDER_KEY}, + "model_info": dict(info), + } + + +def _routed_config(directory: Path, api_base: str, names: Mapping[str, str]) -> Path: + shared: Final = _CONFIG.validate_python(yaml.safe_load((_TESTS_ROOT / "integration/proxy_config.yaml").read_text())) + config: Final = { + **shared, + "model_list": [ + _routed_entry(names["embed"], api_base, f"openai/{_EMBEDDING_MODEL}", _EMBED_ONLY), + _routed_entry(names["chat"], api_base, f"openai/{_CHAT_MODEL}", {"mode": "chat"}), + ], + "router_settings": { + **object_value(shared["router_settings"]), + "num_retries": 0, + "default_fallbacks": [names["chat"]], + "model_group_alias": {names["alias_embed"]: names["embed"], names["alias_chat"]: names["chat"]}, + }, + } + path: Final = directory / "gate-routed-config.yaml" + path.write_text(yaml.safe_dump(config, sort_keys=False)) + return path + + +@pytest.fixture(scope="module") +def routed(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Routed]: + directory: Final = tmp_path_factory.mktemp("gate-routed") + suffix: Final = uuid.uuid4().hex[:12] + names: Final = MappingProxyType( + {label: f"gate-routed-{label}-{suffix}" for label in ("embed", "chat", "alias_embed", "alias_chat")} + ) + with ( + wire_server(answering_model_discovery(_respond)) as wire, + gateway_from_environment() as rig_gateway, + owned_proxy_process( + rig_gateway, + directory, + {}, + config=_routed_config(directory, wire.url + "/v1", names), + workers=2, + state=OwnedState.SHARED_WITH_WORKER, + ) as owned, + ): + eventually( + lambda: owned.gateway.request("POST", "/v1/chat/completions", _chat_body(names["chat"], "probe")), + lambda response: response.status_code == 200, + seconds=30, + ) + yield _Routed(owned.gateway, _Upstream(wire), names) + + +_ROUTED_TIMEOUT: Final = pytest.mark.timeout(2 * graceful_stop_seconds() + 120) +_DEFAULT_FALLBACK_ROUTES: Final[Mapping[str, str]] = MappingProxyType( + {"chat_v1": "/v1/chat/completions", "messages": "/v1/responses", "responses_v1": "/v1/responses"} +) + + +@_ROUTED_TIMEOUT +@pytest.mark.parametrize("route", tuple(_DEFAULT_FALLBACK_ROUTES)) +def test_router_default_fallbacks_answer_for_an_embedding_only_group(routed: _Routed, route: str) -> None: + marker: Final = _marker() + response: Final = _send(routed.gateway, _ROUTES[route], routed.names["embed"], marker) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert routed.upstream.targets(marker) == (_DEFAULT_FALLBACK_ROUTES[route],) + + +_OPT_OUTS: Final[Mapping[str, Mapping[str, JsonValue]]] = MappingProxyType( + { + "empty_fallbacks": {"fallbacks": []}, + "null_fallbacks": {"fallbacks": None}, + "disable_fallbacks": {"disable_fallbacks": True}, + } +) + + +@_ROUTED_TIMEOUT +@pytest.mark.parametrize("opt_out", tuple(_OPT_OUTS)) +def test_request_opt_outs_keep_the_rejection_over_router_default_fallbacks(routed: _Routed, opt_out: str) -> None: + model: Final = routed.names["embed"] + marker: Final = _marker() + response: Final = routed.gateway.request( + "POST", "/v1/chat/completions", {**_chat_body(model, marker), **_OPT_OUTS[opt_out]} + ) + _assert_rejected_before( + routed.upstream, response, marker, _openai_error("/v1/chat/completions", model, "/v1/embeddings") + ) + + +@_ROUTED_TIMEOUT +def test_model_group_alias_to_an_embedding_only_group_falls_back_to_the_default(routed: _Routed) -> None: + marker: Final = _marker() + response: Final = routed.gateway.request( + "POST", "/v1/chat/completions", _chat_body(routed.names["alias_embed"], marker) + ) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert routed.upstream.targets(marker) == ("/v1/chat/completions",) + + +@_ROUTED_TIMEOUT +def test_model_group_alias_to_an_embedding_only_group_is_rejected_without_fallbacks(routed: _Routed) -> None: + alias: Final = routed.names["alias_embed"] + marker: Final = _marker() + response: Final = routed.gateway.request( + "POST", "/v1/chat/completions", {**_chat_body(alias, marker), "disable_fallbacks": True} + ) + _assert_rejected_before( + routed.upstream, response, marker, _openai_error("/v1/chat/completions", alias, "/v1/embeddings") + ) + + +@_ROUTED_TIMEOUT +def test_model_group_alias_to_a_chat_group_keeps_reaching_the_upstream(routed: _Routed) -> None: + marker: Final = _marker() + response: Final = routed.gateway.request( + "POST", "/v1/chat/completions", _chat_body(routed.names["alias_chat"], marker) + ) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert routed.upstream.targets(marker) == ("/v1/chat/completions",) + + +@_ROUTED_TIMEOUT +def test_model_group_alias_to_an_embedding_only_group_keeps_serving_embeddings(routed: _Routed) -> None: + marker: Final = _marker() + response: Final = _owned_embedding(routed.gateway, routed.names["alias_embed"], marker) + assert response.status_code == 200, response.text + assert list_value( + object_value(list_value(JSON_OBJECT.validate_json(response.content)["data"])[0])["embedding"] + ) == list(_VECTOR), response.text + assert routed.upstream.targets(marker) == ("/v1/embeddings",) + + +_REWRITER: Final = """from litellm.integrations.custom_logger import CustomLogger + + +class WebSocketModelRewriter(CustomLogger): + async def async_pre_call_hook(self, user_api_key_dict, cache, data, call_type): + if call_type == "_aresponses_websocket" and data.get("model") == "__SOURCE__": + data["model"] = "__TARGET__" + return data + + +rewriter = WebSocketModelRewriter() +""" + + +@dataclass(frozen=True, slots=True) +class _Rewritten: + gateway: Gateway + upstream: _Upstream + source: str + target: str + + +def _rewritten_config(directory: Path, api_base: str, source: str, target: str, module: str) -> Path: + shared: Final = _CONFIG.validate_python(yaml.safe_load((_TESTS_ROOT / "integration/proxy_config.yaml").read_text())) + config: Final = { + **shared, + "model_list": [ + _routed_entry(source, api_base, f"openai/{_CHAT_MODEL}", {"mode": "chat"}), + _routed_entry(target, api_base, f"openai/{_EMBEDDING_MODEL}", _EMBED_ONLY), + ], + "router_settings": {**object_value(shared["router_settings"]), "num_retries": 0}, + "litellm_settings": {**object_value(shared["litellm_settings"]), "callbacks": [f"{module}.rewriter"]}, + } + path: Final = directory / "gate-rewritten-config.yaml" + path.write_text(yaml.safe_dump(config, sort_keys=False)) + return path + + +@pytest.fixture(scope="module") +def rewritten(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Rewritten]: + directory: Final = tmp_path_factory.mktemp("gate-rewritten") + suffix: Final = uuid.uuid4().hex[:12] + source: Final = f"gate-rewrite-chat-{suffix}" + target: Final = f"gate-rewrite-embed-{suffix}" + module: Final = f"gate_ws_rewriter_{suffix}" + (directory / f"{module}.py").write_text(_REWRITER.replace("__SOURCE__", source).replace("__TARGET__", target)) + with ( + wire_server(answering_model_discovery(_respond)) as wire, + gateway_from_environment() as rig_gateway, + owned_proxy_process( + rig_gateway, + directory, + {}, + config=_rewritten_config(directory, wire.url + "/v1", source, target, module), + workers=2, + state=OwnedState.SHARED_WITH_WORKER, + ) as owned, + ): + eventually( + lambda: owned.gateway.request("POST", "/v1/chat/completions", _chat_body(source, "probe")), + lambda response: response.status_code == 200, + seconds=30, + ) + yield _Rewritten(owned.gateway, _Upstream(wire), source, target) + + +@_ROUTED_TIMEOUT +@pytest.mark.parametrize("limit", tuple(_RESERVING_KEYS)) +def test_a_websocket_model_a_hook_rewrites_to_an_embedding_only_group_releases_the_key_reservation( + rewritten: _Rewritten, limit: str +) -> None: + markers: Final = tuple(_marker() for _ in range(3)) + served_marker: Final = _marker() + with rewritten.gateway.scenario() as scenario: + key: Final = scenario.key(**_RESERVING_KEYS[limit]) + closed: Final = tuple( + asyncio.run(_websocket_first_frame_and_close(rewritten.gateway, key, rewritten.source, marker)) + for marker in markers + ) + assert closed == (_websocket_rejection(rewritten.target),) * 3 + served: Final = rewritten.gateway.request( + "POST", "/v1/chat/completions", _chat_body(rewritten.source, served_marker), key=key + ) + assert served.status_code == 200, served.text + assert _SERVED_TEXT in served.text, served.text + assert [rewritten.upstream.calls(marker) for marker in markers] == [()] * 3 + assert rewritten.upstream.targets(served_marker) == ("/v1/chat/completions",) + + +@dataclass(frozen=True, slots=True) +class _Addressed: + gateway: Gateway + rig: Gateway + upstream: _Upstream + names: Mapping[str, str] + addresses: Mapping[str, str] + api_base: str + + +def _wildcard_backend_entry(name: str, api_base: str, backend: str) -> dict[str, JsonValue]: + entry: Final = _routed_entry(name, api_base, backend, _EMBED_ONLY) + return {**entry, "litellm_params": {**object_value(entry["litellm_params"]), "custom_llm_provider": "openai"}} + + +def _addressed_config( + directory: Path, api_base: str, names: Mapping[str, str], suffix: str, with_wildcards: bool +) -> Path: + shared: Final = _CONFIG.validate_python(yaml.safe_load((_TESTS_ROOT / "integration/proxy_config.yaml").read_text())) + wildcard: Final = f"gate-addr-wild-{suffix}" + wildcard_entries: Final = ( + ( + _routed_entry(f"{wildcard}/*", api_base, "openai/*", {}), + _wildcard_backend_entry(names["wildcard_backend"], api_base, f"{wildcard}/{_CHAT_MODEL}"), + ) + if with_wildcards + else () + ) + aliases: Final = {names["wildcard_alias"]: f"{wildcard}/{_CHAT_BACKEND_PREFIX}{suffix}"} if with_wildcards else {} + config: Final = { + **shared, + "model_list": [ + _routed_entry( + names["embed"], api_base, f"openai/gate-addr-backend-{suffix}", {**_EMBED_ONLY, "id": names["embed_id"]} + ), + _routed_entry(names["chat"], api_base, f"openai/{_CHAT_MODEL}", {"mode": "chat", "id": names["chat_id"]}), + _routed_entry(names["shared_embed"], api_base, f"openai/{_CHAT_MODEL}", _EMBED_ONLY), + *wildcard_entries, + _routed_entry(names["ghost_primary"], api_base, f"openai/{_EMBEDDING_MODEL}", _EMBED_ONLY), + ], + "router_settings": { + **object_value(shared["router_settings"]), + "num_retries": 0, + "model_group_alias": aliases, + "fallbacks": [{names["ghost_primary"]: [names["ghost"]]}, {names["ghost"]: [names["chat"]]}], + }, + "general_settings": { + **object_value(shared["general_settings"]), + "proxy_config_reload_interval_seconds": 3600, + }, + } + path: Final = directory / "gate-addressed-config.yaml" + path.write_text(yaml.safe_dump(config, sort_keys=False)) + return path + + +@contextlib.contextmanager +def _addressed_proxy(tmp_path_factory: pytest.TempPathFactory, with_wildcards: bool) -> Iterator[_Addressed]: + directory: Final = tmp_path_factory.mktemp("gate-addressed") + suffix: Final = uuid.uuid4().hex[:12] + names: Final = MappingProxyType( + { + label: f"gate-addr-{label}-{suffix}" + for label in ( + "embed", + "embed_id", + "chat", + "chat_id", + "shared_embed", + "wildcard_backend", + "wildcard_alias", + "ghost_primary", + "ghost", + ) + } + ) + with ( + wire_server(answering_model_discovery(_respond)) as wire, + gateway_from_environment() as rig_gateway, + owned_proxy_process( + rig_gateway, + directory, + {}, + config=_addressed_config(directory, wire.url + "/v1", names, suffix, with_wildcards), + workers=2, + state=OwnedState.SHARED_WITH_WORKER, + ) as owned, + ): + eventually( + lambda: owned.gateway.request("POST", "/v1/chat/completions", _chat_body(names["chat"], "probe")), + lambda response: response.status_code == 200, + seconds=30, + ) + yield _Addressed( + owned.gateway, + rig_gateway, + _Upstream(wire), + names, + MappingProxyType( + { + "deployment_id": names["embed_id"], + "backend_name": f"openai/gate-addr-backend-{suffix}", + "wildcard_backend_name": f"gate-addr-wild-{suffix}/{_CHAT_MODEL}", + } + ), + wire.url + "/v1", + ) + + +@pytest.fixture(scope="module") +def addressed(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Addressed]: + with _addressed_proxy(tmp_path_factory, with_wildcards=True) as proxy: + yield proxy + + +@pytest.fixture(scope="module") +def addressed_without_wildcards(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Addressed]: + with _addressed_proxy(tmp_path_factory, with_wildcards=False) as proxy: + yield proxy + + +_ADDRESSED_ROUTES: Final = ("chat_v1", "messages", "responses_v1") + + +@_ROUTED_TIMEOUT +@pytest.mark.parametrize("route", _ADDRESSED_ROUTES) +@pytest.mark.parametrize("address", ("deployment_id", "backend_name")) +def test_chat_sent_to_an_embedding_only_deployment_id_or_backend_name_is_rejected_before_the_upstream( + addressed: _Addressed, address: str, route: str +) -> None: + spec: Final = _ROUTES[route] + model: Final = addressed.addresses[address] + marker: Final = _marker() + response: Final = _send(addressed.gateway, spec, model, marker) + expected: Final = ( + _anthropic_error(model, "/v1/embeddings") + if spec.anthropic_shape + else _openai_error(spec.endpoint, model, "/v1/embeddings") + ) + _assert_rejected_before(addressed.upstream, response, marker, expected) + + +@_ROUTED_TIMEOUT +@pytest.mark.parametrize("address", ("deployment_id", "backend_name")) +def test_embeddings_sent_to_an_embedding_only_deployment_id_or_backend_name_keep_working( + addressed: _Addressed, address: str +) -> None: + marker: Final = _marker() + response: Final = _owned_embedding(addressed.gateway, addressed.addresses[address], marker) + assert response.status_code == 200, response.text + assert list_value( + object_value(list_value(JSON_OBJECT.validate_json(response.content)["data"])[0])["embedding"] + ) == list(_VECTOR), response.text + assert addressed.upstream.targets(marker) == ("/v1/embeddings",) + + +@_ROUTED_TIMEOUT +@pytest.mark.parametrize("address", ("chat_deployment_id", "backend_name_a_chat_deployment_shares")) +def test_chat_sent_to_a_chat_capable_deployment_id_or_backend_name_keeps_reaching_the_upstream( + addressed: _Addressed, address: str +) -> None: + model: Final = addressed.names["chat_id"] if address == "chat_deployment_id" else f"openai/{_CHAT_MODEL}" + marker: Final = _marker() + response: Final = addressed.gateway.request("POST", "/v1/chat/completions", _chat_body(model, marker)) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert addressed.upstream.targets(marker) == ("/v1/chat/completions",) + + +@_ROUTED_TIMEOUT +def test_request_fallbacks_naming_a_chat_deployment_id_answer_for_an_embedding_only_primary( + addressed: _Addressed, +) -> None: + marker: Final = _marker() + response: Final = addressed.gateway.request( + "POST", + "/v1/chat/completions", + {**_chat_body(addressed.names["embed"], marker), "fallbacks": [addressed.names["chat_id"]]}, + ) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert addressed.upstream.targets(marker) == ("/v1/chat/completions",) + + +@_ROUTED_TIMEOUT +def test_chat_sent_to_an_embedding_only_backend_name_a_wildcard_also_matches_is_served_by_the_wildcard( + addressed: _Addressed, +) -> None: + marker: Final = _marker() + response: Final = addressed.gateway.request( + "POST", "/v1/chat/completions", _chat_body(addressed.addresses["wildcard_backend_name"], marker) + ) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert [JSON_OBJECT.validate_json(call.body)["model"] for call in addressed.upstream.calls(marker)] == [_CHAT_MODEL] + + +@_ROUTED_TIMEOUT +def test_request_fallbacks_naming_an_alias_only_a_wildcard_serves_answer_for_an_embedding_only_primary( + addressed: _Addressed, +) -> None: + marker: Final = _marker() + response: Final = addressed.gateway.request( + "POST", + "/v1/chat/completions", + {**_chat_body(addressed.names["embed"], marker), "fallbacks": [addressed.names["wildcard_alias"]]}, + ) + _assert_fallback_answered(addressed.upstream, response, marker) + + +@_ROUTED_TIMEOUT +def test_yaml_fallbacks_through_a_group_with_no_deployments_answer_for_an_embedding_only_group( + addressed: _Addressed, +) -> None: + marker: Final = _marker() + response: Final = addressed.gateway.request( + "POST", "/v1/chat/completions", _chat_body(addressed.names["ghost_primary"], marker) + ) + _assert_fallback_answered(addressed.upstream, response, marker) + + +@_ROUTED_TIMEOUT +def test_an_embedding_only_model_the_read_through_loads_from_the_database_is_rejected_on_every_call( + addressed_without_wildcards: _Addressed, +) -> None: + addressed: Final = addressed_without_wildcards + name: Final = f"gate-addr-loaded-{uuid.uuid4().hex[:12]}" + markers: Final = tuple(_marker() for _ in range(6)) + with addressed.rig.scenario() as scenario: + identity: Final = _create( + addressed.rig, name, {"model": f"openai/{_EMBEDDING_MODEL}", "api_base": addressed.api_base}, _EMBED_ONLY + ) + scenario.cleanups.callback(scenario.delete_model, identity) + responses: Final = tuple( + addressed.gateway.request("POST", "/v1/chat/completions", _chat_body(name, marker), headers=_CLOSE) + for marker in markers + ) + expected: Final = _openai_error("/v1/chat/completions", name, "/v1/embeddings") + for response in responses: + _assert_error(response, expected) + assert [addressed.upstream.calls(marker) for marker in markers] == [()] * len(markers) + + +@pytest.mark.parametrize("metadata", ({"model_group": 7}, {"model_group": ["wrong-group"]})) +def test_rejection_with_invalid_metadata_is_logged_without_blocking_embeddings( + rig: _Rig, metadata: dict[str, JsonValue] +) -> None: + marker: Final = _marker() + model: Final = rig.names["embedding_only"] + rejected: Final = _chat(rig, model, marker, metadata=metadata) + _assert_rejected_chat(rig, rejected, marker, model) + embedded_marker: Final = _marker() + embedded: Final = rig.gateway.request( + "POST", "/v1/embeddings", {"model": model, "input": embedded_marker, "encoding_format": "float"} + ) + assert embedded.status_code == 200, embedded.text + assert rig.upstream.targets(embedded_marker) == ("/v1/embeddings",) + + +def test_existing_non_chat_metadata_on_a_chat_backend_remains_advisory(rig: _Rig) -> None: + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + name, identities = _deploy( + scenario, + rig.api_base, + "legacy-audio", + _single({"mode": "audio_speech", "supported_endpoints": ["/v1/audio/speech"]}), + ) + _served_by_every_worker(rig.gateway, frozenset(identities)) + response: Final = _chat(rig, name, marker) + _assert_served(rig, response, marker, "/v1/chat/completions") + + +def test_a_mixed_group_skips_its_restricted_embedding_deployment(rig: _Rig) -> None: + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + name: Final = f"gate-mixed-{uuid.uuid4().hex[:12]}" + embed: Final = _create( + scenario.gateway, + name, + {"model": f"openai/{_EMBEDDING_MODEL}", "api_base": rig.api_base, "order": 0}, + _EMBED_ONLY, + ) + scenario.cleanups.callback(scenario.delete_model, embed) + chat: Final = _create( + scenario.gateway, name, {"model": f"openai/{_CHAT_MODEL}", "api_base": rig.api_base, "order": 1}, {} + ) + scenario.cleanups.callback(scenario.delete_model, chat) + _served_by_every_worker(rig.gateway, frozenset({embed, chat})) + response: Final = _chat(rig, name, marker) + _assert_served(rig, response, marker, "/v1/chat/completions") + calls: Final = rig.upstream.calls(marker) + assert [JSON_OBJECT.validate_json(call.body)["model"] for call in calls] == [_CHAT_MODEL] + + +@pytest.mark.parametrize( + "fallbacks, expected_status", + [(("embedding_only",), 429), (("embedding_only", "chat_target"), 200), (("missing",), 400)], +) +def test_local_rate_limit_skips_restricted_fallbacks_and_still_tries_unloaded_ones( + rig: _Rig, fallbacks: tuple[str, ...], expected_status: int +) -> None: + limited: Final = rig.names["chat_rate_limited"] + targets: Final = tuple(rig.names[name] if name != "missing" else _marker() for name in fallbacks) + with rig.gateway.scenario() as scenario: + key: Final = scenario.key( + model_rpm_limit={limited: 1}, router_settings={"fallbacks": [{limited: list(targets)}]} + ) + first: Final = _chat(rig, limited, _marker(), key=key) + assert first.status_code == 200, first.text + marker: Final = _marker() + second: Final = _chat(rig, limited, marker, key=key) + assert second.status_code == expected_status, second.text + if expected_status == 200: + _assert_served(rig, second, marker, "/v1/chat/completions") + return + assert "Supported endpoints" not in second.text, second.text + assert rig.upstream.calls(marker) == () + + +def _websocket_fallback_config(directory: Path, primary: tuple[str, str], fallback: tuple[str, str]) -> Path: + shared: Final = _CONFIG.validate_python(yaml.safe_load((_TESTS_ROOT / "integration/proxy_config.yaml").read_text())) + config: Final = { + **shared, + "model_list": [ + _routed_entry(primary[0], primary[1], f"openai/{_EMBEDDING_MODEL}", _EMBED_ONLY), + _routed_entry(fallback[0], fallback[1], f"openai/{_CHAT_MODEL}", {"mode": "chat"}), + ], + "router_settings": { + **object_value(shared["router_settings"]), + "num_retries": 0, + "fallbacks": [{primary[0]: [fallback[0]]}], + }, + } + path: Final = directory / "gate-websocket-fallback.yaml" + path.write_text(yaml.safe_dump(config, sort_keys=False)) + return path + + +def test_responses_websocket_uses_a_configured_chat_fallback_without_calling_the_embedding_provider( + rig: _Rig, tmp_path: Path +) -> None: + marker: Final = _marker() + suffix: Final = uuid.uuid4().hex[:12] + primary: Final = f"gate-ws-embed-{suffix}" + fallback: Final = f"gate-ws-chat-{suffix}" + certificate: Final = write_self_signed_cert(tmp_path) + with ( + responses_ws_peer.responses_peer( + certificate, lambda identity: responses_ws_peer.events(identity, _SERVED_TEXT) + ) as peer, + owned_proxy_process( + rig.gateway, + tmp_path, + {"SSL_CERT_FILE": str(certificate[0])}, + config=_websocket_fallback_config(tmp_path, (primary, f"{rig.api_base}/{marker}"), (fallback, peer.url)), + workers=2, + state=OwnedState.SHARED_WITH_WORKER, + ) as owned, + owned.gateway.scenario() as scenario, + ): + key: Final = scenario.key() + (received,) = asyncio.run( + responses_ws_peer.session(str(owned.gateway.client.base_url), key, primary, (marker,)) + ) + assert received[-1]["type"] == "response.completed", received + assert responses_ws_peer.completed(received)["status"] == "completed" + assert any(event.get("delta") == _SERVED_TEXT for event in received), received + (forwarded,) = responses_ws_peer.drain(peer.frames) + assert marker in json.dumps(forwarded), forwarded + assert rig.upstream.calls(marker) == () + + +def test_client_routing_fields_cannot_disable_an_explicit_restriction(rig: _Rig) -> None: + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + model, identities = _deploy(scenario, rig.api_base, "restricted-chat", _single(_EMBED_ONLY)) + _served_by_every_worker(rig.gateway, frozenset(identities)) + response: Final = _chat( + rig, + model, + marker, + original_generic_function="caller-value", + additional_drop_params=["original_generic_function"], + ) + _assert_rejected_chat(rig, response, marker, model) + + +def test_a_compaction_endpoint_declaration_keeps_the_model_available_for_compaction(rig: _Rig) -> None: + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + model, identities = _deploy( + scenario, + rig.api_base, + "compact-declaration", + _single({"supported_endpoints": ["/responses/compact"], "enforce_supported_endpoints": True}), + ) + _served_by_every_worker(rig.gateway, frozenset(identities)) + response: Final = rig.gateway.request("POST", "/responses/compact", {"model": model, "input": marker}) + _assert_served(rig, response, marker, "/v1/responses/compact") + + +@pytest.mark.parametrize("alias_source", ["key_alias", "key_router", "team_router"]) +def test_local_rate_limit_fallbacks_resolve_saved_aliases(rig: _Rig, tmp_path: Path, alias_source: str) -> None: + suffix: Final = uuid.uuid4().hex[:12] + names: Final = {label: f"gate-fallback-alias-{label}-{suffix}" for label in ("primary", "target", "alias")} + shared: Final = _CONFIG.validate_python(yaml.safe_load((_TESTS_ROOT / "integration/proxy_config.yaml").read_text())) + config: Final = { + **shared, + "model_list": [ + _routed_entry(names["primary"], rig.api_base, f"openai/{_CHAT_MODEL}", {"mode": "chat"}), + _routed_entry(names["target"], rig.api_base, f"openai/{_CHAT_MODEL}", {"mode": "chat"}), + ], + "router_settings": { + **object_value(shared["router_settings"]), + "num_retries": 0, + "fallbacks": [{names["primary"]: [names["alias"]]}], + }, + } + path: Final = tmp_path / "fallback-alias.yaml" + path.write_text(yaml.safe_dump(config, sort_keys=False)) + aliases: Final = {names["alias"]: names["target"]} + with ( + owned_proxy_process( + rig.gateway, tmp_path, {}, config=path, workers=2, state=OwnedState.SHARED_WITH_WORKER + ) as owned, + owned.gateway.scenario() as scenario, + ): + team: Final = ( + scenario.team(router_settings={"model_group_alias": aliases}) if alias_source == "team_router" else None + ) + key: Final = scenario.key( + team_id=team, + aliases=aliases if alias_source == "key_alias" else {}, + router_settings={"model_group_alias": aliases} if alias_source == "key_router" else None, + model_rpm_limit={names["primary"]: 1}, + ) + first: Final = owned.gateway.request( + "POST", "/v1/chat/completions", _chat_body(names["primary"], _marker()), key=key + ) + assert first.status_code == 200, first.text + marker: Final = _marker() + second: Final = owned.gateway.request( + "POST", "/v1/chat/completions", _chat_body(names["primary"], marker), key=key + ) + _assert_fallback_answered(rig.upstream, second, marker) diff --git a/tests/unit/proxy/common_utils/test_callback_utils.py b/tests/unit/proxy/common_utils/test_callback_utils.py index ecb2375d495..c8d909ac2eb 100644 --- a/tests/unit/proxy/common_utils/test_callback_utils.py +++ b/tests/unit/proxy/common_utils/test_callback_utils.py @@ -49,6 +49,19 @@ def test_get_remaining_tokens_and_requests_from_request_data(): } +@pytest.mark.parametrize("client_model_group", [7, ["embed-only"], {"name": "embed-only"}]) +def test_remaining_limit_headers_ignore_a_model_group_that_is_not_a_string(client_model_group: object): + casedata: Final = { + "metadata": { + "model_group": client_model_group, + f"litellm-key-remaining-requests-{client_model_group}": 100, + f"litellm-key-remaining-tokens-{client_model_group}": 200, + } + } + + assert get_remaining_tokens_and_requests_from_request_data(casedata) == {} + + @patch( "litellm.proxy.common_utils.callback_utils.CustomLogger.get_callback_env_vars", return_value=["API_KEY", "MISSING_VAR"], diff --git a/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py b/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py index b892dc97e5c..835e7605955 100644 --- a/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py +++ b/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py @@ -3617,6 +3617,33 @@ class TestDeleteTeamModels: assert any(e[0] == "delete_many" for e in prisma.events) +@pytest.mark.parametrize( + ("model_info_patch", "expected_gate"), + [ + pytest.param({"tpm": 5}, True, id="unrelated-field"), + pytest.param({"supported_endpoints": ["/v1/embeddings", "/v1/rerank"]}, True, id="endpoint-list-only"), + pytest.param({"enforce_supported_endpoints": False}, False, id="explicit-false"), + ], +) +def test_model_info_patch_changes_the_endpoint_gate_only_when_it_sends_the_flag( + model_info_patch: dict[str, object], expected_gate: bool +): + from litellm.proxy.management_endpoints.model_management_endpoints import update_db_model + + result: Final = update_db_model( + db_model=Deployment( + model_name="embedding-only", + litellm_params=LiteLLM_Params(model="openai/text-embedding-3-small"), + model_info=ModelInfo( + id="embedding-only", enforce_supported_endpoints=True, supported_endpoints=["/v1/embeddings"] + ), + ), + updated_patch=updateDeployment.model_validate({"model_info": model_info_patch}), + ) + + assert json.loads(result["model_info"]).get("enforce_supported_endpoints") is expected_gate + + def _build_db_model_for_blocked_test(): from litellm.types.router import Deployment, LiteLLM_Params, ModelInfo diff --git a/tests/unit/proxy/response_api_endpoints/test_endpoints.py b/tests/unit/proxy/response_api_endpoints/test_endpoints.py index e3d514b2ea9..b11dcc58ef6 100644 --- a/tests/unit/proxy/response_api_endpoints/test_endpoints.py +++ b/tests/unit/proxy/response_api_endpoints/test_endpoints.py @@ -2055,6 +2055,22 @@ def test_responses_ws_failure_frame_uses_the_http_status_for_the_same_failure( assert frame["error"]["message"] == handle_exception_on_proxy(failure).message +def test_responses_ws_failure_frame_shows_only_the_refusal_for_a_model_that_refuses_the_endpoint(): + from litellm.exceptions import EndpointNotSupportedByModelError + from litellm.proxy.response_api_endpoints.endpoints import _responses_ws_failure_frame + + refusal: Final = "/v1/responses is not supported by model=embed-only. Supported endpoints: /v1/embeddings" + failure: Final = EndpointNotSupportedByModelError( + message=f"{refusal}\n\nLiteLLM: model group 'embed-only' failed with the error above", model="embed-only" + ) + + assert json.loads(_responses_ws_failure_frame(failure)) == { + "type": "error", + "status": 400, + "error": {"type": "invalid_request_error", "message": refusal}, + } + + class TestWSSessionCostTracking: @pytest.mark.asyncio async def test_router_budget_limiter_skips_arealtime_call_type(self): diff --git a/tests/unit/proxy/spend_tracking/test_spend_tracking_utils.py b/tests/unit/proxy/spend_tracking/test_spend_tracking_utils.py index 23f9ac0e2d8..740334252d2 100644 --- a/tests/unit/proxy/spend_tracking/test_spend_tracking_utils.py +++ b/tests/unit/proxy/spend_tracking/test_spend_tracking_utils.py @@ -5384,6 +5384,28 @@ def test_get_logging_payload_router_rejected_request_takes_provider_from_model_g assert payload["custom_llm_provider"] == expected_provider +@pytest.mark.parametrize("client_model_group", [7, ["embed-only"], {"name": "embed-only"}]) +def test_get_logging_payload_drops_a_client_model_group_that_is_not_a_string(client_model_group: object): + payload: Final = get_logging_payload( + kwargs={ + "model": "embed-only", + "litellm_params": { + "metadata": {"user_api_key": "test-key", "model_group": client_model_group, "status": "failure"} + }, + "standard_logging_object": { + **_make_failed_request_standard_logging_payload(), + "model_group": client_model_group, + }, + }, + response_obj={}, + start_time=datetime.datetime.now(timezone.utc), + end_time=datetime.datetime.now(timezone.utc), + llm_router=None, + ) + + assert payload["model_group"] == "" + + def test_get_logging_payload_router_rejected_request_without_router_leaves_provider_empty(): assert _router_rejected_failure_payload("openai-group", None)["custom_llm_provider"] == "" diff --git a/tests/unit/proxy/test_common_request_processing.py b/tests/unit/proxy/test_common_request_processing.py index fdcbcbfcd05..e58a1de9eb5 100644 --- a/tests/unit/proxy/test_common_request_processing.py +++ b/tests/unit/proxy/test_common_request_processing.py @@ -456,7 +456,7 @@ class TestProxyBaseLLMRequestProcessing: mock_add_litellm_data_to_request, ) mock_general_settings = {} - mock_user_api_key_dict = MagicMock(spec=UserAPIKeyAuth) + mock_user_api_key_dict = ProxyUserAPIKeyAuth() mock_proxy_config = MagicMock(spec=ProxyConfig) route_type = "acompletion" @@ -740,7 +740,7 @@ class TestProxyBaseLLMRequestProcessing: returned_data, logging_obj = await processing_obj.common_processing_pre_call_logic( request=mock_request, general_settings={}, - user_api_key_dict=MagicMock(spec=UserAPIKeyAuth), + user_api_key_dict=ProxyUserAPIKeyAuth(), proxy_logging_obj=mock_proxy_logging_obj, proxy_config=MagicMock(spec=ProxyConfig), route_type=route_type, @@ -1186,7 +1186,7 @@ class TestProxyBaseLLMRequestProcessing: ) mock_general_settings = {} - mock_user_api_key_dict = MagicMock(spec=UserAPIKeyAuth) + mock_user_api_key_dict = ProxyUserAPIKeyAuth() mock_proxy_config = MagicMock(spec=ProxyConfig) mock_router_settings = { @@ -2208,7 +2208,7 @@ class TestProxyBaseLLMRequestProcessing: mock_add_litellm_data_to_request, ) mock_general_settings = {} - mock_user_api_key_dict = MagicMock(spec=UserAPIKeyAuth) + mock_user_api_key_dict = ProxyUserAPIKeyAuth() mock_proxy_config = MagicMock(spec=ProxyConfig) route_type = "acompletion" @@ -7706,6 +7706,56 @@ class TestPreCallWithFallbacksOnLocalRateLimit: assert processor.data["model"] == fallback_model assert call_count == 2 + @pytest.mark.asyncio + async def test_local_rate_limit_fallback_skips_an_aliased_embedding_only_model(self): + from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing + from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError + + router: Final = Router( + model_list=[ + { + "model_name": "embed-only", + "litellm_params": {"model": "openai/text-embedding-3-small", "api_key": "sk-synthetic"}, + "model_info": {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}, + }, + {"model_name": "chat-model", "litellm_params": {"model": "openai/gpt-5.4-nano", "api_key": "sk-synthetic"}}, + ], + fallbacks=[{"rate-limited": ["embed-alias", "chat-model"]}], + ) + processor: Final = ProxyBaseLLMRequestProcessing( + data={ + "model": "rate-limited", + "router_settings_override": {"model_group_alias": {"embed-alias": "embed-only"}}, + } + ) + attempted: Final[list[object]] = [] + + async def pre_call_logic(**_: object): + attempted.append(processor.data.get("model")) + if processor.data.get("model") == "rate-limited": + raise ProxyRateLimitError(detail="key rpm limit exceeded", headers={"retry-after": "30"}) + return processor.data, MagicMock() + + with patch.object(processor, "common_processing_pre_call_logic", side_effect=pre_call_logic): + await processor._pre_call_with_fallbacks( + request=MagicMock(), + general_settings={}, + proxy_logging_obj=MagicMock(), + user_api_key_dict=ProxyUserAPIKeyAuth(), + version=None, + proxy_config=MagicMock(), + user_model=None, + user_temperature=None, + user_request_timeout=None, + user_max_tokens=None, + user_api_base=None, + model="rate-limited", + route_type="acompletion", + llm_router=router, + ) + + assert attempted == ["rate-limited", "chat-model"] + @pytest.mark.asyncio async def test_raises_when_no_fallbacks_configured(self): from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError diff --git a/tests/unit/proxy/test_route_llm_request.py b/tests/unit/proxy/test_route_llm_request.py index 9e31d3774f9..8f364e51e2b 100644 --- a/tests/unit/proxy/test_route_llm_request.py +++ b/tests/unit/proxy/test_route_llm_request.py @@ -28,6 +28,7 @@ from litellm.proxy.route_llm_request import ( JSON_OBJECT_ADAPTER, ProxyModelNotFoundError, UserConfigRouters, + endpoint_unsupported_by_model, route_request, user_config_routers, ) @@ -3027,3 +3028,27 @@ async def test_user_config_routers_discard_the_least_recently_used_router_past_c assert first_handler() is not None assert second_handler() is None routers.clear() + + +def test_fallback_probe_skips_only_a_loaded_model_that_refuses_the_route() -> None: + router: Final = litellm.Router( + model_list=[ + { + "model_name": "chat-fallback", + "litellm_params": {"model": "openai/gpt-5.4-nano", "api_key": "unit-provider-key"}, + }, + { + "model_name": "embed-only", + "litellm_params": {"model": "openai/text-embedding-3-small", "api_key": "unit-provider-key"}, + "model_info": {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}, + }, + ], + num_retries=0, + ) + + def skipped(model: str) -> bool: + return endpoint_unsupported_by_model("acompletion", {"model": model}, router, None) + + assert skipped("embed-only") is True + assert skipped("chat-fallback") is False + assert skipped("added-on-another-worker") is False diff --git a/tests/unit/router_utils/test_endpoint_eligibility.py b/tests/unit/router_utils/test_endpoint_eligibility.py new file mode 100644 index 00000000000..5e5d77254d0 --- /dev/null +++ b/tests/unit/router_utils/test_endpoint_eligibility.py @@ -0,0 +1,139 @@ +from collections.abc import Mapping +from typing import Final + +import pytest +from pydantic import TypeAdapter + +from litellm.exceptions import EndpointNotSupportedByModelError +from litellm.router_utils.endpoint_eligibility import ( + eligible_deployments, + endpoint_selection, + warn_on_unenforced_endpoint_flag, +) + +_DEPLOYMENT: Final = TypeAdapter(list[dict[str, object]]) + + +def _deployment(info: Mapping[str, object], identity: str = "embedding") -> dict[str, object]: + return _DEPLOYMENT.validate_python( + [ + { + "model_name": "shared", + "litellm_params": {"model": "openai/test-model"}, + "model_info": {"id": identity, **info}, + } + ] + )[0] + + +@pytest.mark.parametrize("call_type", ["acompletion", "anthropic_messages", "aresponses", "acreate_interaction"]) +def test_explicit_restriction_excludes_an_embedding_deployment_from_a_mixed_group(call_type: str) -> None: + embedding: Final = _deployment({"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}) + chat: Final = _deployment({"supported_endpoints": ["/v1/chat/completions"]}, "chat") + with endpoint_selection(call_type): + result: Final = eligible_deployments("shared", (embedding, chat)) + assert result == (chat,) + + +@pytest.mark.parametrize( + "info", + [ + {"mode": "audio_speech", "supported_endpoints": ["/v1/audio/speech"]}, + {"enforce_supported_endpoints": False, "supported_endpoints": ["/v1/embeddings"]}, + {"enforce_supported_endpoints": True}, + {"enforce_supported_endpoints": True, "supported_endpoints": None}, + {"enforce_supported_endpoints": True, "supported_endpoints": 7}, + {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/messages"]}, + ], +) +def test_advisory_or_incomplete_metadata_preserves_routing(info: Mapping[str, object]) -> None: + deployment: Final = _deployment(info) + with endpoint_selection("acompletion"): + result: Final = eligible_deployments("shared", (deployment,)) + assert result == (deployment,) + + +def test_a_single_endpoint_string_is_enforced_like_a_one_item_list() -> None: + deployment: Final = _deployment({"enforce_supported_endpoints": True, "supported_endpoints": "/v1/embeddings"}) + with endpoint_selection("acompletion"): + result: Final = eligible_deployments("shared", (deployment,)) + assert isinstance(result, EndpointNotSupportedByModelError) + assert result.message.endswith("Supported endpoints: /v1/embeddings") + + +def test_unsupported_error_explains_the_requested_operation_and_available_endpoints() -> None: + deployment: Final = _deployment({"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}) + with endpoint_selection("acreate_interaction"): + result: Final = eligible_deployments("shared", (deployment,)) + assert isinstance(result, EndpointNotSupportedByModelError) + assert result.status_code == 400 + assert "/v1beta/interactions" in result.message + assert "/v1/embeddings" in result.message + + +@pytest.mark.parametrize("call_type", ["aembedding", "aspeech", "_arealtime", "aget_responses"]) +def test_chat_restriction_does_not_block_other_operations(call_type: str) -> None: + deployment: Final = _deployment({"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}) + with endpoint_selection(call_type): + result: Final = eligible_deployments("shared", (deployment,)) + assert result == (deployment,) + + +def test_nested_non_chat_calls_do_not_inherit_the_outer_chat_restriction() -> None: + deployment: Final = _deployment({"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}) + with endpoint_selection("acompletion"): + with endpoint_selection("aembedding"): + embedding_result: Final = eligible_deployments("shared", (deployment,)) + chat_result: Final = eligible_deployments("shared", (deployment,)) + assert embedding_result == (deployment,) + assert isinstance(chat_result, EndpointNotSupportedByModelError) + + +@pytest.mark.parametrize( + "endpoint", + [ + "/messages", + "/interactions", + "/responses/compact", + "/openai/v1/responses/compact", + "/openai/v1/chat/completions", + "chat/completions", + "v1/responses", + " /v1/chat/completions ", + "/v1/chat/completions/", + "/V1/Chat/Completions", + "/v1beta/models/gemini-3-flash:generateContent", + ], +) +def test_chat_endpoint_declarations_keep_their_meaning_across_prefixes(endpoint: str) -> None: + deployment: Final = _deployment({"enforce_supported_endpoints": True, "supported_endpoints": [endpoint]}) + with endpoint_selection("acompletion"): + result: Final = eligible_deployments("shared", (deployment,)) + assert result == (deployment,) + + +@pytest.mark.parametrize( + "info", + [ + {"enforce_supported_endpoints": True}, + {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings", None]}, + {"enforce_supported_endpoints": "true", "supported_endpoints": 7}, + ], +) +def test_a_flag_without_a_usable_endpoint_list_is_reported(info: Mapping[str, object]) -> None: + warning: Final = warn_on_unenforced_endpoint_flag("embed-only", info) + assert warning is not None + assert "'embed-only'" in warning + + +@pytest.mark.parametrize( + "info", + [ + {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}, + {"enforce_supported_endpoints": True, "supported_endpoints": "/v1/embeddings"}, + {"enforce_supported_endpoints": False}, + {"supported_endpoints": ["/v1/embeddings", None]}, + ], +) +def test_an_enforceable_or_unflagged_deployment_is_not_reported(info: Mapping[str, object]) -> None: + assert warn_on_unenforced_endpoint_flag("embed-only", info) is None diff --git a/tests/unit/test_router/test_router.py b/tests/unit/test_router/test_router.py index ee3a70c7c1c..62510dac26d 100644 --- a/tests/unit/test_router/test_router.py +++ b/tests/unit/test_router/test_router.py @@ -12270,6 +12270,67 @@ def test_is_deployment_blocked_static_helper_reflects_blocked_flag(): ) +def test_synchronous_text_completion_rejects_a_restricted_model_before_the_provider() -> None: + model: Final = "text-completion-openai/unit-text-model" + router: Final = Router( + model_list=[ + { + "model_name": model, + "litellm_params": { + "model": model, + "api_base": "https://sync-router.invalid/v1", + "api_key": "unit-provider-key", + }, + "model_info": {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}, + } + ], + num_retries=0, + ) + with respx.mock(assert_all_called=False) as upstream: + provider: Final = upstream.post("https://sync-router.invalid/v1/completions").respond( + 200, + json={ + "id": "cmpl-unit", + "object": "text_completion", + "created": 1, + "model": "unit-text-model", + "choices": [{"text": "provider called", "index": 0, "finish_reason": "stop", "logprobs": None}], + "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}, + }, + ) + with pytest.raises(litellm.BadRequestError, match="/v1/completions is not supported"): + router.text_completion(model=model, prompt="test") + assert provider.call_count == 0 + + +def test_routable_deployments_leave_out_models_that_refuse_the_call_type() -> None: + router: Final = Router( + model_list=[ + { + "model_name": "available", + "litellm_params": {"model": "openai/gpt-5.4-nano", "api_key": "unit-provider-key"}, + }, + { + "model_name": "embed-only", + "litellm_params": {"model": "openai/text-embedding-3-small", "api_key": "unit-provider-key"}, + "model_info": {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}, + }, + ], + num_retries=0, + ) + + def routable(model: str, call_type: str | None) -> tuple[str, ...]: + return tuple( + deployment["model_name"] for deployment in router.get_routable_deployments(model, {}, call_type=call_type) + ) + + assert routable("unregistered-fallback", "acompletion") == () + assert routable("available", "acompletion") == ("available",) + assert routable("embed-only", "acompletion") == () + assert routable("embed-only", "aembedding") == ("embed-only",) + assert routable("embed-only", None) == ("embed-only",) + + def test_routing_groups_legacy_property_forwards_to_public_attribute(): router = _router_with_two_deployments([False, False]) @@ -18803,6 +18864,35 @@ async def test_anthropic_messages_stream_dropped_before_content_is_retried_withi assert [(attempted, budget) for _, attempted, budget in provider.calls] == [(0, 2), (1, 2)] +@pytest.mark.asyncio +async def test_anthropic_messages_same_group_retry_skips_an_embedding_only_deployment(): + """The retry of a /v1/messages stream dropped before content must not land on an enforced embedding-only + deployment of the same group, even when that deployment is the group's preferred order.""" + router: Final = Router( + model_list=[ + {"model_name": "glm", "litellm_params": {"model": "anthropic/glm-chat", "api_key": "sk-test", "order": 2}}, + { + "model_name": "glm", + "litellm_params": {"model": "anthropic/glm-embed", "api_key": "sk-test", "order": 1}, + "model_info": {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}, + }, + ], + num_retries=1, + ) + provider: Final = _AnthropicMessagesScriptedProvider( + _anthropic_messages_dropped_before_content, _anthropic_messages_retried_stream + ) + + async def anthropic_messages(**kwargs: object) -> object: + return await provider(**kwargs) + + stream: Final = await _anthropic_messages_stream_through_router(router, anthropic_messages) + body: Final = [chunk async for chunk in stream] + + assert body == [_anthropic_messages_message_start_chunk(), _anthropic_messages_content_chunk("pong")] + assert [model for model, _, _ in provider.calls] == ["anthropic/glm-chat", "anthropic/glm-chat"] + + class _AnthropicMessagesRetryStashRecorder: def __init__(self) -> None: self.observed: tuple[Usage | None, ...] = () @@ -31431,3 +31521,26 @@ def test_router_keeps_supported_fallback_shapes( router: Final = Router(model_list=list(_FALLBACK_VALIDATION_MODEL_LIST), **router_kwargs) assert router.fallbacks == expected_fallbacks + + +def test_router_warns_when_a_deployment_enforces_endpoints_without_a_list(caplog: pytest.LogCaptureFixture) -> None: + with caplog.at_level(logging.WARNING, logger="LiteLLM Router"): + router: Final = Router( + model_list=[ + { + "model_name": "embed-from-config", + "litellm_params": {"model": "openai/text-embedding-3-small", "api_key": "sk-fake"}, + "model_info": {"enforce_supported_endpoints": True}, + } + ] + ) + router.add_deployment( + Deployment( + model_name="embed-from-api", + litellm_params=LiteLLM_Params(model="openai/text-embedding-3-small", api_key="sk-fake"), + model_info=ModelInfo(enforce_supported_endpoints=True, supported_endpoints=["/v1/embeddings", 7]), + ) + ) + + assert "Deployment 'embed-from-config' sets enforce_supported_endpoints without" in caplog.text + assert "Deployment 'embed-from-api' sets enforce_supported_endpoints without" in caplog.text diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index f114c1c73d2..6faffd7a050 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -50433,6 +50433,8 @@ export interface components { discoverable?: boolean | null; /** Enable Tag Filtering */ enable_tag_filtering?: boolean | null; + /** Enforce Supported Endpoints */ + enforce_supported_endpoints?: boolean | null; /** Explicit Cache Read Input Token Cost */ explicit_cache_read_input_token_cost?: number | null; /** Id */ From 961fabbb76b253884469c1af81121e392b8987ef Mon Sep 17 00:00:00 2001 From: Mateo Wang <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 29 Oct 2026 08:44:46 -0700 Subject: [PATCH 2/7] feat(secret_managers): resolve oidc/google api_key references per request (#47703) * test(secret_managers): cover oidc/google api_key references resolved per request Unit tests send a metadata server ID token for an admin oidc/google/ api_key, reuse it within its lifetime, refresh it before it expires, fetch it off the event loop, fail closed with a 401 that names only the kind, and honor GCE_METADATA_HOST. The integration file scripts a metadata server and checks the bearer the upstream receives for config, stored credential and DB credential sources, plus a metadata failure * feat(secret_managers): resolve oidc/google api_key references per request An admin-set oidc/google/ api_key now resolves on each call to the GCE metadata server's ID token for that audience, cached until a minute before it expires. Async callers fetch on a cache miss in a worker thread, so the event loop never waits on the metadata server. A failed fetch answers 401 naming only the kind, and GCE_METADATA_HOST points the fetch at another metadata server. * fix(oidc): bound the Google metadata fetch and answer 503 when it fails A Google OIDC api_key fetch now waits at most GCE_METADATA_TIMEOUT seconds (default 3), runs on its own thread pool instead of the loop's default executor, and sends the token read from the cache at call start even if it expires mid-call. A metadata timeout, network error, or 5xx answers a retryable 503, while a 4xx or empty token stays a 401. The expiry and empty-token warnings log the secret name on one line. * test(oidc): expect 503 when the metadata server answers 5xx in the integration test * fix(oidc): answer 503 on token-count routes during a metadata outage and fall back on a bad GCE_METADATA_TIMEOUT * fix(secret_managers): keep get_secret's metadata wait, share concurrent google token fetches, refuse oidc/google on pass-through A model's api_key fetch keeps the 3 second GCE_METADATA_TIMEOUT wait while get_secret callers keep the handler default. Any metadata transport error answers a retryable 503. Async cache misses for one reference share one fetch, and a call the reference may not go to is refused before any fetch. Pass-through routes answer 401 for an oidc/google credential instead of sending the reference text. The metadata client is closed after each fetch, and the unused sync resolver is removed. * fix(utils): refuse an oidc/google api_key on pass-through calls before any fetch A router-model or credential-name pass-through call with an oidc/google reference now fails before the metadata fetch and sends nothing upstream, since the pass-through config never sends the resolved key. The single-flight test also waits on the fetch-start signal with a bounded window instead of a fixed sleep. --- .../litellm_executed_batches.py | 8 +- .../passthrough_endpoint_router.py | 9 + litellm/proxy/proxy_server.py | 22 +- litellm/realtime_api/main.py | 4 +- litellm/router.py | 4 +- litellm/secret_managers/main.py | 135 +++++-- litellm/utils.py | 120 +++++- .../security/test_api_key_secret_reference.py | 183 ++++++++- .../test_passthrough_endpoint_router.py | 19 +- tests/unit/proxy/test_proxy_token_counter.py | 33 ++ .../test_secret_managers_main.py | 42 +- tests/unit/test_utils.py | 376 +++++++++++++++++- 12 files changed, 884 insertions(+), 71 deletions(-) diff --git a/litellm/proxy/batches_endpoints/litellm_executed_batches.py b/litellm/proxy/batches_endpoints/litellm_executed_batches.py index 098dd2f9816..67728e0fd63 100644 --- a/litellm/proxy/batches_endpoints/litellm_executed_batches.py +++ b/litellm/proxy/batches_endpoints/litellm_executed_batches.py @@ -35,7 +35,7 @@ from litellm.repositories.managed_batch_repository import ManagedBatchRepository from litellm.types.llms.base import LiteLLMBaseModel from litellm.types.llms.openai import LiteLLMBatchCreateRequest, OpenAIFileObject, OpenAIFilesPurpose from litellm.types.utils import LITELLM_EXECUTED_BATCH_PROVIDERS, ExtractedFileData, LiteLLMBatch, LlmProviders -from litellm.utils import resolve_api_key_secret_reference +from litellm.utils import aresolve_api_key_secret_reference if TYPE_CHECKING: from prisma import types as prisma_types @@ -227,7 +227,7 @@ async def upstream_lacks_files_api(api_base: str, api_key: str | None, http_clie return response.status_code == httpx.codes.NOT_FOUND -def _upstream_of(credentials: Mapping[str, object], provider: str) -> tuple[str, str | None] | None: +async def _upstream_of(credentials: Mapping[str, object], provider: str) -> tuple[str, str | None] | None: model: Final = credentials.get("model") api_base: Final = credentials.get("api_base") api_key: Final = credentials.get("api_key") @@ -244,7 +244,7 @@ def _upstream_of(credentials: Mapping[str, object], provider: str) -> tuple[str, return None if resolved_api_base is None: return None - return resolved_api_base, resolve_api_key_secret_reference( + return resolved_api_base, await aresolve_api_key_secret_reference( resolved_api_key, model=model, llm_provider=provider, upstream=credentials ) @@ -255,7 +255,7 @@ async def litellm_executed_provider_for( provider: Final = litellm_executed_provider_of(credentials) if provider is None: return None - upstream: Final = _upstream_of(credentials, provider) + upstream: Final = await _upstream_of(credentials, provider) if upstream is None: return None return provider if await lacks_files_api(*upstream) else None diff --git a/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py b/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py index 1d0b3d00434..739d4b3a01a 100644 --- a/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py +++ b/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py @@ -12,6 +12,7 @@ from litellm.integrations.vector_store_integrations.vector_store_pre_call_hook i ) from litellm.litellm_core_utils.credential_accessor import CredentialAccessor from litellm.secret_managers.main import ( + GOOGLE_OIDC_SECRET_PREFIX, UnresolvableSecretReference, get_secret_str, resolve_local_oidc_reference, @@ -47,6 +48,14 @@ def _credential_identity(credentials: VERTEX_CREDENTIALS_TYPES | None) -> str | def _resolve_oidc_reference(credential: str | None) -> str | None: if credential is None: return None + if credential.startswith(GOOGLE_OIDC_SECRET_PREFIX): + raise HTTPException( + status_code=401, + detail=( + "pass-through credential oidc/google secret reference could not be resolved: pass-through routes " + "do not fetch Google ID tokens. Use a static key, oidc/file/ or oidc/env/." + ), + ) match resolve_local_oidc_reference(credential): case UnresolvableSecretReference(kind=kind, reason=reason): raise HTTPException( diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 091e05610ba..a75ef593b7a 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -207,7 +207,7 @@ from litellm.types.utils import ( is_server_derived_pricing_key, pricing_override_fields, ) -from litellm.utils import cost_map_omits_token_price, load_credentials_from_list, with_named_credential +from litellm.utils import acredential_updates, cost_map_omits_token_price, with_named_credential if TYPE_CHECKING: from aiohttp import ClientSession @@ -14764,10 +14764,10 @@ def _provider_count_outcome( return result -def _with_resolved_api_key(litellm_params: object) -> Mapping[str, object]: +async def _with_resolved_api_key(litellm_params: object) -> Mapping[str, object]: call_params: Final = _LITELLM_PARAMS_COPY.validate_python(litellm_params) try: - load_credentials_from_list(call_params) + return {**call_params, **(await acredential_updates(call_params))} except litellm.AuthenticationError as e: raise ProxyException( message=e.message, @@ -14775,13 +14775,19 @@ def _with_resolved_api_key(litellm_params: object) -> Mapping[str, object]: param="model", code=401, ) from e - return call_params + except litellm.ServiceUnavailableError as e: + raise ProxyException( + message=e.message, + type="service_unavailable", + param="model", + code=503, + ) from e -def _with_resolved_credentials(deployment: Mapping[str, object] | None) -> dict[str, object] | None: +async def _with_resolved_credentials(deployment: Mapping[str, object] | None) -> dict[str, object] | None: if deployment is None: return None - return {**deployment, "litellm_params": _with_resolved_api_key(deployment.get("litellm_params", {}))} + return {**deployment, "litellm_params": await _with_resolved_api_key(deployment.get("litellm_params", {}))} _DEPLOYMENT_LITELLM_PARAMS: Final = TypeAdapter(Mapping[str, object]) @@ -14844,7 +14850,7 @@ async def _try_provider_token_count( model_to_use=model_to_use or "", messages=messages, contents=contents, - deployment=_with_resolved_credentials(deployment_in_region), + deployment=await _with_resolved_credentials(deployment_in_region), request_model=request_model, tools=tools, system=system, @@ -14878,7 +14884,7 @@ async def _try_native_responses_count( model_to_use=model_to_use, request=responses_request, policy=policy, - deployment=_with_resolved_credentials(deployment), + deployment=await _with_resolved_credentials(deployment), request_model=request_model, ) ) diff --git a/litellm/realtime_api/main.py b/litellm/realtime_api/main.py index 919fd54e7d6..e03cb44efda 100644 --- a/litellm/realtime_api/main.py +++ b/litellm/realtime_api/main.py @@ -35,8 +35,8 @@ from litellm.types.utils import CallTypes, LlmProviders from litellm.utils import ( ModelInfoLookupMiss, ProviderConfigManager, + aresolve_api_key_secret_reference, lookup_model_info_for_registration, - resolve_api_key_secret_reference, ) from ..litellm_core_utils.core_helpers import as_str_mapping, normalize_drop_params @@ -845,7 +845,7 @@ async def realtime_health_check( resolved_api_base: Final = cast( # cast-ok: provider parameters expose optional string endpoints str | None, api_base or resolved_params.get("api_base") ) - resolved_api_key: Final = resolve_api_key_secret_reference( + resolved_api_key: Final = await aresolve_api_key_secret_reference( cast( # cast-ok: provider parameters expose optional string credentials str | None, api_key or resolved_params.get("api_key") ), diff --git a/litellm/router.py b/litellm/router.py index 31d679f616e..7d566bd0744 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -487,6 +487,7 @@ from litellm.utils import ( ProviderConfigManager, Rules, TextCompletionResponse, + aresolve_api_key_secret_reference, function_setup, get_llm_provider, get_non_default_completion_params, @@ -497,7 +498,6 @@ from litellm.utils import ( lookup_model_info_for_registration, model_info_with_pricing_overrides, provider_rejectable_params, - resolve_api_key_secret_reference, set_live_deployment_replay, token_counter, ) @@ -12079,7 +12079,7 @@ class Router: return if api_base is None or "*" in model or params.get("use_clientside_credentials"): return - api_key: Final = resolve_api_key_secret_reference( + api_key: Final = await aresolve_api_key_secret_reference( params.api_key or dynamic_api_key, model=model, llm_provider=provider, diff --git a/litellm/secret_managers/main.py b/litellm/secret_managers/main.py index 588bdb1409c..02cf53d00ec 100644 --- a/litellm/secret_managers/main.py +++ b/litellm/secret_managers/main.py @@ -5,7 +5,7 @@ import time import traceback from collections.abc import Callable from dataclasses import dataclass -from typing import Final +from typing import Final, TypeAlias import httpx from pydantic import TypeAdapter, ValidationError @@ -13,6 +13,7 @@ from pydantic import TypeAdapter, ValidationError import litellm from litellm._logging import verbose_logger from litellm.caching.caching import DualCache +from litellm.litellm_core_utils.env_utils import get_env_float from litellm.litellm_core_utils.single_flight import single_flight_lru_cache from litellm.llms.custom_httpx.http_handler import HTTPHandler from litellm.secret_managers.get_azure_ad_token_provider import ( @@ -167,12 +168,37 @@ def get_secret_str( LOCAL_OIDC_SECRET_PREFIXES: Final = ("oidc/file/", "oidc/env/") +GOOGLE_OIDC_SECRET_PREFIX: Final = "oidc/google/" +API_KEY_OIDC_SECRET_PREFIXES: Final = (*LOCAL_OIDC_SECRET_PREFIXES, GOOGLE_OIDC_SECRET_PREFIX) +_GOOGLE_METADATA_DEFAULT_HOST: Final = "metadata.google.internal" +_GOOGLE_METADATA_DEFAULT_TIMEOUT_SECONDS: Final = 3.0 +_GOOGLE_TOKEN_UNAVAILABLE: Final = "the metadata server did not return an ID token" + + +def _google_metadata_identity_url() -> str: + host: Final = os.getenv("GCE_METADATA_HOST") or _GOOGLE_METADATA_DEFAULT_HOST + return f"http://{host}/computeMetadata/v1/instance/service-accounts/default/identity" + + +def _google_metadata_timeout() -> float: + seconds: Final = get_env_float("GCE_METADATA_TIMEOUT", _GOOGLE_METADATA_DEFAULT_TIMEOUT_SECONDS) + return seconds if seconds > 0 else _GOOGLE_METADATA_DEFAULT_TIMEOUT_SECONDS + + +class GoogleMetadataStatusError(ValueError): + def __init__(self, status_code: int) -> None: + super().__init__(f"Google OIDC provider failed with HTTP {status_code}") + self.status_code: Final = status_code @dataclass(frozen=True, slots=True) class UnresolvableSecretReference: kind: str reason: str + retryable: bool = False + + +OidcResolver: TypeAlias = Callable[[str], "str | UnresolvableSecretReference"] _OIDC_PROVIDERS: Final = frozenset({"google", "circleci", "circleci_v2", "github", "azure", "file", "env", "env_path"}) @@ -198,11 +224,86 @@ def resolve_local_oidc_reference(value: str) -> str | UnresolvableSecretReferenc return secret -def _logged_unresolvable(reference: str, kind: str, reason: str) -> UnresolvableSecretReference: +def resolve_api_key_oidc_reference(value: str) -> str | UnresolvableSecretReference: + if not value.startswith(GOOGLE_OIDC_SECRET_PREFIX): + return resolve_local_oidc_reference(value) + try: + token: Final = _google_id_token( + value, value.removeprefix(GOOGLE_OIDC_SECRET_PREFIX), timeout=_google_metadata_timeout() + ) + except (httpx.HTTPError, httpx.InvalidURL, ValueError) as exc: + return _logged_unresolvable( + value, + "oidc/google", + _GOOGLE_TOKEN_UNAVAILABLE, + f"{type(exc).__name__}: {exc}", + retryable=_metadata_server_failed_to_answer(exc), + ) + if not token: + return _logged_unresolvable(value, "oidc/google", "it resolved to an empty value") + return token + + +def _metadata_server_failed_to_answer(exc: httpx.HTTPError | httpx.InvalidURL | ValueError) -> bool: + match exc: + case httpx.TransportError(): + return True + case GoogleMetadataStatusError(status_code=status_code): + return status_code >= 500 + case _: + return False + + +def _google_id_token(reference: str, audience: str, timeout: float | None) -> str: + cached: Final = cached_oidc_token(reference) + if cached is not None: + return cached + oidc_client: Final = _get_oidc_http_handler() + try: + # https://cloud.google.com/compute/docs/instances/verifying-instance-identity#request_signature + response: Final = oidc_client.get( + _google_metadata_identity_url(), + params={"audience": audience}, + headers={"Metadata-Flavor": "Google"}, + timeout=timeout, + ) + finally: + oidc_client.close() + if response.status_code != 200: + raise GoogleMetadataStatusError(response.status_code) + token: Final = response.text + ttl: Final = _oidc_token_cache_ttl(token, 3600 - 60) + if not token: + verbose_logger.warning("Google OIDC token for %s is empty; not caching it", _single_line(reference)) + elif ttl > 0: + oidc_cache.set_cache(key=reference, value=token, ttl=ttl) + else: + verbose_logger.warning( + "Google OIDC token for %s is already expired or expires within %ss; not caching it", + _single_line(reference), + _OIDC_TOKEN_EXPIRY_MARGIN_SECONDS, + ) + return token + + +def cached_oidc_token(reference: str) -> str | None: + cached: Final = oidc_cache.get_cache(key=reference) # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType] # DualCache.get_cache is untyped + return cached if isinstance(cached, str) and cached else None + + +def _logged_unresolvable( + reference: str, kind: str, reason: str, detail: str | None = None, retryable: bool = False +) -> UnresolvableSecretReference: verbose_logger.warning( - "Secret reference %s could not be resolved: %s", reference.replace("\r", "").replace("\n", ""), reason + "Secret reference %s could not be resolved: %s", + _single_line(reference), + reason if detail is None else f"{reason} ({_single_line(detail)})", ) - return UnresolvableSecretReference(kind=kind, reason=reason) + return UnresolvableSecretReference(kind=kind, reason=reason, retryable=retryable) + + +def _single_line(text: str) -> str: + return text.replace("\r", "").replace("\n", "") def _unresolvable_reason(kind: str, exc: OSError | ValueError) -> str: @@ -273,31 +374,7 @@ def get_secret( oidc_aud = "/".join(secret_name_split.split("/")[1:]) # TODO: Add caching for HTTP requests if oidc_provider == "google": - oidc_token = oidc_cache.get_cache(key=secret_name) - if oidc_token is not None: - return oidc_token - - oidc_client = _get_oidc_http_handler() - # https://cloud.google.com/compute/docs/instances/verifying-instance-identity#request_signature - response = oidc_client.get( - "http://metadata.google.internal/computeMetadata/v1/instance/service-accounts/default/identity", - params={"audience": oidc_aud}, - headers={"Metadata-Flavor": "Google"}, - ) - if response.status_code == 200: - oidc_token = response.text - ttl: Final = _oidc_token_cache_ttl(oidc_token, 3600 - 60) - if ttl > 0: - oidc_cache.set_cache(key=secret_name, value=oidc_token, ttl=ttl) - else: - verbose_logger.warning( - "Google OIDC token for %s is already expired or expires within %ss; not caching it", - secret_name, - _OIDC_TOKEN_EXPIRY_MARGIN_SECONDS, - ) - return oidc_token - else: - raise ValueError("Google OIDC provider failed") + return _google_id_token(secret_name, oidc_aud, timeout=None) elif oidc_provider == "circleci": # https://circleci.com/docs/openid-connect-tokens/ env_secret = os.getenv("CIRCLE_OIDC_TOKEN") diff --git a/litellm/utils.py b/litellm/utils.py index 5fe3e63f31a..8ea4f6ceda0 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -28,6 +28,7 @@ import textwrap import threading import time import traceback +from concurrent.futures import ThreadPoolExecutor from contextlib import nullcontext from dataclasses import dataclass, field, replace from functools import lru_cache, partial, wraps @@ -477,11 +478,14 @@ from litellm.llms.base_llm.evals.transformation import BaseEvalsAPIConfig from litellm.llms.base_llm.responses.transformation import BaseResponsesAPIConfig from litellm.llms.base_llm.skills.transformation import BaseSkillsAPIConfig from litellm.secret_managers.main import ( - LOCAL_OIDC_SECRET_PREFIXES, + API_KEY_OIDC_SECRET_PREFIXES, + GOOGLE_OIDC_SECRET_PREFIX, + OidcResolver, UnresolvableSecretReference, + cached_oidc_token, get_secret, oidc_reference_kind, - resolve_local_oidc_reference, + resolve_api_key_oidc_reference, ) from ._logging import is_debugging_on, verbose_logger @@ -784,27 +788,97 @@ def _named_credential_values(kwargs: Mapping[str, object]) -> Mapping[str, objec ) -def resolve_api_key_secret_reference( +async def aresolve_api_key_secret_reference( api_key: str | None, model: object, llm_provider: object, upstream: Mapping[str, object] ) -> str | None: if not isinstance(api_key, AdminApiKeySecretReference): return api_key - return _secret_behind_api_key_reference(api_key, model=model, llm_provider=llm_provider, upstream=upstream) + return await _resolved_off_the_event_loop( + partial(_secret_behind_api_key_reference, api_key, model, llm_provider, upstream), api_key + ) + + +_OIDC_TOKEN_FETCH_EXECUTOR: Final = ThreadPoolExecutor(thread_name_prefix="litellm-oidc-token-fetch") +_PASS_THROUGH_CALL_TYPES: Final = frozenset( + {CallTypes.allm_passthrough_route.value, CallTypes.llm_passthrough_route.value} +) +_Resolved: Final = TypeVar("_Resolved") + + +async def _resolved_off_the_event_loop(resolve: Callable[[OidcResolver], _Resolved], api_key: object) -> _Resolved: + if not isinstance(api_key, AdminApiKeySecretReference) or not api_key.startswith(GOOGLE_OIDC_SECRET_PREFIX): + return resolve(resolve_api_key_oidc_reference) + match cached_oidc_token(api_key): + case str() as token: + return resolve(lambda _reference: token) + case None: + _raise_if_the_reference_may_not_be_used(resolve) + fetched: Final = await asyncio.shield( + _google_token_fetch(asyncio.get_running_loop(), api_key).start_or_join() + ) + return resolve(lambda _reference: fetched) + + +def _raise_if_the_reference_may_not_be_used(resolve: Callable[[OidcResolver], object]) -> None: + _ = resolve(lambda _reference: "") + + +@dataclass(slots=True) +class _GoogleTokenFetch: + reference: str + in_flight: asyncio.Future[str | UnresolvableSecretReference] | None = None + + def start_or_join(self) -> asyncio.Future[str | UnresolvableSecretReference]: + if self.in_flight is not None: + return self.in_flight + caller_context: Final = contextvars.copy_context() + fetch: Final = asyncio.get_running_loop().run_in_executor( + _OIDC_TOKEN_FETCH_EXECUTOR, lambda: caller_context.run(resolve_api_key_oidc_reference, self.reference) + ) + fetch.add_done_callback(self._land) + self.in_flight = fetch + return fetch + + def _land(self, _fetch: asyncio.Future[str | UnresolvableSecretReference]) -> None: + self.in_flight = None + + +@lru_cache(maxsize=256) +def _google_token_fetch(loop: asyncio.AbstractEventLoop, reference: str) -> _GoogleTokenFetch: + return _GoogleTokenFetch(reference) def _secret_behind_api_key_reference( - reference: AdminApiKeySecretReference, model: object, llm_provider: object, upstream: Mapping[str, object] + reference: AdminApiKeySecretReference, + model: object, + llm_provider: object, + upstream: Mapping[str, object], + resolve_oidc: OidcResolver = resolve_api_key_oidc_reference, + call_type: str | None = None, ) -> str: if not reference.may_reach(upstream): raise _secret_reference_error(_unreachable_upstream_message(reference), llm_provider=llm_provider, model=model) - if not reference.startswith(LOCAL_OIDC_SECRET_PREFIXES): + if not reference.startswith(API_KEY_OIDC_SECRET_PREFIXES): raise _secret_reference_error( f"api_key is an {oidc_reference_kind(reference)} secret reference, a kind api_key does not support. " - "Use oidc/file/ or oidc/env/.", + "Use oidc/file/, oidc/env/ or oidc/google/.", llm_provider=llm_provider, model=model, ) - match resolve_local_oidc_reference(reference): + if call_type in _PASS_THROUGH_CALL_TYPES and reference.startswith(GOOGLE_OIDC_SECRET_PREFIX): + raise _secret_reference_error( + "api_key oidc/google secret reference could not be resolved: pass-through routes do not fetch Google ID " + "tokens. Use a static key, oidc/file/ or oidc/env/.", + llm_provider=llm_provider, + model=model, + ) + match resolve_oidc(reference): + case UnresolvableSecretReference(kind=kind, reason=reason, retryable=True): + raise ServiceUnavailableError( + message=f"api_key {kind} secret reference could not be resolved: {reason}", + llm_provider=llm_provider, + model=model, + ) case UnresolvableSecretReference(kind=kind, reason=reason): raise _secret_reference_error( f"api_key {kind} secret reference could not be resolved: {reason}", @@ -836,21 +910,39 @@ def with_named_credential(kwargs: Mapping[str, object]) -> Mapping[str, object]: return {**kwargs, **_named_credential_values(kwargs)} -def load_credentials_from_list(kwargs: dict[str, object]) -> None: +def load_credentials_from_list(kwargs: dict[str, object], call_type: str | None = None) -> None: """ Updates kwargs with the credentials if credential_name in kwarg, then swaps an admin-configured oidc/ api_key reference for the secret it points to, read now so a rotated token is used on the next call """ - kwargs.update(_credential_updates(kwargs)) + kwargs.update(_credential_updates(kwargs, _named_credential_values(kwargs), call_type=call_type)) -def _credential_updates(kwargs: Mapping[str, object]) -> Mapping[str, object]: +async def acredential_updates(kwargs: Mapping[str, object], call_type: str | None = None) -> Mapping[str, object]: credential_values: Final = _named_credential_values(kwargs) - api_key: Final = kwargs.get("api_key", credential_values.get("api_key")) + return await _resolved_off_the_event_loop( + partial(_credential_updates, kwargs, credential_values, call_type=call_type), + _api_key_after_credentials(kwargs, credential_values), + ) + + +def _api_key_after_credentials(kwargs: Mapping[str, object], credential_values: Mapping[str, object]) -> object: + return kwargs.get("api_key", credential_values.get("api_key")) + + +def _credential_updates( + kwargs: Mapping[str, object], + credential_values: Mapping[str, object], + resolve_oidc: OidcResolver = resolve_api_key_oidc_reference, + call_type: str | None = None, +) -> Mapping[str, object]: + api_key: Final = _api_key_after_credentials(kwargs, credential_values) if not isinstance(api_key, AdminApiKeySecretReference): return credential_values secret: Final = _secret_behind_api_key_reference( api_key, + resolve_oidc=resolve_oidc, + call_type=call_type, model=kwargs.get("model"), llm_provider=kwargs.get("custom_llm_provider"), upstream=MappingProxyType( @@ -1936,7 +2028,7 @@ def client(original_function): assert logging_obj is not None, "logging_obj should not be None after function_setup" ## LOAD CREDENTIALS - load_credentials_from_list(kwargs) + load_credentials_from_list(kwargs, call_type) from litellm.llms.github_copilot.per_user_auth import attach_github_copilot_user_session attach_github_copilot_user_session(kwargs) # pyright: ignore[reportUnknownArgumentType] # kwargs is the untyped request dict @@ -2230,7 +2322,7 @@ def client(original_function): kwargs["litellm_logging_obj"] = logging_obj ## LOAD CREDENTIALS - load_credentials_from_list(kwargs) + kwargs.update(await acredential_updates(kwargs, call_type)) # pyright: ignore[reportUnknownMemberType, reportUnknownArgumentType] # kwargs is the untyped request dict from litellm.llms.github_copilot.per_user_auth import aattach_github_copilot_user_session await aattach_github_copilot_user_session(kwargs) # pyright: ignore[reportUnknownArgumentType] # kwargs is the untyped request dict diff --git a/tests/integration/security/test_api_key_secret_reference.py b/tests/integration/security/test_api_key_secret_reference.py index 0b1187d2e0e..86a7d58e0d3 100644 --- a/tests/integration/security/test_api_key_secret_reference.py +++ b/tests/integration/security/test_api_key_secret_reference.py @@ -1,6 +1,9 @@ from __future__ import annotations +import base64 import json +import threading +import time import uuid from collections.abc import Callable, Iterator, Mapping, Sequence from dataclasses import dataclass @@ -46,7 +49,18 @@ _SEARCH_PATH: Final = f"/v1/vector_stores/{_VECTOR_STORE}/search" _BATCH_DEPLOYMENT: Final = "wif-vllm-batch" _BEDROCK_MODEL: Final = "bedrock/converse/us.anthropic.claude-sonnet-5-5" _PROVIDER_INPUT_TOKENS: Final = 7919 -_UNSUPPORTED_REFERENCE: Final = "oidc/google/https://example.invalid" +_UNSUPPORTED_REFERENCE: Final = "oidc/github/https://example.invalid" +_GOOGLE_MODEL: Final = "hosted_vllm/wif-google" +_GOOGLE_AUDIENCES: Final = MappingProxyType( + { + "config_inline": "https://vllm-inline.example.test", + "config_credential": "https://vllm-credential.example.test", + "db_credential": "https://vllm-db-credential.example.test", + } +) +_UNAVAILABLE_AUDIENCE: Final = "https://vllm-unavailable.example.test" +_HUNG_AUDIENCE: Final = "https://vllm-hung.example.test" +_IDENTITY_PATH: Final = "/computeMetadata/v1/instance/service-accounts/default/identity" _RUN_SECRETS_TOKEN: Final = "/run/secrets/wif-absent-token" _SOURCES: Final = ("config_credential", "config_inline", "db_credential", "db_inline") _UNREADABLE: Final = MappingProxyType( @@ -108,6 +122,45 @@ def _reference(path: Path) -> str: return f"oidc/file/{path}" +def _google_reference(audience: str) -> str: + return f"oidc/google/{audience}" + + +def _jwt_segment(part: Mapping[str, JsonValue]) -> str: + return base64.urlsafe_b64encode(json.dumps(part).encode()).decode().rstrip("=") + + +@dataclass(frozen=True, slots=True) +class _Google: + expires: int + + def token(self, audience: str) -> str: + claims: Final = _jwt_segment({"iss": "https://accounts.google.com", "aud": audience, "exp": self.expires}) + return f"{_jwt_segment({'alg': 'RS256', 'typ': 'JWT'})}.{claims}.signature" + + def bearers(self) -> frozenset[str]: + return frozenset(f"Bearer {self.token(audience)}" for audience in _GOOGLE_AUDIENCES.values()) + + +def _metadata_server(google: _Google) -> Callable[[Request], Reply]: + def respond(request: Request) -> Reply: + target: Final = httpx.URL(request.target) + audience: Final = target.params.get("audience") + if audience is None or (request.method, target.path, request.headers.get("metadata-flavor")) != ( + "GET", + _IDENTITY_PATH, + "Google", + ): + return Reply(status=404, body=b"not found", content_type="text/plain") + if audience == _UNAVAILABLE_AUDIENCE: + return Reply(status=503, body=b"metadata server unavailable", content_type="text/plain") + if audience == _HUNG_AUDIENCE: + return Reply(chunks=(b"never sent",), gate_before_first=threading.Event(), content_type="text/plain") + return Reply(body=google.token(audience).encode(), content_type="text/plain") + + return respond + + def _secrets(directory: Path) -> _Secrets: allowed: Final = directory / "secrets" outside_dir: Final = directory / "outside" @@ -169,9 +222,9 @@ def _served(request: Request) -> Reply: return openai_error(404) -def _provider(secrets: _Secrets) -> Callable[[Request], Reply]: +def _provider(secrets: _Secrets, google: _Google) -> Callable[[Request], Reply]: def respond(request: Request) -> Reply: - if request.headers.get("authorization") != f"Bearer {secrets.current()}": + if request.headers.get("authorization") not in {f"Bearer {secrets.current()}", *google.bearers()}: return openai_error(401) return _served(request) @@ -224,6 +277,27 @@ def _config(directory: Path, secrets: _Secrets, provider: str) -> Path: "aws_bedrock_runtime_endpoint": provider.removesuffix("/v1"), }, ), + _deployment( + "wif-google-inline", + { + "model": _GOOGLE_MODEL, + "api_key": _google_reference(_GOOGLE_AUDIENCES["config_inline"]), + "api_base": provider, + }, + ), + _deployment("wif-google-credential", _named(_GOOGLE_MODEL, "wif-google")), + _deployment( + "wif-google-unavailable", + {"model": _GOOGLE_MODEL, "api_key": _google_reference(_UNAVAILABLE_AUDIENCE), "api_base": provider}, + ), + _deployment( + "wif-google-unavailable-count", + {"model": _CHAT_MODEL, "api_key": _google_reference(_UNAVAILABLE_AUDIENCE), "api_base": provider}, + ), + _deployment( + "wif-google-hung", + {"model": _GOOGLE_MODEL, "api_key": _google_reference(_HUNG_AUDIENCE), "api_base": provider}, + ), _deployment( "wif-mistral-passthrough", { @@ -239,6 +313,7 @@ def _config(directory: Path, secrets: _Secrets, provider: str) -> Path: _credential("wif-outside", _reference(secrets.outside), provider), _credential("wif-unsupported", _UNSUPPORTED_REFERENCE, provider), _credential("wif-run-secrets", f"oidc/file/{_RUN_SECRETS_TOKEN}", provider), + _credential("wif-google", _google_reference(_GOOGLE_AUDIENCES["config_credential"]), provider), ], "vector_store_registry": [ { @@ -312,6 +387,8 @@ class _Rig: owned: OwnedProxy provider: Recorder intruder: Recorder + metadata: Recorder + google: _Google provider_base: str intruder_base: str secrets: _Secrets @@ -339,17 +416,24 @@ def _is_discovery(request: Request) -> bool: def rig(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Rig]: directory: Final = tmp_path_factory.mktemp("api_key_secret_reference").resolve() secrets: Final = _secrets(directory) + google: Final = _Google(expires=int(time.time()) + 3600) with ( gateway_from_environment() as gateway, - wire_server(_provider(secrets)) as provider_wire, + wire_server(_provider(secrets, google)) as provider_wire, wire_server(_intruder) as intruder_wire, + wire_server(_metadata_server(google)) as metadata_wire, ): provider: Final = Recorder(provider_wire) provider_base: Final = f"{provider_wire.url}/v1" with owned_proxy_process( gateway, directory, - {**secrets.overrides(), "MISTRAL_API_BASE": provider_wire.url}, + { + **secrets.overrides(), + "MISTRAL_API_BASE": provider_wire.url, + "GCE_METADATA_HOST": httpx.URL(metadata_wire.url).netloc.decode(), + "GCE_METADATA_TIMEOUT": "1", + }, config=_config(directory, secrets, provider_base), workers=2, healthcheck=WorkerHealthcheck.LENIENT, @@ -361,6 +445,8 @@ def rig(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Rig]: owned=owned, provider=provider, intruder=Recorder(intruder_wire), + metadata=Recorder(metadata_wire), + google=google, provider_base=provider_base, intruder_base=f"{intruder_wire.url}/v1", secrets=secrets, @@ -708,7 +794,92 @@ def test_team_row_naming_the_credential_sends_the_token_only_to_its_host(rig: _R def test_startup_discovery_sends_the_boot_token(rig: _Rig) -> None: sent: Final = frozenset(_bearers(tuple(request for request in rig.boot if _is_discovery(request)))) - assert sent == {f"Bearer {rig.boot_token}"}, sent + assert sent - rig.google.bearers() == {f"Bearer {rig.boot_token}"}, sent + + +def _discovery_bearers(rig: _Rig) -> frozenset[str | None]: + return frozenset(_bearers(tuple(request for request in rig.provider.requests() if _is_discovery(request)))) + + +def test_startup_discovery_sends_the_google_token(rig: _Rig) -> None: + expected: Final = frozenset( + f"Bearer {rig.google.token(_GOOGLE_AUDIENCES[source])}" for source in ("config_inline", "config_credential") + ) + sent: Final = eventually( + partial(_discovery_bearers, rig), lambda seen: expected <= seen, seconds=30, return_last_on_timeout=True + ) + assert expected <= sent, sent + + +def _google_model(rig: _Rig, scenario: Scenario, source: str) -> str: + match source: + case "config_inline": + return "wif-google-inline" + case "config_credential": + return "wif-google-credential" + case "db_credential": + credential: Final = f"db-google-{uuid.uuid4().hex}" + rig.owned.gateway.post( + "/credentials", + _credential(credential, _google_reference(_GOOGLE_AUDIENCES[source]), rig.provider_base), + ) + scenario.cleanups.callback(_delete_credential, rig.owned.gateway, credential) + name, identity = _created_model( + rig.owned.gateway, + scenario, + _deployment(f"db-google-{uuid.uuid4().hex}", _named(_GOOGLE_MODEL, credential)), + ) + eventually(partial(rig.owned.serves_only_on_every_worker, name, identity), bool, seconds=120) + return name + case _: + pytest.fail(f"unknown source {source}") + + +def _asked_audiences(rig: _Rig) -> frozenset[str | None]: + return frozenset(httpx.URL(request.target).params.get("audience") for request in rig.metadata.requests()) + + +@pytest.mark.parametrize("source", tuple(_GOOGLE_AUDIENCES)) +def test_google_reference_sends_the_metadata_server_token_for_its_audience(rig: _Rig, source: str) -> None: + audience: Final = _GOOGLE_AUDIENCES[source] + with rig.owned.gateway.scenario() as scenario: + marker: Final = _marker() + response: Final = _chat(rig.owned.gateway, _google_model(rig, scenario, source), marker) + sent: Final = _bearers(_carrying(rig.provider.requests(), marker)) + assert (response.status_code, sent) == (200, (f"Bearer {rig.google.token(audience)}",)), response.text + assert audience in _asked_audiences(rig) + + +def test_google_reference_fails_closed_when_the_metadata_server_fails(rig: _Rig) -> None: + marker: Final = _marker() + response: Final = _chat(rig.owned.gateway, "wif-google-unavailable", marker) + reached: Final = _carrying(rig.provider.requests(), marker) + assert (response.status_code, reached) == (503, ()), (response.text, _bearers(reached)) + assert "oidc/google secret reference could not be resolved" in response.text, response.text + assert _UNAVAILABLE_AUDIENCE.removeprefix("https://") not in response.text, response.text + assert _UNAVAILABLE_AUDIENCE in _asked_audiences(rig) + + +def test_google_reference_answers_503_when_the_metadata_server_never_finishes_its_reply(rig: _Rig) -> None: + marker: Final = _marker() + response: Final = _chat(rig.owned.gateway, "wif-google-hung", marker) + reached: Final = _carrying(rig.provider.requests(), marker) + assert (response.status_code, reached) == (503, ()), (response.text, _bearers(reached)) + assert "the metadata server did not return an ID token" in response.text, response.text + assert _HUNG_AUDIENCE in _asked_audiences(rig) + + +def test_input_token_count_answers_503_when_the_metadata_server_fails(rig: _Rig) -> None: + marker: Final = _marker() + response: Final = rig.owned.gateway.request( + "POST", + "/v1/responses/input_tokens", + {"model": "wif-google-unavailable-count", "input": f"say {marker}"}, + headers=_CLOSE, + ) + reached: Final = _carrying(rig.provider.requests(), marker) + assert (response.status_code, reached) == (503, ()), (response.text, _bearers(reached)) + assert "oidc/google secret reference could not be resolved" in response.text, response.text def _side_call(rig: _Rig, seam: str, marker: str) -> tuple[httpx.Response, Callable[[Request], bool]]: diff --git a/tests/unit/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py b/tests/unit/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py index 9c50a6d4df6..ab4f3f7a675 100644 --- a/tests/unit/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py +++ b/tests/unit/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py @@ -485,7 +485,7 @@ def test_oidc_env_path_reference_is_not_resolved(tmp_path, monkeypatch): ) -def test_network_backed_oidc_reference_is_not_fetched_inline(monkeypatch): +def test_oidc_google_reference_is_refused_without_a_metadata_fetch(monkeypatch): def fail(*args, **kwargs): raise AssertionError("network-backed oidc reference was resolved inside the request path") @@ -494,11 +494,24 @@ def test_network_backed_oidc_reference_is_not_fetched_inline(monkeypatch): lambda name, *a, **k: "oidc/google/https://example.com" if name == "OPENAI_API_KEY" else fail(), ) - assert ( + with pytest.raises(HTTPException) as excinfo: _passthrough_router(None).get_credentials(custom_llm_provider="openai", region_name=None) - == "oidc/google/https://example.com" + + assert excinfo.value.status_code == 401 + assert "pass-through credential oidc/google" in str(excinfo.value.detail) + assert "example.com" not in str(excinfo.value.detail) + + +def test_deployment_oidc_google_reference_is_not_sent_upstream(monkeypatch): + llm_router = litellm.Router( + model_list=[_flagged_deployment("openai/gpt-4o", api_key="oidc/google/https://vllm.example.test")] ) + with pytest.raises(HTTPException) as excinfo: + _passthrough_router(llm_router).get_credentials(custom_llm_provider="openai", region_name=None) + + assert excinfo.value.status_code == 401 + @pytest.fixture() async def _drain_logging_worker(): diff --git a/tests/unit/proxy/test_proxy_token_counter.py b/tests/unit/proxy/test_proxy_token_counter.py index b21730c64f6..d0cf7be794a 100644 --- a/tests/unit/proxy/test_proxy_token_counter.py +++ b/tests/unit/proxy/test_proxy_token_counter.py @@ -279,6 +279,39 @@ async def test_token_counter_reads_an_oidc_api_key_only_when_it_calls_the_provid assert str(tmp_path) not in excinfo.value.message +@pytest.mark.asyncio +async def test_token_counter_answers_503_when_the_metadata_server_behind_an_oidc_google_api_key_fails( + monkeypatch: pytest.MonkeyPatch, +): + audience: Final = f"https://count.test/{uuid.uuid4().hex}" + monkeypatch.setattr( + litellm.proxy.proxy_server, + "llm_router", + Router( + model_list=[ + { + "model_name": "claude", + "litellm_params": {"model": "anthropic/claude-sonnet-5-5", "api_key": f"oidc/google/{audience}"}, + } + ] + ), + ) + + with respx.mock(assert_all_called=True) as upstream: + upstream.get(url__regex=r"/computeMetadata/v1/instance/service-accounts/default/identity").respond( + 500, text="metadata server unavailable" + ) + with pytest.raises(ProxyException) as excinfo: + await token_counter( + request=TokenCountRequest(model="claude", messages=[{"role": "user", "content": "hello"}]), + call_endpoint=True, + ) + + assert excinfo.value.code == "503" + assert "oidc/google secret reference could not be resolved" in excinfo.value.message + assert audience.removeprefix("https://") not in excinfo.value.message + + @pytest.mark.asyncio async def test_native_openai_responses_count_sends_the_token_an_oidc_api_key_points_to( tmp_path: Path, monkeypatch: pytest.MonkeyPatch diff --git a/tests/unit/secret_managers/test_secret_managers_main.py b/tests/unit/secret_managers/test_secret_managers_main.py index 862658d9169..a2e5b498f34 100644 --- a/tests/unit/secret_managers/test_secret_managers_main.py +++ b/tests/unit/secret_managers/test_secret_managers_main.py @@ -4,6 +4,7 @@ import logging import os import threading import time +import uuid from collections.abc import Callable from concurrent.futures import ThreadPoolExecutor from queue import SimpleQueue @@ -11,6 +12,7 @@ from typing import Final from unittest.mock import Mock, patch import pytest +import respx import litellm from litellm._logging import redact_internal_details_from_client_message @@ -37,7 +39,7 @@ class MockHTTPHandler: self.text = "mocked_token" self.json_data = {"value": "mocked_token"} - def get(self, url, params=None, headers=None): + def get(self, url, params=None, headers=None, timeout=None): # Store params for audience verification self.last_params = params logger.debug( @@ -49,6 +51,9 @@ class MockHTTPHandler: mock_response.json.return_value = self.json_data return mock_response + def close(self): + pass + @pytest.fixture def mock_oidc_cache(): @@ -229,6 +234,41 @@ def test_oidc_google_failure(): get_secret(secret_name) + +def test_oidc_google_fetches_the_token_from_the_metadata_host_gce_metadata_host_names( + monkeypatch: pytest.MonkeyPatch, +) -> None: + monkeypatch.setenv("GCE_METADATA_HOST", "127.0.0.1:8930") + audience: Final = f"https://vllm.example.test/{uuid.uuid4().hex}" + + with respx.mock(assert_all_called=True) as metadata: + metadata.get( + "http://127.0.0.1:8930/computeMetadata/v1/instance/service-accounts/default/identity", + params={"audience": audience}, + headers={"Metadata-Flavor": "Google"}, + ).respond(200, text="metadata-host-token") + token: Final = get_secret(f"oidc/google/{audience}") + + assert token == "metadata-host-token" + +def test_oidc_google_secret_outside_a_model_api_key_keeps_its_metadata_wait_when_gce_metadata_timeout_is_set( + monkeypatch: pytest.MonkeyPatch, +) -> None: + monkeypatch.setenv("GCE_METADATA_TIMEOUT", "1") + audience: Final = f"https://sts.example.test/{uuid.uuid4().hex}" + + with respx.mock(assert_all_called=True) as metadata: + identity: Final = metadata.get( + "http://metadata.google.internal/computeMetadata/v1/instance/service-accounts/default/identity", + params={"audience": audience}, + ).respond(200, text="web-identity-token") + token: Final = get_secret(f"oidc/google/{audience}") + waits: Final = identity.calls.last.request.extensions["timeout"] + + assert token == "web-identity-token" + assert waits == {"connect": 5.0, "read": 600.0, "write": 600.0, "pool": 600.0} + + def test_oidc_circleci_success(monkeypatch): monkeypatch.setenv("CIRCLE_OIDC_TOKEN", "circleci_token") diff --git a/tests/unit/test_utils.py b/tests/unit/test_utils.py index 2870016a5e9..cd56b6e1b46 100644 --- a/tests/unit/test_utils.py +++ b/tests/unit/test_utils.py @@ -5,14 +5,18 @@ import copy import contextlib import contextvars import io +import itertools import json import logging import os import queue import threading +import time +import uuid from collections.abc import Callable, Iterator, Mapping from concurrent.futures import Future, ThreadPoolExecutor from datetime import datetime, timedelta, timezone +from functools import partial from pathlib import Path, PurePath from typing import Final, SupportsIndex, cast from unittest import mock @@ -79,8 +83,10 @@ from litellm.types.utils import ( all_litellm_params, bedrock_batch_litellm_params, ) +from litellm.secret_managers.main import oidc_cache from litellm.types.videos.main import VideoObject from litellm.utils import ( + acredential_updates, _invalidate_model_cost_lowercase_map, add_custom_logger_callback_to_specific_event, check_valid_key, @@ -7320,7 +7326,6 @@ def test_unresolvable_oidc_api_key_fails_closed_naming_the_reason_without_its_ta ("reference", "kind"), [ ("oidc/env_path/MISTRAL_SA_TOKEN_PATH", "oidc/env_path"), - ("oidc/google/https://api.mistral.ai", "oidc/google"), ("oidc/github/mistral", "oidc/github"), ("oidc/env_MISTRAL_SA_TOKEN_PATH", "oidc/"), ("oidc/LIT-UNKNOWN-KIND/MISTRAL_SA_TOKEN_PATH", "oidc/"), @@ -7341,6 +7346,7 @@ def test_oidc_api_key_of_a_kind_api_key_cannot_resolve_fails_closed_as_an_authen assert not route.called assert excinfo.value.status_code == 401 assert f"{kind} secret reference" in excinfo.value.message + assert "oidc/google/" in excinfo.value.message assert reference.removeprefix(kind).lstrip("/") not in excinfo.value.message assert "MISTRAL_SA_TOKEN_PATH" not in excinfo.value.message @@ -7356,7 +7362,7 @@ async def test_router_fails_over_from_a_deployment_whose_oidc_api_key_kind_canno "model_name": "devstral", "litellm_params": { "model": "mistral/devstral-latest", - "api_key": "oidc/google/https://api.mistral.ai", + "api_key": "oidc/github/mistral", "order": 1, }, }, @@ -7395,6 +7401,372 @@ def test_an_oidc_api_key_sent_with_the_call_goes_upstream_as_written( assert sent == f"Bearer {reference}" +_GOOGLE_METADATA_IDENTITY: Final = ( + "http://metadata.google.internal/computeMetadata/v1/instance/service-accounts/default/identity" +) +_VLLM_BASE: Final = "https://vllm.example.test/v1" +_VLLM_CHAT_URL: Final = f"{_VLLM_BASE}/chat/completions" +_VLLM_MODEL: Final = "hosted_vllm/qwen3-0.6b" +_VLLM_REPLY: Final = {**_MISTRAL_REPLY, "model": "qwen3-0.6b"} +_CLOCK_START: Final = 1_900_000_000.0 +_GOOGLE_TOKEN_LIFETIME: Final = 3600 + + +def _jwt_segment(claims: Mapping[str, object]) -> str: + return base64.urlsafe_b64encode(json.dumps(claims).encode()).decode().rstrip("=") + + +def _google_id_token(audience: str, issued_at: float) -> str: + claims: Final = {"aud": audience, "iat": int(issued_at), "exp": int(issued_at) + _GOOGLE_TOKEN_LIFETIME} + return f"{_jwt_segment({'alg': 'RS256', 'typ': 'JWT'})}.{_jwt_segment(claims)}.signature" + + +def _google_reference() -> tuple[str, str]: + audience: Final = f"https://vllm.example.test/{uuid.uuid4().hex}" + return audience, f"oidc/google/{audience}" + + +def _metadata_identity(fetched_on: queue.SimpleQueue[int], request: httpx.Request) -> httpx.Response: + fetched_on.put(threading.get_ident()) + if request.headers.get("Metadata-Flavor") != "Google": + return httpx.Response(403) + return httpx.Response(200, text=_google_id_token(request.url.params["audience"], time.time())) + + +def _vllm_credential(api_key: str) -> CredentialItem: + return CredentialItem( + credential_name="vllm-sa", credential_values={"api_key": api_key, "api_base": _VLLM_BASE}, credential_info={} + ) + + +async def _ping_vllm_through(call_path: str, credential_name: str) -> None: + if call_path == "async": + await litellm.acompletion(model=_VLLM_MODEL, messages=_PING, litellm_credential_name=credential_name) + return + litellm.completion(model=_VLLM_MODEL, messages=_PING, litellm_credential_name=credential_name) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("call_path", ["sync", "async"]) +async def test_named_credential_oidc_google_api_key_sends_the_metadata_token_and_refreshes_it_before_it_expires( + monkeypatch: pytest.MonkeyPatch, call_path: str +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + monkeypatch.setattr(time, "time", lambda: _CLOCK_START) + audience, reference = _google_reference() + monkeypatch.setattr(litellm, "credential_list", [_vllm_credential(reference)]) + refreshed_at: Final = _CLOCK_START + _GOOGLE_TOKEN_LIFETIME - 59 + + with respx.mock(assert_all_called=True) as upstream: + metadata: Final = upstream.get(_GOOGLE_METADATA_IDENTITY).mock( + side_effect=partial(_metadata_identity, queue.SimpleQueue()) + ) + chat: Final = upstream.post(_VLLM_CHAT_URL).respond(200, json=_VLLM_REPLY) + await _ping_vllm_through(call_path, "vllm-sa") + await _ping_vllm_through(call_path, "vllm-sa") + fetches_within_lifetime: Final = metadata.call_count + monkeypatch.setattr(time, "time", lambda: refreshed_at) + await _ping_vllm_through(call_path, "vllm-sa") + sent: Final = [call.request.headers["Authorization"] for call in chat.calls] + audiences: Final = [call.request.url.params["audience"] for call in metadata.calls] + + first_token: Final = f"Bearer {_google_id_token(audience, _CLOCK_START)}" + assert sent == [first_token, first_token, f"Bearer {_google_id_token(audience, refreshed_at)}"] + assert (fetches_within_lifetime, audiences) == (1, [audience, audience]) + assert litellm.credential_list[0].credential_values["api_key"] == reference + + +@pytest.mark.asyncio +async def test_router_fetches_an_oidc_google_api_key_token_off_the_event_loop_and_keeps_the_reference( + monkeypatch: pytest.MonkeyPatch, +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + monkeypatch.setattr(time, "time", lambda: _CLOCK_START) + audience, reference = _google_reference() + fetched_on: Final[queue.SimpleQueue[int]] = queue.SimpleQueue() + router: Final = litellm.Router( + model_list=[ + { + "model_name": "vllm", + "litellm_params": {"model": _VLLM_MODEL, "api_base": _VLLM_BASE, "api_key": reference}, + } + ] + ) + + try: + with respx.mock(assert_all_called=True) as upstream: + upstream.get(_GOOGLE_METADATA_IDENTITY).mock(side_effect=partial(_metadata_identity, fetched_on)) + chat: Final = upstream.post(_VLLM_CHAT_URL).respond(200, json=_VLLM_REPLY) + await router.acompletion(model="vllm", messages=_PING) + sent: Final = chat.calls.last.request.headers["Authorization"] + finally: + router.discard() + + assert sent == f"Bearer {_google_id_token(audience, _CLOCK_START)}" + assert fetched_on.qsize() == 1 + assert fetched_on.get_nowait() != threading.get_ident() + assert router.get_model_list(model_name="vllm")[0]["litellm_params"]["api_key"] == reference + + +@pytest.mark.asyncio +@pytest.mark.parametrize("call_path", ["sync", "async"]) +@pytest.mark.parametrize( + ("metadata_failure", "status_code"), + [ + (lambda: httpx.Response(500, text="metadata server unavailable"), 503), + (lambda: httpx.Response(403, text="Missing Metadata-Flavor:Google header"), 401), + (lambda: httpx.Response(200, text=""), 401), + (lambda: httpx.ConnectError("[Errno -2] Name or service not known"), 503), + (lambda: httpx.ReadTimeout("timed out"), 503), + (lambda: httpx.RemoteProtocolError("Server disconnected without sending a response."), 503), + ], + ids=["server error", "refused", "empty token", "not on GCP", "timeout", "disconnected"], +) +async def test_oidc_google_api_key_fails_closed_when_the_metadata_server_fails_and_recovers_on_the_next_call( + monkeypatch: pytest.MonkeyPatch, + caplog: pytest.LogCaptureFixture, + call_path: str, + metadata_failure: Callable[[], httpx.Response | httpx.HTTPError], + status_code: int, +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + monkeypatch.setattr(time, "time", lambda: _CLOCK_START) + audience, reference = _google_reference() + monkeypatch.setattr(litellm, "credential_list", [_vllm_credential(reference)]) + recovered_token: Final = _google_id_token(audience, _CLOCK_START) + + with respx.mock(assert_all_called=True) as upstream: + upstream.get(_GOOGLE_METADATA_IDENTITY).mock( + side_effect=[metadata_failure(), httpx.Response(200, text=recovered_token)] + ) + chat: Final = upstream.post(_VLLM_CHAT_URL).respond(200, json=_VLLM_REPLY) + with caplog.at_level(logging.WARNING, logger="LiteLLM"), pytest.raises( + (litellm.AuthenticationError, litellm.ServiceUnavailableError) + ) as excinfo: + await _ping_vllm_through(call_path, "vllm-sa") + sent_while_failing: Final = chat.call_count + await _ping_vllm_through(call_path, "vllm-sa") + sent_after_recovery: Final = [call.request.headers["Authorization"] for call in chat.calls] + + assert excinfo.value.status_code == status_code + assert "api_key oidc/google secret reference could not be resolved" in excinfo.value.message + assert audience.removeprefix("https://") not in excinfo.value.message + assert reference in caplog.text + assert sent_while_failing == 0 + assert sent_after_recovery == [f"Bearer {recovered_token}"] + + +@pytest.mark.asyncio +async def test_an_oidc_google_api_key_token_cached_when_the_call_starts_is_sent_though_it_expires_mid_call( + monkeypatch: pytest.MonkeyPatch, +) -> None: + monkeypatch.setattr(time, "time", lambda: _CLOCK_START) + audience, reference = _google_reference() + cached_token: Final = _google_id_token(audience, _CLOCK_START) + oidc_cache.set_cache(key=reference, value=cached_token, ttl=60) + clock: Final = itertools.chain([_CLOCK_START + 60], itertools.repeat(_CLOCK_START + 61)) + monkeypatch.setattr(time, "time", lambda: next(clock)) + monkeypatch.setattr(litellm, "credential_list", [_vllm_credential(reference)]) + fetched_on: Final[queue.SimpleQueue[int]] = queue.SimpleQueue() + + with respx.mock(assert_all_called=False) as upstream: + upstream.get(_GOOGLE_METADATA_IDENTITY).mock(side_effect=partial(_metadata_identity, fetched_on)) + updates: Final = await acredential_updates({"model": _VLLM_MODEL, "litellm_credential_name": "vllm-sa"}) + + assert updates["api_key"] == cached_token + assert fetched_on.empty() + + +@pytest.mark.asyncio +async def test_a_pending_oidc_google_metadata_fetch_leaves_the_default_executor_free( + monkeypatch: pytest.MonkeyPatch, +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + _, reference = _google_reference() + monkeypatch.setattr(litellm, "credential_list", [_vllm_credential(reference)]) + loop: Final = asyncio.get_running_loop() + loop.set_default_executor(ThreadPoolExecutor(max_workers=1)) + fetch_started: Final[queue.SimpleQueue[int]] = queue.SimpleQueue() + release: Final = threading.Event() + + def pending_metadata(_request: httpx.Request) -> httpx.Response: + fetch_started.put(threading.get_ident()) + release.wait(timeout=30) + return httpx.Response(500, text="metadata server unavailable") + + with ThreadPoolExecutor(max_workers=1) as observer, respx.mock(assert_all_called=False) as upstream: + upstream.get(_GOOGLE_METADATA_IDENTITY).mock(side_effect=pending_metadata) + chat: Final = upstream.post(_VLLM_CHAT_URL).respond(200, json=_VLLM_REPLY) + call: Final = asyncio.ensure_future(_ping_vllm_through("async", "vllm-sa")) + try: + await asyncio.wait_for(loop.run_in_executor(observer, fetch_started.get), timeout=30) + other_work: Final = await asyncio.wait_for(loop.run_in_executor(None, lambda: "done"), timeout=30) + finally: + release.set() + with pytest.raises(litellm.ServiceUnavailableError) as excinfo: + await call + + assert other_work == "done" + assert excinfo.value.status_code == 503 + assert chat.call_count == 0 + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "metadata_answer", + [lambda token: httpx.Response(200, text=token), lambda _token: httpx.Response(500, text="metadata server down")], + ids=["token", "outage"], +) +async def test_concurrent_calls_that_miss_the_oidc_google_token_cache_share_one_metadata_fetch( + monkeypatch: pytest.MonkeyPatch, metadata_answer: Callable[[str], httpx.Response] +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + monkeypatch.setattr(time, "time", lambda: _CLOCK_START) + audience, reference = _google_reference() + monkeypatch.setattr(litellm, "credential_list", [_vllm_credential(reference)]) + token: Final = _google_id_token(audience, _CLOCK_START) + fetch_started: Final[queue.SimpleQueue[int]] = queue.SimpleQueue() + release: Final = threading.Event() + + def held_metadata(_request: httpx.Request) -> httpx.Response: + fetch_started.put(threading.get_ident()) + release.wait(timeout=30) + return metadata_answer(token) + + with ThreadPoolExecutor(max_workers=1) as observer, respx.mock(assert_all_called=False) as upstream: + metadata: Final = upstream.get(_GOOGLE_METADATA_IDENTITY).mock(side_effect=held_metadata) + chat: Final = upstream.post(_VLLM_CHAT_URL).respond(200, json=_VLLM_REPLY) + calls: Final = [asyncio.ensure_future(_ping_vllm_through("async", "vllm-sa")) for _ in range(24)] + loop: Final = asyncio.get_running_loop() + try: + await asyncio.wait_for(loop.run_in_executor(observer, fetch_started.get), timeout=30) + fetch_started_while_held: Final = await loop.run_in_executor( + observer, partial(_next_fetch_within, fetch_started, 0.5) + ) + finally: + release.set() + outcomes: Final = await asyncio.gather(*calls, return_exceptions=True) + sent: Final = [call.request.headers["Authorization"] for call in chat.calls] + + assert fetch_started_while_held is None + if metadata.calls[0].response.status_code == 200: + assert metadata.call_count == 1 + assert outcomes == [None] * 24 + assert sent == [f"Bearer {token}"] * 24 + else: + assert [getattr(outcome, "status_code", None) for outcome in outcomes] == [503] * 24 + assert sent == [] + + +def _next_fetch_within(fetches: queue.SimpleQueue[int], seconds: float) -> int | None: + try: + return fetches.get(timeout=seconds) + except queue.Empty: + return None + + +@pytest.mark.asyncio +async def test_a_call_an_oidc_google_api_key_may_not_go_to_is_refused_without_a_metadata_fetch( + monkeypatch: pytest.MonkeyPatch, +) -> None: + _, reference = _google_reference() + monkeypatch.setattr(litellm, "credential_list", [_vllm_credential(reference)]) + + with respx.mock(assert_all_called=False) as upstream: + metadata: Final = upstream.get(_GOOGLE_METADATA_IDENTITY).respond(500) + with pytest.raises(litellm.AuthenticationError) as excinfo: + await acredential_updates( + { + "model": _VLLM_MODEL, + "litellm_credential_name": "vllm-sa", + "api_base": "https://elsewhere.example.test/v1", + } + ) + + assert "only to that credential's own api_base" in excinfo.value.message + assert metadata.call_count == 0 + + +async def _pass_through_chat_call(call_path: str, reference: str) -> None: + pass_through: Final = {"method": "POST", "endpoint": "v1/chat/completions", "json": {"messages": _PING}} + if call_path == "sync": + litellm.llm_passthrough_route(model=_VLLM_MODEL, litellm_credential_name="vllm-sa", **pass_through) + return + router: Final = litellm.Router( + model_list=[ + { + "model_name": "vllm-model", + "litellm_params": {"model": _VLLM_MODEL, "api_base": _VLLM_BASE, "api_key": reference}, + } + ], + num_retries=0, + ) + await router.allm_passthrough_route(model="vllm-model", **pass_through) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("call_path", ["sync", "async router"]) +async def test_a_pass_through_call_with_an_oidc_google_api_key_is_refused_without_a_metadata_fetch( + monkeypatch: pytest.MonkeyPatch, call_path: str +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + monkeypatch.setattr(time, "time", lambda: _CLOCK_START) + audience, reference = _google_reference() + monkeypatch.setattr(litellm, "credential_list", [_vllm_credential(reference)]) + + with respx.mock(assert_all_called=False) as upstream: + metadata: Final = upstream.get(_GOOGLE_METADATA_IDENTITY).respond( + 200, text=_google_id_token(audience, _CLOCK_START) + ) + vllm: Final = upstream.route(host="vllm.example.test").respond(200, json=_VLLM_REPLY) + with pytest.raises(litellm.AuthenticationError) as excinfo: + await _pass_through_chat_call(call_path, reference) + + assert "pass-through routes do not fetch Google ID tokens" in excinfo.value.message + assert metadata.call_count == 0 + assert vllm.call_count == 0 + + +@pytest.mark.parametrize( + ("configured", "bound"), + [(None, 3.0), ("1.5", 1.5), ("three seconds", 3.0), ("0", 3.0), ("-2", 3.0)], + ids=["unset", "configured", "malformed", "zero", "negative"], +) +def test_oidc_google_api_key_metadata_fetch_waits_at_most_gce_metadata_timeout_seconds( + monkeypatch: pytest.MonkeyPatch, configured: str | None, bound: float +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + if configured is None: + monkeypatch.delenv("GCE_METADATA_TIMEOUT", raising=False) + else: + monkeypatch.setenv("GCE_METADATA_TIMEOUT", configured) + monkeypatch.setattr(time, "time", lambda: _CLOCK_START) + audience, reference = _google_reference() + monkeypatch.setattr(litellm, "credential_list", [_vllm_credential(reference)]) + token: Final = _google_id_token(audience, _CLOCK_START) + + with respx.mock(assert_all_called=True) as upstream: + metadata: Final = upstream.get(_GOOGLE_METADATA_IDENTITY).respond(200, text=token) + chat: Final = upstream.post(_VLLM_CHAT_URL).respond(200, json=_VLLM_REPLY) + litellm.completion(model=_VLLM_MODEL, messages=_PING, litellm_credential_name="vllm-sa") + sent: Final = chat.calls.last.request.headers["Authorization"] + waits: Final = metadata.calls.last.request.extensions["timeout"] + + assert sent == f"Bearer {token}" + assert waits == {"connect": bound, "read": bound, "write": bound, "pool": bound} + + +def test_an_oidc_google_api_key_sent_with_the_call_goes_upstream_as_written_without_a_metadata_fetch() -> None: + _, reference = _google_reference() + + with respx.mock(assert_all_called=True) as upstream: + chat: Final = upstream.post(_VLLM_CHAT_URL).respond(200, json=_VLLM_REPLY) + litellm.completion(model=_VLLM_MODEL, api_base=_VLLM_BASE, messages=_PING, api_key=reference) + sent: Final = chat.calls.last.request.headers["Authorization"] + + assert sent == f"Bearer {reference}" + + @pytest.mark.asyncio @pytest.mark.parametrize( ("model_info", "expected"), From 06cdee91daf61c47bdf7cef1497a6bb001f23831 Mon Sep 17 00:00:00 2001 From: paveln83 <349044532+paveln83@users.noreply.github.com> Date: Thu, 29 Oct 2026 08:46:49 -0700 Subject: [PATCH 3/7] fix(guardrails): scan and rewrite streamed /v1/completions under post_call guardrails (#49035) * feat(guardrails): scan and rewrite streamed /v1/completions in post_call pipelines Gives OpenAITextCompletionHandler a process_output_streaming_response that assembles the buffered TextCompletionResponse chunks per choice index at end of stream, runs the guardrail over the assembled text, and writes text rewrites back into the first text-carrying chunk of each choice. Tests and live proof still pending. * fix(guardrails): scan and rewrite streamed /v1/completions under post_call guardrails Assemble each choice's streamed text once the stream ends, run the guardrail on it, write a mask back into the buffered chunks, and end the stream with the guardrail's error on a block. User key metadata now goes through the shared metadata bucket so applied guardrails reach spend logs. * test(guardrails): cover streamed /v1/completions under post_call guardrails * fix(guardrails): keep Gray Swan key metadata, scan a cut-short text stream, and pass non-chunk frames through * fix(guardrails): replace blocked text completion streams like chat, keep Gray Swan's vendor payload, scan text streams in the content filter * fix(guardrails): release an unfinished choice's held text and omit usage the upstream never sent The content filter's streaming hook held back the end of a choice until its finish_reason, which litellm sends only for the first choice of an n>1 stream, so choice 1 lost its last characters on text and chat streams. A closing chunk now releases every unfinished choice when the stream ends. A blocked text stream's closing chunk reported 0/0 usage when the upstream sent none; it now carries usage only when the upstream reported it. * fix(guardrails): send no usage on the content filter's end-of-stream chunk * fix(guardrails): keep the usage chunk last and log the masked text of streamed text completions * fix(guardrails): stream Responses API events through the content filter as they arrive * fix(guardrails): treat a null text completion choice index as choice zero when assembling streams --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../litellm_core_utils/served_output_texts.py | 60 +++-- .../guardrail_translation/handler.py | 198 +++++++++++++++- litellm/proxy/common_request_processing.py | 20 +- .../guardrail_hooks/grayswan/grayswan.py | 36 ++- .../litellm_content_filter/content_filter.py | 194 ++++++++++----- ...ext_completion_stream_guardrail_rebuild.py | 95 +++++++- .../test_served_output_texts.py | 35 ++- .../test_text_completion_guardrail_handler.py | 223 +++++++++++++++++- .../content_filter/test_content_filter.py | 190 ++++++++++++++- .../guardrail_hooks/test_grayswan.py | 28 +++ .../test_deferred_guardrail_logging.py | 21 ++ .../proxy_logging/test_guardrail_pipeline.py | 156 +++++++++++- 12 files changed, 1132 insertions(+), 124 deletions(-) diff --git a/litellm/litellm_core_utils/served_output_texts.py b/litellm/litellm_core_utils/served_output_texts.py index 854785f304b..b19973e06c3 100644 --- a/litellm/litellm_core_utils/served_output_texts.py +++ b/litellm/litellm_core_utils/served_output_texts.py @@ -4,12 +4,13 @@ post-call guardrail rewrote rather than the provider response the proxy assemble from __future__ import annotations from collections.abc import Sequence +from itertools import chain from typing import Final, Literal -from pydantic import BaseModel, Field, TypeAdapter, ValidationError +from pydantic import BaseModel, ConfigDict, Field, TypeAdapter, ValidationError from litellm.types.llms.base import LiteLLMBaseModel -from litellm.types.utils import ModelResponse, ModelResponseStream +from litellm.types.utils import ModelResponse, ModelResponseStream, TextCompletionResponse SERVED_OUTPUT_TEXTS_KEY: Final = "served_output_texts" @@ -44,6 +45,19 @@ class _ChatChoices(LiteLLMBaseModel): choices: list[object] +class _TextCompletionChoice(LiteLLMBaseModel): + model_config = ConfigDict(frozen=True) + + index: int | None = None + text: str | None = None + + +class _TextCompletionChunk(LiteLLMBaseModel): + model_config = ConfigDict(frozen=True) + + choices: tuple[_TextCompletionChoice, ...] + + def _as_json_object(response: object) -> dict[str, object] | None: candidate: Final = response.model_dump() if isinstance(response, BaseModel) else response try: @@ -108,8 +122,13 @@ def served_output_texts(response: object) -> ServedTexts | None: def served_stream_output_texts(chunks: Sequence[object]) -> ServedTexts | None: - if chunks and all(isinstance(chunk, ModelResponseStream) for chunk in chunks): - return _chat_stream_texts(tuple(chunk for chunk in chunks if isinstance(chunk, ModelResponseStream))) + chat_chunks: Final = tuple(chunk for chunk in chunks if isinstance(chunk, ModelResponseStream)) + if chunks and len(chat_chunks) == len(chunks): + return _indexed_texts_by_choice(tuple(chain.from_iterable(_chat_chunk_texts(chunk) for chunk in chat_chunks))) + if chunks and all(isinstance(chunk, TextCompletionResponse) for chunk in chunks): + return _indexed_texts_by_choice( + tuple(chain.from_iterable(_text_completion_chunk_texts(chunk) for chunk in chunks)) + ) from litellm.llms.anthropic.messages.stream_assembly import ( assemble_anthropic_sse_stream, is_anthropic_sse_stream, @@ -121,19 +140,26 @@ def served_stream_output_texts(chunks: Sequence[object]) -> ServedTexts | None: return _chat_texts(assembled) if assembled is not None else None -def _chat_stream_choice_text(chunks: Sequence[ModelResponseStream], index: int) -> str | None: - contents: Final = tuple( - content - for chunk in chunks +def _chat_chunk_texts(chunk: ModelResponseStream) -> tuple[tuple[int, str | None], ...]: + return tuple( + (choice.index or 0, content if isinstance(content := choice.delta.content, str) else None) for choice in chunk.choices - if (choice.index or 0) == index and isinstance(content := choice.delta.content, str) ) - return "".join(contents) if contents else None -def _chat_stream_texts(chunks: Sequence[ModelResponseStream]) -> ServedTexts | None: - choice_count: Final = max(((choice.index or 0) + 1 for chunk in chunks for choice in chunk.choices), default=0) - texts: Final = tuple(_chat_stream_choice_text(chunks, index) for index in range(choice_count)) +def _text_completion_chunk_texts(chunk: object) -> tuple[tuple[int, str | None], ...]: + parsed: Final = _TextCompletionChunk.model_validate(_as_json_object(chunk)) + return tuple((choice.index or 0, choice.text) for choice in parsed.choices) + + +def _joined_choice_text(indexed_texts: Sequence[tuple[int, str | None]], index: int) -> str | None: + texts: Final = tuple(text for text_index, text in indexed_texts if text_index == index and text is not None) + return "".join(texts) if texts else None + + +def _indexed_texts_by_choice(indexed_texts: Sequence[tuple[int, str | None]]) -> ServedTexts | None: + choice_count: Final = max((index + 1 for index, _ in indexed_texts), default=0) + texts: Final = tuple(_joined_choice_text(indexed_texts, index) for index in range(choice_count)) return texts if any(text is not None for text in texts) else None @@ -170,6 +196,8 @@ def _choice_with_text(choice: object, text: str | None) -> object: if choice_obj is None or text is None: return choice message: Final = _as_json_object(choice_obj.get("message")) - if message is None or message.get("content") == text: - return choice - return {**choice_obj, "message": {**message, "content": text}} + if message is not None: + return choice if message.get("content") == text else {**choice_obj, "message": {**message, "content": text}} + if isinstance(choice_obj.get("text"), str): + return {**choice_obj, "text": text} + return choice diff --git a/litellm/llms/openai/completion/guardrail_translation/handler.py b/litellm/llms/openai/completion/guardrail_translation/handler.py index c4bc04d4627..27c82cfc004 100644 --- a/litellm/llms/openai/completion/guardrail_translation/handler.py +++ b/litellm/llms/openai/completion/guardrail_translation/handler.py @@ -5,18 +5,30 @@ This module provides guardrail translation support for OpenAI's text completion The handler processes the 'prompt' parameter for guardrails. """ -from collections.abc import Mapping +import json +from collections.abc import Mapping, Sequence +from itertools import chain from typing import TYPE_CHECKING, Final from litellm._logging import verbose_proxy_logger -from litellm.llms.base_llm.guardrail_translation.base_translation import BaseTranslation -from litellm.types.utils import GenericGuardrailAPIInputs +from litellm.llms.base_llm.guardrail_translation.base_translation import ( + BaseTranslation, + StreamingScanKey, + StreamTransformSink, +) +from litellm.llms.base_llm.guardrail_translation.utils import ( + blocked_chat_stream_usage, + stream_item_field, + stream_item_items, +) +from litellm.types.utils import GenericGuardrailAPIInputs, TextChoices, TextCompletionResponse, Usage if TYPE_CHECKING: - from litellm.integrations.custom_guardrail import CustomGuardrail + from fastapi import HTTPException + + from litellm.integrations.custom_guardrail import CustomGuardrail, ModifyResponseException from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.types.proxy.auth.user_api_key_auth import UserAPIKeyAuth - from litellm.types.utils import TextCompletionResponse class OpenAITextCompletionHandler(BaseTranslation): @@ -32,6 +44,9 @@ class OpenAITextCompletionHandler(BaseTranslation): - A list of strings (for batch completions) """ + delivers_ended_stream_rewrites = True + assembles_streamed_response = True + async def process_input_messages( self, data: dict, @@ -125,7 +140,7 @@ class OpenAITextCompletionHandler(BaseTranslation): guardrail_to_apply: "CustomGuardrail", litellm_logging_obj: "LiteLLMLoggingObj | None" = None, user_api_key_dict: "UserAPIKeyAuth | None" = None, - request_data: dict | None = None, + request_data: dict[str, object] | None = None, ) -> "TextCompletionResponse": """ Process output response by applying guardrails to completion text. @@ -162,11 +177,7 @@ class OpenAITextCompletionHandler(BaseTranslation): if "response" not in request_data: request_data["response"] = response - # Add user API key metadata with prefixed keys - if "litellm_metadata" not in request_data: - user_metadata: Final = self.transform_user_api_key_dict_to_metadata(user_api_key_dict) - if user_metadata: - request_data["litellm_metadata"] = user_metadata + self.merge_user_api_key_metadata_into_request(request_data, user_api_key_dict) inputs: Final = GenericGuardrailAPIInputs(texts=texts_to_check) # Include model information from the response if available @@ -195,3 +206,168 @@ class OpenAITextCompletionHandler(BaseTranslation): ) return response + + async def process_output_streaming_response( + self, + responses_so_far: list[object], # mutable-ok: rewrites the caller's buffered chunks in place + guardrail_to_apply: "CustomGuardrail", + litellm_logging_obj: "LiteLLMLoggingObj | None" = None, + user_api_key_dict: "UserAPIKeyAuth | None" = None, + request_data: dict[str, object] | None = None, # mutable-ok: the hooks share and mutate the request payload + stream_transform_sink: StreamTransformSink | None = None, + deliver_ended_stream_rewrites: bool = False, + ) -> Sequence[object]: + chunks: Final = _text_completion_chunks(responses_so_far) + if not self._check_streaming_has_ended(chunks): + return responses_so_far + assembled: Final = self._assemble_ended_stream(chunks) + pre_guardrail_texts: Final = _choice_texts(assembled) + await self.process_output_response( + response=assembled, + guardrail_to_apply=guardrail_to_apply, + litellm_logging_obj=litellm_logging_obj, + user_api_key_dict=user_api_key_dict, + request_data=request_data, + ) + if not deliver_ended_stream_rewrites: + return responses_so_far + rewrites: Final = tuple( + after for before, after in zip(pre_guardrail_texts, _choice_texts(assembled)) if after != before + ) + for choice_index, text in rewrites: + self._write_choice_text(chunks, choice_index, text) + return responses_so_far + + def get_streaming_scan_key(self, responses_so_far: Sequence[object]) -> StreamingScanKey | None: + chunks: Final = _text_completion_chunks(responses_so_far) + if not self._check_streaming_has_ended(chunks): + return None + assembled: Final = self._assemble_ended_stream(chunks) + return StreamingScanKey(texts=tuple(text for _, text in _choice_texts(assembled)), stream_ended=True) + + def released_stream_as_ended(self, responses_so_far: Sequence[object]) -> tuple[object, ...]: + chunks: Final = _text_completion_chunks(responses_so_far) + streamed_choices: Final = tuple(chain.from_iterable(chunk.choices for chunk in chunks)) + finished_indices: Final = frozenset( + _choice_index(choice) for choice in streamed_choices if choice.finish_reason + ) + unfinished_indices: Final = tuple( + dict.fromkeys( + _choice_index(choice) for choice in streamed_choices if _choice_index(choice) not in finished_indices + ) + ) + if not unfinished_indices: + return tuple(responses_so_far) + terminator: Final = TextCompletionResponse( + id=chunks[0].id, + model=chunks[0].model, + object="text_completion", + choices=[TextChoices(index=index, text="", finish_reason="length") for index in unfinished_indices], + ) + return (*responses_so_far, terminator) + + def build_block_sse_chunks( + self, + exc: "ModifyResponseException", + stream_started: bool = False, + responses_so_far: Sequence[object] | None = None, + ) -> Sequence[bytes]: + streamed: Final = _text_completion_chunks(responses_so_far or ()) + source: Final = streamed[0] if streamed else TextCompletionResponse(model=exc.model) + message_chunk: Final = TextCompletionResponse( + id=source.id, + created=source.created, + model=source.model or exc.model, + object="text_completion", + choices=[TextChoices(index=0, text=exc.message)], + ).model_copy(update={"usage": None}) + final_chunk: Final = TextCompletionResponse( + id=source.id, + created=source.created, + model=source.model or exc.model, + object="text_completion", + choices=[TextChoices(index=0, text="", finish_reason="content_filter")], + ).model_copy(update={"usage": _blocked_stream_usage(exc.original_response)}) + return tuple( + f"data: {chunk.model_dump_json(exclude_none=True)}\n\n".encode() for chunk in (message_chunk, final_chunk) + ) + + def build_stream_error_items( + self, + exc: "HTTPException", + responses_so_far: Sequence[object] | None = None, + ) -> Sequence[bytes] | None: + from litellm.proxy.common_request_processing import sse_error_payload + + _, error_obj = sse_error_payload(exc) + return (f'data: {{"error": {json.dumps(error_obj)}}}\n\n'.encode(),) + + def _check_streaming_has_ended(self, responses_so_far: Sequence[object]) -> bool: + streamed_choices: Final = chain.from_iterable(stream_item_items(item, "choices") for item in responses_so_far) + return any(stream_item_field(choice, "finish_reason") is not None for choice in streamed_choices) + + @staticmethod + def _assemble_ended_stream(responses_so_far: Sequence[TextCompletionResponse]) -> TextCompletionResponse: + streamed_choices: Final = tuple(chain.from_iterable(item.choices for item in responses_so_far)) + choice_indices: Final = tuple(dict.fromkeys(_choice_index(choice) for choice in streamed_choices)) + assembled_choices: Final = [ + TextChoices( + index=index, + text="".join(_choice_text(choice) for choice in streamed_choices if _choice_index(choice) == index), + finish_reason=next( + ( + choice.finish_reason + for choice in streamed_choices + if _choice_index(choice) == index and choice.finish_reason + ), + None, + ), + ) + for index in choice_indices + ] + first: Final = responses_so_far[0] + return TextCompletionResponse( + id=first.id, + created=first.created, + model=first.model, + object="text_completion", + choices=assembled_choices, + usage=next((item.usage for item in reversed(responses_so_far) if item.usage is not None), None), + ) + + @staticmethod + def _write_choice_text(responses_so_far: Sequence[TextCompletionResponse], choice_index: int, text: str) -> None: + streamed_choices: Final = chain.from_iterable(item.choices for item in responses_so_far) + head, *tail = (choice for choice in streamed_choices if _choice_index(choice) == choice_index) + head.text = text + for choice in tail: + if isinstance(stream_item_field(choice, "text"), str): + choice.text = "" + + +def _blocked_stream_usage(original_response: object) -> Usage | None: + prompt_tokens, completion_tokens = blocked_chat_stream_usage(original_response) + if isinstance(original_response, list) and prompt_tokens == completion_tokens == 0: + return None + return Usage( + prompt_tokens=prompt_tokens, + completion_tokens=completion_tokens, + total_tokens=prompt_tokens + completion_tokens, + ) + + +def _text_completion_chunks(responses_so_far: Sequence[object]) -> tuple[TextCompletionResponse, ...]: + return tuple(item for item in responses_so_far if isinstance(item, TextCompletionResponse)) + + +def _choice_text(choice: TextChoices) -> str: + text: Final = stream_item_field(choice, "text") + return text if isinstance(text, str) else "" + + +def _choice_index(choice: TextChoices) -> int: + return choice.index or 0 + + +def _choice_texts(response: TextCompletionResponse) -> tuple[tuple[int, str], ...]: + return tuple((_choice_index(choice), _choice_text(choice)) for choice in response.choices) diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py index 4884e75af4e..7e487ded569 100644 --- a/litellm/proxy/common_request_processing.py +++ b/litellm/proxy/common_request_processing.py @@ -4059,13 +4059,13 @@ class ProxyBaseLLMRequestProcessing: (assembled_response, cache_hit); the closure also runs non-apply_guardrail post-call hooks via _run_deferred_stream_guardrails. - - Every other anthropic_messages/aresponses stream gets a closure - that dispatches on the stored args shape, because the arming site - cannot tell the producers apart: native iterators store a single - ready-made logging coroutine to enqueue, while bridged streams - (LiteLLMCompletionStreamingIterator, and the plain SSE generator - AnthropicStreamWrapper returns for bridged /v1/messages) share - their inner CustomStreamWrapper's logging_obj and so store + - Every other anthropic_messages/aresponses/atext_completion stream + gets a closure that dispatches on the stored args shape, because the + arming site cannot tell the producers apart: native iterators store + a single ready-made logging coroutine to enqueue, while bridged + streams (LiteLLMCompletionStreamingIterator, TextCompletionStreamWrapper, + and the plain SSE generator AnthropicStreamWrapper returns for bridged + /v1/messages) share their inner CustomStreamWrapper's logging_obj and so store (assembled_response, cache_hit); for those the closure only dispatches success logging, matching the route's pre-existing hook surface. @@ -4102,7 +4102,11 @@ class ProxyBaseLLMRequestProcessing: logging_obj._on_deferred_stream_complete = _on_deferred_stream_complete return - if route_type not in ("anthropic_messages", "aresponses") or not self._is_streaming_response(response): + if route_type not in ( + "anthropic_messages", + "aresponses", + "atext_completion", + ) or not self._is_streaming_response(response): return from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER diff --git a/litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py b/litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py index 7ff5aa2466b..e57d6e78378 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py +++ b/litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py @@ -82,6 +82,22 @@ def _as_plain_dict(item: object) -> Mapping[str, object]: return cast("Mapping[str, object]", item) # cast-ok: wire rows are message/tool-call dicts +_STR_KEYED_MAPPING: Final = TypeAdapter(Mapping[str, object]) + + +def _str_keyed(value: object) -> Mapping[str, object]: + return _STR_KEYED_MAPPING.validate_python(value) if isinstance(value, Mapping) else {} + + +def _request_route(request_data: Mapping[str, object]) -> str | None: + litellm_metadata: Final = request_data.get("litellm_metadata") + metadata_bucket: Final = _str_keyed( + litellm_metadata if litellm_metadata is not None else request_data.get("metadata") + ) + route: Final = metadata_bucket.get("user_api_key_request_route") + return route if isinstance(route, str) else None + + class GraySwanGuardrailMissingSecrets(Exception): """Raised when the Gray Swan API key is missing.""" @@ -625,11 +641,8 @@ class GraySwanGuardrail(CustomGuardrail): from litellm.litellm_core_utils.api_route_to_call_types import get_call_types_for_route from litellm.llms import load_guardrail_translation_mappings - litellm_metadata: Final = request_data.get("litellm_metadata") - request_route: Final = ( - litellm_metadata.get("user_api_key_request_route") if isinstance(litellm_metadata, Mapping) else None - ) - route_call_types: Final = get_call_types_for_route(request_route) if isinstance(request_route, str) else None + request_route: Final = _request_route(request_data) + route_call_types: Final = get_call_types_for_route(request_route) if request_route is not None else None call_type: Final = ( (route_call_types[0].value if route_call_types else None) or (logging_obj.call_type if logging_obj is not None else None) @@ -698,13 +711,12 @@ class GraySwanGuardrail(CustomGuardrail): inbound_headers: Final = self._extract_inbound_headers(request_data, logging_obj) - litellm_metadata: Final = request_data.get("litellm_metadata") - cleaned_litellm_metadata: Final = dict(litellm_metadata) if isinstance(litellm_metadata, dict) else {} - if inbound_headers: - existing_headers: Final = cleaned_litellm_metadata.get("headers") - cleaned_litellm_metadata["headers"] = ( - {**existing_headers, **inbound_headers} if isinstance(existing_headers, dict) else inbound_headers - ) + litellm_metadata: Final = _str_keyed(request_data.get("litellm_metadata")) + cleaned_litellm_metadata: Final = ( + {**litellm_metadata, "headers": {**_str_keyed(litellm_metadata.get("headers")), **inbound_headers}} + if inbound_headers + else litellm_metadata + ) if cleaned_litellm_metadata: sanitized: Final = _DECODED_JSON.validate_python( safe_json_loads(safe_dumps(cleaned_litellm_metadata), default={}) diff --git a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py index f58c8705815..71b8976bebf 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py +++ b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py @@ -12,11 +12,11 @@ import json import os import re import time -from collections.abc import AsyncGenerator, Coroutine, Mapping, Sequence +from collections.abc import AsyncGenerator, AsyncIterable, Callable, Coroutine, Mapping, Sequence from dataclasses import dataclass, replace from datetime import datetime from re import Pattern -from typing import TYPE_CHECKING, Any, Final, Literal, Optional, TypedDict, cast +from typing import TYPE_CHECKING, Any, Final, Literal, Optional, TypedDict, TypeVar, cast import yaml from fastapi import HTTPException @@ -40,6 +40,7 @@ from litellm.proxy.guardrails.content_filter_data import ( ) from litellm.types.utils import ( CallTypes, + Delta, Function, GenericGuardrailAPIInputs, GuardrailStatus, @@ -47,6 +48,9 @@ from litellm.types.utils import ( Message, ModelResponse, ModelResponseStream, + StreamingChoices, + TextChoices, + TextCompletionResponse, ) if TYPE_CHECKING: @@ -78,6 +82,7 @@ from .patterns import PATTERN_EXTRA_CONFIG, get_compiled_pattern MAX_KEYWORD_VALUE_GAP_WORDS: Final = 1 _DECODED_ARGUMENTS: Final = TypeAdapter(object) +_StreamItemT: Final = TypeVar("_StreamItemT") _JSON_ENTRIES: Final = TypeAdapter(list[object]) GAP_WORD_TOKENIZER: Final = re.compile(r"\b\w+\b") SENTENCE_TERMINATORS: Final = re.compile(r"[.!?]+") @@ -139,6 +144,26 @@ class _CategoryConfigView(TypedDict): category_file: str | None +def _streamed_choice_content(choice: StreamingChoices | TextChoices) -> str | None: + if isinstance(choice, TextChoices): + return choice.text if isinstance(choice.text, str) else "" + delta: Final = getattr(choice, "delta", None) + if not delta: + return None + content: Final = getattr(delta, "content", None) + return content if isinstance(content, str) else "" + + +def _streamed_choices(item: object) -> Sequence[StreamingChoices | TextChoices]: + if isinstance(item, (ModelResponseStream, TextCompletionResponse)): + return item.choices + return () + + +def _is_choiceless_chunk(item: object) -> bool: + return isinstance(item, (ModelResponseStream, TextCompletionResponse)) and not item.choices + + @dataclass(frozen=True, slots=True) class _StreamedChoiceState: buffered_text: str = "" @@ -146,6 +171,52 @@ class _StreamedChoiceState: committed_detections: tuple[ContentFilterDetection, ...] = () latest_detections: tuple[ContentFilterDetection, ...] = () next_trim_len: int = 0 + finished: bool = False + + +def _end_of_stream_chunk( + last_item: object, state_by_choice: Mapping[int, _StreamedChoiceState] +) -> ModelResponseStream | TextCompletionResponse | None: + unfinished: Final = tuple(index for index, state in state_by_choice.items() if not state.finished) + if not unfinished: + return None + if isinstance(last_item, TextCompletionResponse): + return TextCompletionResponse( + id=last_item.id, + created=last_item.created, + model=last_item.model, + object="text_completion", + choices=[TextChoices(index=index, text="") for index in unfinished], + ).model_copy(update={"usage": None}) + if isinstance(last_item, ModelResponseStream): + return ModelResponseStream( + id=last_item.id, + created=last_item.created, + model=last_item.model, + choices=[StreamingChoices(index=index, delta=Delta(content="")) for index in unfinished], + ) + return None + + +async def _stream_then_end_of_stream_chunk( + response: AsyncIterable[_StreamItemT], end_of_stream_chunk: Callable[[_StreamItemT], _StreamItemT | None] +) -> AsyncGenerator[tuple[_StreamItemT, bool], None]: + last_item: _StreamItemT | None = None # rebind-ok: the end-of-stream chunk copies the last chunk's id and model + held_items: tuple[_StreamItemT, ...] = () # rebind-ok: the usage chunk must stay after the end-of-stream chunk + async for item in response: + last_item = item + if _is_choiceless_chunk(item): + held_items = (*held_items, item) + continue + for held_item in held_items: + yield held_item, False + held_items = () + yield item, False + closing_item: Final = None if last_item is None else end_of_stream_chunk(last_item) + if closing_item is not None: + yield closing_item, True + for held_item in held_items: + yield held_item, False @dataclass(frozen=True, slots=True) @@ -2104,72 +2175,75 @@ class ContentFilterGuardrail(CustomGuardrail): ) try: - async for item in response: - if isinstance(item, ModelResponseStream) and item.choices: - for choice in item.choices: - if not (hasattr(choice, "delta") and choice.delta): - continue + async for item, stream_ended in _stream_then_end_of_stream_chunk( + response, lambda last_item: _end_of_stream_chunk(last_item, state_by_choice) + ): + for choice in _streamed_choices(item): + new_content = _streamed_choice_content(choice) + if new_content is None: + continue - choice_index = getattr(choice, "index", 0) - if not isinstance(choice_index, int): - choice_index = 0 + choice_index = getattr(choice, "index", 0) + if not isinstance(choice_index, int): + choice_index = 0 - content = getattr(choice.delta, "content", None) - is_final = bool(getattr(choice, "finish_reason", None)) - new_content = content if isinstance(content, str) else "" - if not new_content and not is_final: - continue + is_final = stream_ended or bool(getattr(choice, "finish_reason", None)) + if not new_content and not is_final: + continue - previous_state = state_by_choice.get(choice_index, _StreamedChoiceState()) - buffered_text = previous_state.buffered_text + new_content - if not buffered_text: - continue + previous_state = state_by_choice.get(choice_index, _StreamedChoiceState()) + buffered_text = previous_state.buffered_text + new_content + if not buffered_text: + continue - # Add a space at the end if it's the final chunk to trigger word boundaries (\b) - text_to_scan = buffered_text + (" " if is_final else "") - choice_detections: list[ContentFilterDetection] = [] - scan_started = time.perf_counter() + # Add a space at the end if it's the final chunk to trigger word boundaries (\b) + text_to_scan = buffered_text + (" " if is_final else "") + choice_detections: list[ContentFilterDetection] = [] + scan_started = time.perf_counter() - try: - masked_text = self._filter_single_text(text_to_scan, detections=choice_detections) - if is_final and masked_text.endswith(" "): - masked_text = masked_text[:-1] - latest_detections = tuple(choice_detections) - except HTTPException: - state_by_choice[choice_index] = replace( - previous_state, latest_detections=tuple(choice_detections) - ) - raise - except Exception as e: - verbose_proxy_logger.error("ContentFilterGuardrail: Error in masking: %s", e) - masked_text = text_to_scan # Fallback to current text - latest_detections = previous_state.latest_detections - finally: - scan_seconds += time.perf_counter() - scan_started - - safe_to_yield_len = max( - previous_state.yielded_masked_text_len, - len(masked_text) - (0 if is_final else CONTENT_FILTER_STREAMING_HOLDBACK_CHARS), + try: + masked_text = self._filter_single_text(text_to_scan, detections=choice_detections) + if is_final and masked_text.endswith(" "): + masked_text = masked_text[:-1] + latest_detections = tuple(choice_detections) + except HTTPException: + state_by_choice[choice_index] = replace( + previous_state, latest_detections=tuple(choice_detections) ) - choice.delta.content = masked_text[previous_state.yielded_masked_text_len : safe_to_yield_len] - next_state = replace( - previous_state, - buffered_text=buffered_text, - yielded_masked_text_len=safe_to_yield_len, - latest_detections=latest_detections, - ) - if is_final: - state_by_choice[choice_index] = next_state - continue + raise + except Exception as e: + verbose_proxy_logger.error("ContentFilterGuardrail: Error in masking: %s", e) + masked_text = text_to_scan # Fallback to current text + latest_detections = previous_state.latest_detections + finally: + scan_seconds += time.perf_counter() - scan_started - trim_started = time.perf_counter() - state_by_choice[choice_index] = self._trim_streamed_choice_buffer(next_state, masked_text, plan) - scan_seconds += time.perf_counter() - trim_started + safe_to_yield_len = max( + previous_state.yielded_masked_text_len, + len(masked_text) - (0 if is_final else CONTENT_FILTER_STREAMING_HOLDBACK_CHARS), + ) + released_text = masked_text[previous_state.yielded_masked_text_len : safe_to_yield_len] + if isinstance(choice, TextChoices): + choice.text = released_text + else: + choice.delta.content = released_text + next_state = replace( + previous_state, + buffered_text=buffered_text, + yielded_masked_text_len=safe_to_yield_len, + latest_detections=latest_detections, + ) + if is_final: + state_by_choice[choice_index] = replace(next_state, finished=True) + continue - yield item - else: - # Not a ModelResponseStream or no choices - yield as is - yield item + trim_started = time.perf_counter() + state_by_choice[choice_index] = self._trim_streamed_choice_buffer(next_state, masked_text, plan) + scan_seconds += time.perf_counter() - trim_started + + if stream_ended and not any(_streamed_choice_content(choice) for choice in _streamed_choices(item)): + continue + yield item except HTTPException: status = "guardrail_intervened" raise diff --git a/tests/integration/streaming/test_text_completion_stream_guardrail_rebuild.py b/tests/integration/streaming/test_text_completion_stream_guardrail_rebuild.py index 229a9be6d20..ea0a820a705 100644 --- a/tests/integration/streaming/test_text_completion_stream_guardrail_rebuild.py +++ b/tests/integration/streaming/test_text_completion_stream_guardrail_rebuild.py @@ -27,7 +27,12 @@ _PROXY_CONFIG: Final = Path(__file__).resolve().parents[1] / "proxy_config.yaml" _PROVIDER_KEY: Final = "synthetic-provider-key" _GUARDRAIL: Final = "audit-text-rebuild-tool-permission" _TEXT_MODEL: Final = f"audit-text-instruct-{uuid.uuid4().hex}" +_GOVERNED_MODEL: Final = f"audit-text-governed-{uuid.uuid4().hex}" +_KEYWORD_FILTER: Final = "audit-text-keyword-filter" +_POLICY: Final = "audit-text-post-call-governance" _PIECES: Final = ("Hello", " streamed", " world") +_MASKED_PIECES: Final = ("A pers", "im", "mon tart") +_BLOCKED_PIECES: Final = ("A kum", "quat tart") _PROMPT_TOKENS: Final = 11 _COMPLETION_TOKENS: Final = 3 _USAGE: Final[dict[str, JsonValue]] = { @@ -97,6 +102,14 @@ def _text_choice(text: str, finish_reason: str | None) -> dict[str, JsonValue]: return {"text": text, "index": 0, "logprobs": None, "finish_reason": finish_reason} +def _pieces_for(identity: str) -> tuple[str, ...]: + if identity.startswith("cmpl-mask-"): + return _MASKED_PIECES + if identity.startswith("cmpl-block-"): + return _BLOCKED_PIECES + return _PIECES + + def _respond(request: Request) -> Reply: body: Final = JSON_OBJECT.validate_json(request.body) assert body.get("stream") is True, body @@ -106,7 +119,7 @@ def _respond(request: Request) -> Reply: return Reply( content_type="text/event-stream", chunks=( - *(_text_frame(identity, [_text_choice(piece, None)]) for piece in _PIECES), + *(_text_frame(identity, [_text_choice(piece, None)]) for piece in _pieces_for(identity)), _text_frame(identity, [_text_choice("", "length")]), _text_frame(identity, [], usage=_USAGE), b"data: [DONE]\n\n", @@ -114,20 +127,25 @@ def _respond(request: Request) -> Reply: ) +def _text_deployment(model_name: str, wire: Wire) -> dict[str, JsonValue]: + return { + "model_name": model_name, + "litellm_params": { + "model": "text-completion-openai/gpt-3.5-turbo-instruct", + "api_base": f"{wire.url}/v1", + "api_key": _PROVIDER_KEY, + }, + } + + def _config(directory: Path, wire: Wire) -> Path: base: Final = object_value(yaml.safe_load(_PROXY_CONFIG.read_text())) config: Final = { **base, "model_list": [ *list_value(base["model_list"]), - { - "model_name": _TEXT_MODEL, - "litellm_params": { - "model": "text-completion-openai/gpt-3.5-turbo-instruct", - "api_base": f"{wire.url}/v1", - "api_key": _PROVIDER_KEY, - }, - }, + _text_deployment(_TEXT_MODEL, wire), + _text_deployment(_GOVERNED_MODEL, wire), ], "guardrails": [ { @@ -139,8 +157,30 @@ def _config(directory: Path, wire: Wire) -> Path: "default_action": "allow", "rules": [], }, - } + }, + { + "guardrail_name": _KEYWORD_FILTER, + "litellm_params": { + "guardrail": "litellm_content_filter", + "mode": "post_call", + "default_on": False, + "blocked_words": [ + {"keyword": "persimmon", "action": "MASK"}, + {"keyword": "kumquat", "action": "BLOCK"}, + ], + }, + }, ], + "policies": { + _POLICY: { + "guardrails": {"add": [_KEYWORD_FILTER]}, + "pipeline": { + "mode": "post_call", + "steps": [{"guardrail": _KEYWORD_FILTER, "on_pass": "allow", "on_fail": "block"}], + }, + } + }, + "policy_attachments": [{"policy": _POLICY, "models": [_GOVERNED_MODEL]}], } path: Final = directory / f"text-rebuild-{uuid.uuid4().hex}.yaml" path.write_text(yaml.safe_dump(config)) @@ -163,14 +203,14 @@ def rig(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Rig]: yield _Rig(wire, owned) -def _stream(rig: _Rig, identity: str) -> _Stream: +def _stream(rig: _Rig, identity: str, model: str = _TEXT_MODEL) -> _Stream: gateway: Final = rig.proxy.gateway with gateway.client.stream( "POST", "/v1/completions", headers={"Authorization": f"Bearer {gateway.key}"}, json={ - "model": _TEXT_MODEL, + "model": model, "prompt": identity, "max_tokens": 16, "stream": True, @@ -212,6 +252,37 @@ def test_guarded_text_completion_stream_reaches_the_caller_whole_with_its_usage( assert _GUARDRAIL in list_value(object_value(row["metadata"])["applied_guardrails"]), row +def test_post_call_pipeline_masks_a_streamed_text_completion(rig: _Rig) -> None: + identity: Final = f"cmpl-mask-{uuid.uuid4().hex}" + + stream: Final = _stream(rig, identity, _GOVERNED_MODEL) + + assert stream.status == 200, stream.body + assert stream.body.rstrip().endswith(b"data: [DONE]"), stream.body[-400:] + assert stream.text == "A [KEYWORD_REDACTED] tart", stream.frames + assert b"persimmon" not in stream.body, stream.body + assert stream.finish_reasons == ("length",), stream.frames + assert [(usage["prompt_tokens"], usage["completion_tokens"]) for usage in stream.usages] == [ + (_PROMPT_TOKENS, _COMPLETION_TOKENS) + ], stream.frames + (row,) = eventually( + lambda: rig.proxy.gateway.rows('SELECT metadata FROM "LiteLLM_SpendLogs" WHERE request_id = %s', (identity,)), + lambda rows: len(rows) == 1, + seconds=70, + ) + assert _KEYWORD_FILTER in list_value(object_value(row["metadata"])["applied_guardrails"]), row + + +def test_post_call_pipeline_blocks_a_streamed_text_completion(rig: _Rig) -> None: + identity: Final = f"cmpl-block-{uuid.uuid4().hex}" + + stream: Final = _stream(rig, identity, _GOVERNED_MODEL) + + assert stream.status == 400, stream.body + assert b"Content blocked: keyword 'kumquat' detected" in stream.body, stream.body + assert b"quat tart" not in stream.body, stream.body + + def _together_chunk(choices: list[JsonValue], **extra: JsonValue) -> bytes: return _frame( { diff --git a/tests/unit/litellm_core_utils/test_served_output_texts.py b/tests/unit/litellm_core_utils/test_served_output_texts.py index 637a20571c8..e01c710dfdb 100644 --- a/tests/unit/litellm_core_utils/test_served_output_texts.py +++ b/tests/unit/litellm_core_utils/test_served_output_texts.py @@ -9,7 +9,16 @@ from litellm.litellm_core_utils.served_output_texts import ( served_output_texts, served_stream_output_texts, ) -from litellm.types.utils import Choices, Delta, Message, ModelResponse, ModelResponseStream, StreamingChoices +from litellm.types.utils import ( + Choices, + Delta, + Message, + ModelResponse, + ModelResponseStream, + StreamingChoices, + TextChoices, + TextCompletionResponse, +) RAW = "Card: 4111 1111 1111 1111" MASKED = "Card: " @@ -83,6 +92,30 @@ def test_served_stream_output_texts_keeps_every_choice_index_when_chunks_carry_o assert served_stream_output_texts(chunks) == ("first choice", MASKED) +def _text_completion_chunk(text: str | None, index: int = 0) -> TextCompletionResponse: + return TextCompletionResponse(object="text_completion", choices=[TextChoices(index=index, text=text)]) + + +def test_served_stream_output_texts_joins_text_completion_chunks_per_choice(): + chunks = [ + _text_completion_chunk("Card: ", 0), + _text_completion_chunk("second", 1), + _text_completion_chunk("", 0), + _text_completion_chunk(None, 1), + TextCompletionResponse(object="text_completion", choices=[]), + ] + assert served_stream_output_texts(chunks) == (MASKED, "second") + + +def test_overlay_replaces_text_completion_choice_text_with_served_text(): + logged = {"id": "x", "object": "text_completion", "choices": [{"index": 0, "text": RAW}, {"index": 1, "text": RAW}]} + assert overlay_served_output_texts(logged, (MASKED,)) == { + "id": "x", + "object": "text_completion", + "choices": [{"index": 0, "text": MASKED}, {"index": 1, "text": RAW}], + } + + @pytest.mark.parametrize("content", (None, "", "first")) def test_null_choice_index_joins_choice_zero_without_changing_other_choices(content: str | None) -> None: chunks: Final = ( diff --git a/tests/unit/llms/openai/completion/test_text_completion_guardrail_handler.py b/tests/unit/llms/openai/completion/test_text_completion_guardrail_handler.py index c96fbf34fe1..9beb1faf7f4 100644 --- a/tests/unit/llms/openai/completion/test_text_completion_guardrail_handler.py +++ b/tests/unit/llms/openai/completion/test_text_completion_guardrail_handler.py @@ -2,18 +2,25 @@ Unit tests for OpenAI Text Completion Guardrail Translation Handler """ -from typing import List, Optional, Tuple +import json +from typing import List, Literal, Optional, Tuple from unittest.mock import MagicMock import pytest +from fastapi import HTTPException from litellm.integrations.custom_guardrail import CustomGuardrail +from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.llms import get_guardrail_translation_mapping from litellm.llms.openai.completion.guardrail_translation.handler import ( OpenAITextCompletionHandler, ) -from litellm.types.utils import CallTypes, TextChoices, TextCompletionResponse +from litellm.proxy._types import UserAPIKeyAuth +from litellm.proxy.guardrails.guardrail_hooks.litellm_content_filter.content_filter import ContentFilterGuardrail +from litellm.proxy.guardrails.guardrail_hooks.unified_guardrail.unified_guardrail import UnifiedLLMGuardrails +from litellm.types.guardrails import BlockedWord, ContentFilterAction, GuardrailEventHooks +from litellm.types.utils import CallTypes, GenericGuardrailAPIInputs, TextChoices, TextCompletionResponse, Usage class MockGuardrail(CustomGuardrail): @@ -341,3 +348,215 @@ class TestPIIMaskingScenario: assert "[EMAIL_REDACTED]" in result["prompt"][1] assert "alice@example.com" not in result["prompt"][0] assert "bob@test.org" not in result["prompt"][1] + + +class _RecordingKeywordFilter(ContentFilterGuardrail): + def __init__(self): + super().__init__( + guardrail_name="keyword-filter", + blocked_words=[ + BlockedWord(keyword="persimmon", action=ContentFilterAction.MASK), + BlockedWord(keyword="kumquat", action=ContentFilterAction.BLOCK), + ], + event_hook=GuardrailEventHooks.post_call, + default_on=True, + ) + self.streaming_sampling_rate = 1 + self.scanned_texts: tuple[tuple[str, ...], ...] = () + + async def apply_guardrail( + self, + inputs: GenericGuardrailAPIInputs, + request_data: dict, + input_type: Literal["request", "response"], + logging_obj: LiteLLMLoggingObj | None = None, + ) -> GenericGuardrailAPIInputs: + self.scanned_texts = (*self.scanned_texts, tuple(inputs.get("texts") or ())) + return await super().apply_guardrail(inputs, request_data, input_type, logging_obj) + + +class _PassthroughReplaceGuardrail(CustomGuardrail): + async def apply_guardrail( + self, + inputs: GenericGuardrailAPIInputs, + request_data: dict, + input_type: Literal["request", "response"], + logging_obj: LiteLLMLoggingObj | None = None, + ) -> GenericGuardrailAPIInputs: + if any("kumquat" in text for text in inputs.get("texts") or []): + self.raise_passthrough_exception(violation_message="replaced kumquat", request_data=request_data) + return inputs + + +def _text_chunk(index: int | None, text: str | None, finish_reason: str | None = None) -> TextCompletionResponse: + return TextCompletionResponse( + id="cmpl-stream", + created=1589478378, + model="gpt-3.5-turbo-instruct", + object="text_completion", + choices=[TextChoices(index=index, text=text, finish_reason=finish_reason)], + ) + + +def _usage_chunk() -> TextCompletionResponse: + return TextCompletionResponse( + id="cmpl-stream", + created=1589478378, + model="gpt-3.5-turbo-instruct", + object="text_completion", + choices=[], + usage=Usage(prompt_tokens=3, completion_tokens=5, total_tokens=8), + ) + + +def _delivered(chunks: list[TextCompletionResponse]) -> list[tuple[int, str | None, str | None]]: + return [(choice.index, choice.text, choice.finish_reason) for chunk in chunks for choice in chunk.choices] + + +async def _drive_flat_stream(guardrail: CustomGuardrail, chunks: list[TextCompletionResponse]) -> list[object]: + async def upstream(): + for chunk in chunks: + yield chunk + + return [ + item + async for item in UnifiedLLMGuardrails().async_post_call_streaming_iterator_hook( + user_api_key_dict=UserAPIKeyAuth(api_key="test-key", request_route="/v1/completions"), + response=upstream(), + request_data={"guardrail_to_apply": guardrail, "model": "gpt-3.5-turbo-instruct"}, + ) + ] + + +class TestStreamingOutputProcessing: + @pytest.mark.asyncio + async def test_ended_stream_delivers_each_choice_masked_text(self): + guardrail = _RecordingKeywordFilter() + chunks = [ + _text_chunk(0, "A pers"), + _text_chunk(1, "kiwi "), + _text_chunk(0, "im"), + _text_chunk(1, "pie", "stop"), + _text_chunk(0, "mon tart"), + _text_chunk(0, None, "length"), + _usage_chunk(), + ] + + await OpenAITextCompletionHandler().process_output_streaming_response( + responses_so_far=chunks, + guardrail_to_apply=guardrail, + request_data={}, + deliver_ended_stream_rewrites=True, + ) + + assert guardrail.scanned_texts == (("A persimmon tart", "kiwi pie"),) + assert _delivered(chunks) == [ + (0, "A [KEYWORD_REDACTED] tart", None), + (1, "kiwi ", None), + (0, "", None), + (1, "pie", "stop"), + (0, "", None), + (0, None, "length"), + ] + assert chunks[-1].usage.total_tokens == 8 + + @pytest.mark.asyncio + async def test_ended_stream_scanned_without_delivery_keeps_chunks_and_stores_the_assembled_response(self): + chunks = [_text_chunk(0, "A pers"), _text_chunk(0, "immon tart"), _text_chunk(0, "", "stop"), _usage_chunk()] + request_data: dict[str, object] = {} + + await OpenAITextCompletionHandler().process_output_streaming_response( + responses_so_far=chunks, guardrail_to_apply=_RecordingKeywordFilter(), request_data=request_data + ) + + assert _delivered(chunks) == [(0, "A pers", None), (0, "immon tart", None), (0, "", "stop")] + assembled = request_data["response"] + assert isinstance(assembled, TextCompletionResponse) + assert [(c.text, c.finish_reason) for c in assembled.choices] == [("A [KEYWORD_REDACTED] tart", "stop")] + assert assembled.usage.total_tokens == 8 + + @pytest.mark.asyncio + async def test_ended_stream_block_raises_the_guardrail_error(self): + chunks = [_text_chunk(0, "A kum"), _text_chunk(0, "quat tart"), _text_chunk(0, "", "stop")] + + with pytest.raises(HTTPException) as exc_info: + await OpenAITextCompletionHandler().process_output_streaming_response( + responses_so_far=chunks, guardrail_to_apply=_RecordingKeywordFilter(), deliver_ended_stream_rewrites=True + ) + + assert exc_info.value.status_code == 400 + assert "kumquat" in str(exc_info.value.detail) + + @pytest.mark.asyncio + async def test_ended_stream_treats_a_null_choice_index_as_choice_zero(self): + guardrail = _RecordingKeywordFilter() + chunks = [_text_chunk(None, "A pers"), _text_chunk(0, "immon tart"), _text_chunk(0, "", "stop")] + + await OpenAITextCompletionHandler().process_output_streaming_response( + responses_so_far=chunks, guardrail_to_apply=guardrail, request_data={}, deliver_ended_stream_rewrites=True + ) + + assert guardrail.scanned_texts == (("A persimmon tart",),) + assert [choice.text for chunk in chunks for choice in chunk.choices] == ["A [KEYWORD_REDACTED] tart", "", ""] + + @pytest.mark.asyncio + async def test_ended_stream_blocks_a_keyword_split_across_a_null_and_a_zero_choice_index(self): + chunks = [_text_chunk(None, "A kum"), _text_chunk(0, "quat tart"), _text_chunk(0, "", "stop")] + + with pytest.raises(HTTPException) as exc_info: + await OpenAITextCompletionHandler().process_output_streaming_response( + responses_so_far=chunks, guardrail_to_apply=_RecordingKeywordFilter(), deliver_ended_stream_rewrites=True + ) + + assert exc_info.value.status_code == 400 + assert "kumquat" in str(exc_info.value.detail) + + @pytest.mark.asyncio + async def test_flat_stream_is_scanned_once_after_it_ends(self): + guardrail = _RecordingKeywordFilter() + chunks = [_text_chunk(0, "A pers"), _text_chunk(0, "immon tart"), _text_chunk(0, "", "stop"), _usage_chunk()] + + delivered = await _drive_flat_stream(guardrail, chunks) + + assert guardrail.scanned_texts == (("A persimmon tart",),) + assert len(delivered) == len(chunks) + + @pytest.mark.asyncio + async def test_flat_stream_block_ends_with_an_in_band_error_frame(self): + chunks = [_text_chunk(0, "A kum"), _text_chunk(0, "quat tart"), _text_chunk(0, "", "stop"), _usage_chunk()] + + delivered = await _drive_flat_stream(_RecordingKeywordFilter(), chunks) + + error_frame = delivered[-1] + assert isinstance(error_frame, bytes) + assert error_frame.startswith(b'data: {"error": ') + assert b"kumquat" in error_frame + + @pytest.mark.asyncio + async def test_flat_stream_passthrough_block_appends_the_replacement_and_ends_with_content_filter(self): + guardrail = _PassthroughReplaceGuardrail( + guardrail_name="passthrough", event_hook=GuardrailEventHooks.post_call, default_on=True + ) + chunks = [_text_chunk(0, "A kum"), _text_chunk(0, "quat tart"), _text_chunk(0, "", "stop"), _usage_chunk()] + + delivered = await _drive_flat_stream(guardrail, chunks) + + frames = [json.loads(item.decode().removeprefix("data: ")) for item in delivered if isinstance(item, bytes)] + assert [(frame["choices"][0]["text"], frame["choices"][0].get("finish_reason")) for frame in frames] == [ + ("replaced kumquat", None), + ("", "content_filter"), + ] + assert frames[-1]["usage"]["total_tokens"] == 8 + + @pytest.mark.asyncio + async def test_flat_stream_passthrough_block_reports_no_usage_the_upstream_never_sent(self): + guardrail = _PassthroughReplaceGuardrail( + guardrail_name="passthrough", event_hook=GuardrailEventHooks.post_call, default_on=True + ) + chunks = [_text_chunk(0, "A kum"), _text_chunk(0, "quat tart"), _text_chunk(0, "", "stop")] + + delivered = await _drive_flat_stream(guardrail, [chunk.model_copy(update={"usage": None}) for chunk in chunks]) + + frames = [json.loads(item.decode().removeprefix("data: ")) for item in delivered if isinstance(item, bytes)] + assert frames[-1]["choices"][0]["finish_reason"] == "content_filter" + assert [frame for frame in frames if "usage" in frame] == [] diff --git a/tests/unit/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py b/tests/unit/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py index f7fb7e1fed0..9a14222c684 100644 --- a/tests/unit/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py +++ b/tests/unit/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py @@ -4,6 +4,7 @@ Tests for the Content Filter Guardrail import json import os +from collections.abc import AsyncGenerator, AsyncIterator, Callable from typing import Final from unittest.mock import MagicMock @@ -30,7 +31,17 @@ from litellm.types.proxy.guardrails.guardrail_hooks.litellm_content_filter impor ContentFilterCategoryConfig, ContentFilterDetection, ) -from litellm.types.utils import StandardLoggingGuardrailInformation +from litellm.llms.base_llm.guardrail_translation.utils import stream_item_field +from litellm.types.llms.openai import OutputTextDeltaEvent, ResponsesAPIStreamEvents +from litellm.types.utils import ( + Delta, + ModelResponseStream, + StandardLoggingGuardrailInformation, + StreamingChoices, + TextChoices, + TextCompletionResponse, + Usage, +) class TestContentFilterGuardrail: @@ -3523,3 +3534,180 @@ async def test_image_whose_description_is_clean_passes_through_unchanged(): ) assert result == {"images": ["https://example.com/desk.png"], "texts": []} + + +def _fruit_word_filter() -> ContentFilterGuardrail: + return ContentFilterGuardrail( + guardrail_name="fruit-filter", + blocked_words=[ + BlockedWord(keyword="persimmon", action=ContentFilterAction.MASK), + BlockedWord(keyword="kumquat", action=ContentFilterAction.BLOCK), + ], + event_hook=GuardrailEventHooks.post_call, + ) + + +async def _text_completion_stream(*parts: tuple[str | None, str | None]): + for text, finish_reason in parts: + yield TextCompletionResponse( + id="cmpl-stream", + model="gpt-3.5-turbo-instruct", + object="text_completion", + choices=[TextChoices(index=0, text=text, finish_reason=finish_reason)], + ) + + +@pytest.mark.asyncio +async def test_streaming_hook_masks_a_word_split_across_text_completion_chunks(): + request_data: Final[dict] = {} + + delivered: Final = [ + chunk + async for chunk in _fruit_word_filter().async_post_call_streaming_iterator_hook( + user_api_key_dict=MagicMock(), + response=_text_completion_stream(("A pers", None), ("immon tart", None), (None, "stop")), + request_data=request_data, + ) + ] + + assert "".join(chunk.choices[0].text or "" for chunk in delivered) == "A [KEYWORD_REDACTED] tart" + assert request_data["metadata"]["standard_logging_guardrail_information"][0]["guardrail_status"] == "success" + + +_LONG_TAIL: Final = "the quick brown fox jumps over the lazy dog and then eats a ripe persimmon" + + +async def _text_completion_n2_stream_without_choice_1_finish() -> AsyncGenerator[TextCompletionResponse, None]: + for index, text, finish_reason in ( + (0, "A short answer", None), + (1, _LONG_TAIL[:30], None), + (1, _LONG_TAIL[30:], None), + (0, None, "stop"), + ): + yield TextCompletionResponse( + id="cmpl-stream", + model="gpt-3.5-turbo-instruct", + object="text_completion", + choices=[TextChoices(index=index, text=text, finish_reason=finish_reason)], + ).model_copy(update={"usage": None}) + + +async def _chat_n2_stream_without_choice_1_finish() -> AsyncGenerator[ModelResponseStream, None]: + for index, content, finish_reason in ( + (0, "A short answer", None), + (1, _LONG_TAIL[:30], None), + (1, _LONG_TAIL[30:], None), + (0, None, "stop"), + ): + yield ModelResponseStream( + id="chatcmpl-stream", + model="gpt-4.1-mini", + choices=[StreamingChoices(index=index, delta=Delta(content=content), finish_reason=finish_reason)], + ) + + +def _streamed_text(chunk: ModelResponseStream | TextCompletionResponse, index: int) -> str: + return "".join(_choice_text(choice) for choice in chunk.choices if choice.index == index) + + +def _choice_text(choice: TextChoices | StreamingChoices) -> str: + content: Final = stream_item_field(choice, "text") if isinstance(choice, TextChoices) else choice.delta.content + return content if isinstance(content, str) else "" + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "stream", [_text_completion_n2_stream_without_choice_1_finish, _chat_n2_stream_without_choice_1_finish] +) +async def test_streaming_hook_releases_the_held_tail_of_a_choice_the_stream_never_finished( + stream: Callable[[], AsyncIterator[ModelResponseStream | TextCompletionResponse]], +) -> None: + delivered: Final = [ + chunk + async for chunk in _fruit_word_filter().async_post_call_streaming_iterator_hook( + user_api_key_dict=MagicMock(), response=stream(), request_data={} + ) + ] + + assert "".join(_streamed_text(chunk, 0) for chunk in delivered) == "A short answer" + assert "".join(_streamed_text(chunk, 1) for chunk in delivered) == _LONG_TAIL.replace( + "persimmon", "[KEYWORD_REDACTED]" + ) + assert all(getattr(chunk, "usage", None) is None for chunk in delivered) + + +_STREAM_USAGE: Final = Usage(prompt_tokens=5, completion_tokens=20, total_tokens=25) + + +async def _then_usage_chunk( + stream: Callable[[], AsyncIterator[ModelResponseStream | TextCompletionResponse]], + usage_chunk: ModelResponseStream | TextCompletionResponse, +) -> AsyncGenerator[ModelResponseStream | TextCompletionResponse, None]: + async for chunk in stream(): + yield chunk + yield usage_chunk + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + ("stream", "usage_chunk"), + [ + ( + _text_completion_n2_stream_without_choice_1_finish, + TextCompletionResponse(id="cmpl-stream", object="text_completion", choices=[], usage=_STREAM_USAGE), + ), + (_chat_n2_stream_without_choice_1_finish, ModelResponseStream(id="chatcmpl-stream", choices=[], usage=_STREAM_USAGE)), + ], +) +async def test_streaming_hook_keeps_the_usage_chunk_last_after_releasing_a_held_tail( + stream: Callable[[], AsyncIterator[ModelResponseStream | TextCompletionResponse]], + usage_chunk: ModelResponseStream | TextCompletionResponse, +) -> None: + delivered: Final = [ + chunk + async for chunk in _fruit_word_filter().async_post_call_streaming_iterator_hook( + user_api_key_dict=MagicMock(), response=_then_usage_chunk(stream, usage_chunk), request_data={} + ) + ] + + assert delivered[-1] is usage_chunk + assert "".join(_streamed_text(chunk, 1) for chunk in delivered[:-1]) == _LONG_TAIL.replace( + "persimmon", "[KEYWORD_REDACTED]" + ) + + +@pytest.mark.asyncio +async def test_streaming_hook_relays_each_responses_api_event_as_it_arrives() -> None: + timeline: Final[list[str]] = [] + + async def responses_api_events() -> AsyncGenerator[OutputTextDeltaEvent, None]: + for delta in ("first", "second"): + timeline.append(f"upstream sent {delta}") + yield OutputTextDeltaEvent( + type=ResponsesAPIStreamEvents.OUTPUT_TEXT_DELTA, + item_id="msg_1", + output_index=0, + content_index=0, + delta=delta, + ) + + async for event in _fruit_word_filter().async_post_call_streaming_iterator_hook( + user_api_key_dict=MagicMock(), response=responses_api_events(), request_data={} + ): + timeline.append(f"client got {event.delta}") + + assert timeline == ["upstream sent first", "client got first", "upstream sent second", "client got second"] + + +@pytest.mark.asyncio +async def test_streaming_hook_blocks_a_word_split_across_text_completion_chunks(): + with pytest.raises(HTTPException) as exc_info: + async for _ in _fruit_word_filter().async_post_call_streaming_iterator_hook( + user_api_key_dict=MagicMock(), + response=_text_completion_stream(("A kum", None), ("quat tart", None), (None, "stop")), + request_data={}, + ): + pass + + assert exc_info.value.status_code == 400 + assert "kumquat" in str(exc_info.value.detail) diff --git a/tests/unit/proxy/guardrails/guardrail_hooks/test_grayswan.py b/tests/unit/proxy/guardrails/guardrail_hooks/test_grayswan.py index 954c57b5cb6..a5ba9dab0bd 100644 --- a/tests/unit/proxy/guardrails/guardrail_hooks/test_grayswan.py +++ b/tests/unit/proxy/guardrails/guardrail_hooks/test_grayswan.py @@ -5,6 +5,7 @@ import pytest from fastapi import HTTPException from litellm.integrations.custom_guardrail import ModifyResponseException +from litellm.llms.openai.completion.guardrail_translation.handler import OpenAITextCompletionHandler from litellm.proxy._types import UserAPIKeyAuth from litellm.proxy.guardrails.guardrail_hooks.grayswan import grayswan as grayswan_module from litellm.proxy.guardrails.guardrail_hooks.grayswan.grayswan import ( @@ -12,6 +13,7 @@ from litellm.proxy.guardrails.guardrail_hooks.grayswan.grayswan import ( GraySwanGuardrailAPIError, ) from litellm.types.guardrails import GuardrailEventHooks +from litellm.types.utils import TextChoices, TextCompletionResponse @pytest.fixture @@ -884,3 +886,29 @@ async def test_pre_call_payload_unchanged() -> None: {"role": "user", "content": "second"}, ] assert "tools" not in payload + + +@pytest.mark.asyncio +async def test_post_call_never_forwards_the_proxy_key_metadata_of_the_request_metadata() -> None: + guardrail = _post_call_guardrail() + client = _CapturingClient() + guardrail.async_handler = client + + await OpenAITextCompletionHandler().process_output_response( + response=TextCompletionResponse( + id="cmpl-1", model="gpt-4.1-mini", choices=[TextChoices(index=0, text="hello", finish_reason="stop")] + ), + guardrail_to_apply=guardrail, + user_api_key_dict=UserAPIKeyAuth(team_id="team-456", key_alias="ci-key", user_email="dev@example.com"), + request_data={ + "model": "gpt-4.1-mini", + "prompt": "say hello", + "metadata": { + "user_api_key_hash": "hashed-key", + "user_api_key_user_email": "dev@example.com", + "user_api_key_request_route": "/v1/completions", + }, + }, + ) + + assert "litellm_metadata" not in client.calls[0]["json"] diff --git a/tests/unit/proxy/guardrails/test_deferred_guardrail_logging.py b/tests/unit/proxy/guardrails/test_deferred_guardrail_logging.py index 7e5d0f6342e..bd6190fee5f 100644 --- a/tests/unit/proxy/guardrails/test_deferred_guardrail_logging.py +++ b/tests/unit/proxy/guardrails/test_deferred_guardrail_logging.py @@ -1537,6 +1537,27 @@ class TestArmDeferredStreamDispatch: assert recorded["cache_hit"] is False assert recorded["prefer_async_handlers"] is True + @pytest.mark.asyncio + async def test_text_completion_stream_logs_after_the_guardrails_finish(self): + from litellm.text_completions.streaming_iterator import TextCompletionStreamWrapper + + logging_obj, recorded = self._dispatch_recording_logging_obj() + + self._processor()._arm_deferred_stream_dispatch( + response=TextCompletionStreamWrapper(completion_stream=MagicMock(), model="gpt-test"), + route_type="atext_completion", + user_api_key_dict=MagicMock(), + logging_obj=logging_obj, + ) + + assembled = object() + logging_obj._deferred_stream_complete_args = (assembled, False) + ProxyLogging._fire_deferred_stream_logging({"litellm_logging_obj": logging_obj}) + await asyncio.sleep(0) + + assert recorded["result"] is assembled + assert recorded["cache_hit"] is False + @pytest.mark.asyncio async def test_router_wrapped_bridged_iterator_gets_csw_arg_shape(self): """The router wraps iterators without _hidden_params in diff --git a/tests/unit/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/unit/proxy/utils/proxy_logging/test_guardrail_pipeline.py index acc1463b072..cfd4fe197be 100644 --- a/tests/unit/proxy/utils/proxy_logging/test_guardrail_pipeline.py +++ b/tests/unit/proxy/utils/proxy_logging/test_guardrail_pipeline.py @@ -41,6 +41,7 @@ from litellm.types.proxy.policy_engine.pipeline_types import ( GuardrailPipeline, PipelineStep, ) +from litellm.types.utils import TextChoices, TextCompletionResponse @pytest.fixture(autouse=True) @@ -1886,7 +1887,7 @@ def test_streamable_post_call_pipelines_keeps_hook_guardrails_and_drops_iterator @pytest.mark.parametrize( "request_route", - ["/v1/completions", "/v1beta/models/gemini-2.5-flash:streamGenerateContent", "/a2a/agent"], + ["/v1beta/models/gemini-2.5-flash:streamGenerateContent", "/a2a/agent"], ) def test_streamable_post_call_pipelines_keeps_legacy_hooks_off_routes_that_assemble_no_response( make_user_api_key_auth, monkeypatch, caplog, request_route @@ -2036,6 +2037,120 @@ async def test_streaming_iterator_hook_delivers_legacy_hook_rewrite_on_every_cho assert _warnings(caplog) == [] +def _text_completion_chunk(text: str, finish_reason: str | None = None) -> TextCompletionResponse: + return TextCompletionResponse( + id="cmpl-1", model="m", choices=[TextChoices(index=0, text=text, finish_reason=finish_reason)] + ) + + +def _rewritten_text_completion(response: Any) -> TextCompletionResponse: + return TextCompletionResponse( + id=response.id, + model=response.model, + choices=[TextChoices(index=0, text="[REWRITTEN] " + response.choices[0].text, finish_reason="stop")], + ) + + +@pytest.mark.asyncio +async def test_streaming_iterator_hook_delivers_legacy_hook_rewrite_on_a_text_completion_stream( + proxy_logging, make_user_api_key_auth, monkeypatch, caplog +): + seen: Dict[str, Any] = {} + monkeypatch.setattr(litellm, "callbacks", [_legacy_hook_stream_guardrail(seen, rewrite=_rewritten_text_completion)]) + monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False) + data = {**_post_call_pipeline_data(stream=True), "prompt": "hi"} + chunks = [_text_completion_chunk("hello "), _text_completion_chunk("world"), _text_completion_chunk("", "stop")] + auth = make_user_api_key_auth(request_route="/v1/completions") + + with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): + await proxy_logging.pre_call_hook( + user_api_key_dict=auth, data=data, call_type="text_completion", guardrails_only=True + ) + delivered = [ + item + async for item in proxy_logging.async_post_call_streaming_iterator_hook( + user_api_key_dict=auth, response=_async_chunk_iter(chunks), request_data=data + ) + ] + + assert isinstance(seen["response"], TextCompletionResponse) + assert seen["response"].choices[0].text == "hello world" + assert [(item.choices[0].text, item.choices[0].finish_reason) for item in delivered] == [ + ("[REWRITTEN] hello world", None), + ("", None), + ("", "stop"), + ] + assert data["metadata"]["applied_guardrails"] == ["gr-post"] + assert _warnings(caplog) == [] + + +class _FlatBlockGuardrail(CustomGuardrail): + async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None): + if any("kumquat" in text for text in inputs.get("texts") or []): + raise HTTPException(status_code=400, detail="flat guardrail blocked kumquat") + return inputs + + +@pytest.mark.asyncio +async def test_streaming_iterator_hook_delivers_a_flat_guardrail_block_on_a_text_completion_stream_under_a_pipeline( + proxy_logging, make_user_api_key_auth, monkeypatch +): + seen: Final[Dict[str, Any]] = {} + flat: Final = _FlatBlockGuardrail( + guardrail_name="flat-block", event_hook=GuardrailEventHooks.post_call, default_on=True + ) + monkeypatch.setattr(litellm, "callbacks", [flat, _live_scan_guardrail(seen)]) + monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False) + data: Final = {**_post_call_pipeline_data(stream=True), "prompt": "hi"} + chunks: Final = [ + _text_completion_chunk("A kum"), + _text_completion_chunk("quat tart"), + _text_completion_chunk("", "stop"), + ] + auth: Final = make_user_api_key_auth(request_route="/v1/completions") + + await proxy_logging.pre_call_hook(user_api_key_dict=auth, data=data, call_type="text_completion", guardrails_only=True) + delivered: Final = [ + item + async for item in proxy_logging.async_post_call_streaming_iterator_hook( + user_api_key_dict=auth, response=_async_chunk_iter(chunks), request_data=data + ) + ] + + assert seen["texts"] == ["A kumquat tart"] + assert isinstance(delivered[-1], bytes) + assert json.loads(delivered[-1].decode().removeprefix("data: "))["error"]["message"] == ( + "flat guardrail blocked kumquat" + ) + + +@pytest.mark.asyncio +async def test_streaming_iterator_hook_live_pipeline_scans_what_a_disconnected_text_completion_client_received( + proxy_logging, make_user_api_key_auth, monkeypatch +): + seen: Final[Dict[str, Any]] = {} + monkeypatch.setattr(litellm, "callbacks", [_live_scan_guardrail(seen)]) + monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False) + data: Final = {**_post_call_pipeline_data(step=_DETECT_ONLY_STEP, stream=True), "prompt": "hi"} + chunks: Final = [ + _text_completion_chunk("hello "), + _text_completion_chunk("world"), + _text_completion_chunk("", "stop"), + ] + auth: Final = make_user_api_key_auth(request_route="/v1/completions") + + await proxy_logging.pre_call_hook(user_api_key_dict=auth, data=data, call_type="text_completion", guardrails_only=True) + stream: Final = proxy_logging.async_post_call_streaming_iterator_hook( + user_api_key_dict=auth, response=_async_chunk_iter(chunks), request_data=data + ) + first: Final = await stream.__anext__() + await stream.aclose() + + assert first.choices[0].text == "hello " + assert seen["texts"] == ["hello "] + assert _post_call_statuses(data) == ["success"] + + @pytest.mark.asyncio async def test_streaming_iterator_hook_releases_stream_untouched_when_legacy_hook_returns_none( proxy_logging, make_user_api_key_auth, monkeypatch @@ -2615,6 +2730,45 @@ async def test_streaming_iterator_hook_pipeline_modify_response_emits_translated assert not any(item is chunk for item in delivered for chunk in chunks) +@pytest.mark.asyncio +async def test_streaming_iterator_hook_pipeline_modify_response_replaces_a_text_completion_stream( + proxy_logging, make_user_api_key_auth, monkeypatch +): + seen: Final[Dict[str, Any]] = {} + monkeypatch.setattr(litellm, "callbacks", [_unified_stream_guardrail(seen, block=True)]) + monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False) + pipeline: Final = GuardrailPipeline( + mode="post_call", + steps=[ + PipelineStep( + guardrail="gr-post", + on_pass="allow", + on_fail="modify_response", + modify_response_message="content policy block", + ) + ], + ) + data: Final = {**_post_call_pipeline_data(stream=True), "prompt": "hi"} + data["metadata"]["_guardrail_pipelines"] = [("response-governance", pipeline)] + chunks: Final = [_text_completion_chunk("hello "), _text_completion_chunk("world"), _text_completion_chunk("", "stop")] + auth: Final = make_user_api_key_auth(request_route="/v1/completions") + + await proxy_logging.pre_call_hook(user_api_key_dict=auth, data=data, call_type="text_completion", guardrails_only=True) + delivered: Final = [ + item + async for item in proxy_logging.async_post_call_streaming_iterator_hook( + user_api_key_dict=auth, response=_async_chunk_iter(chunks), request_data=data + ) + ] + + frames: Final = [json.loads(item.decode().removeprefix("data: ")) for item in delivered] + assert seen["count"] == 1 + assert [(frame["choices"][0]["text"], frame["choices"][0].get("finish_reason")) for frame in frames] == [ + ("content policy block", None), + ("", "content_filter"), + ] + + @pytest.mark.asyncio async def test_streaming_iterator_hook_pipeline_delivers_text_rewrite_on_anthropic_sse( proxy_logging, make_user_api_key_auth, monkeypatch From c98641fd5ac1093c9196462a968dc634c6f70c94 Mon Sep 17 00:00:00 2001 From: Rick <26716961+Bytechoreographer@users.noreply.github.com> Date: Fri, 30 Oct 2026 00:08:18 +0800 Subject: [PATCH 4/7] feat(dashscope)!: derive protocol bases from one domain (#47812) * feat(dashscope): derive protocol bases from one domain Accept the DashScope domain root, the OpenAI-compatible base, the Anthropic Messages base, or a native /api/v1 base and derive every surface from it Adds a native Anthropic Messages config for /apps/anthropic so a single deployment serves chat, messages, embeddings, rerank, image generation, and video against the same credential Co-Authored-By: AaronHowell <111272993+AaronHowell@users.noreply.github.com> * fix(dashscope): keep Anthropic credentials off DashScope messages Require a DashScope key or a caller auth header before the Messages path runs, and opt out of workload identity, so the base class cannot fall back to ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN Also format the image generation base selection * style(dashscope): keep the messages config within the format gate * test(dashscope): cover caller Authorization on messages * fix(dashscope): keep the chat bridge for protocol-named bases and the image default for unconfigured calls * fix(dashscope): keep configured image and rerank bases, normalize blank keys, derive brand chat bases - image calls fall back to the provider's own chat base env (DASHSCOPE_API_BASE, QWENCLOUD_API_BASE, QWEN_AI_PLATFORM_API_BASE) before the image default - an explicit /compatible-api/v1/reranks URL wins over DASHSCOPE_API_BASE_RERANK again - a blank DashScope key counts as unset on /v1/messages, so it never falls back to ANTHROPIC_API_KEY - QwenCloud and Qianwen AI Platform chat and embeddings derive their compatible-mode base from a domain root - tests annotate new variables with Final and drop the config-class assertion * fix(dashscope): keep an explicit compatible-api rerank base as given * fix(dashscope): derive chat from compatible-api bases, keep the caller's rerank base, and let explicit bases beat per-endpoint env vars * fix(dashscope): keep a full DashScope image endpoint URL as given and type the brand tests * fix(dashscope): let the image env var beat any shared DashScope base * fix(dashscope): keep a full DASHSCOPE_API_BASE image URL ahead of the image env var * fix(dashscope): keep a native DashScope endpoint URL as given for chat and the api base header --------- Co-authored-by: AaronHowell <111272993+AaronHowell@users.noreply.github.com> Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- litellm/images/main.py | 12 +- litellm/llms/dashscope/chat/transformation.py | 9 +- litellm/llms/dashscope/common_utils.py | 136 ++++++++++- .../llms/dashscope/embed/transformation.py | 4 +- .../image_generation/transformation.py | 35 ++- litellm/llms/dashscope/messages/__init__.py | 7 + .../llms/dashscope/messages/transformation.py | 96 ++++++++ litellm/llms/dashscope/qwen_ai_platform.py | 16 +- litellm/llms/dashscope/qwencloud.py | 16 +- .../llms/dashscope/videos/transformation.py | 5 + litellm/rerank_api/main.py | 12 +- litellm/utils.py | 6 + .../test_websearch_short_circuit.py | 1 + .../test_dashscope_chat_transformation.py | 37 ++- .../dashscope/test_dashscope_common_utils.py | 171 ++++++++++++++ ...test_dashscope_embedding_transformation.py | 35 +++ .../test_dashscope_messages_transformation.py | 215 ++++++++++++++++++ .../test_dashscope_rerank_transformation.py | 52 +++++ .../test_dashscope_video_transformation.py | 14 ++ .../llms/dashscope/test_qwen_brand_aliases.py | 46 ++++ tests/unit/test_dashscope_image_generation.py | 127 ++++++++++- 21 files changed, 1015 insertions(+), 37 deletions(-) create mode 100644 litellm/llms/dashscope/messages/__init__.py create mode 100644 litellm/llms/dashscope/messages/transformation.py create mode 100644 tests/unit/llms/dashscope/test_dashscope_common_utils.py create mode 100644 tests/unit/llms/dashscope/test_dashscope_messages_transformation.py diff --git a/litellm/images/main.py b/litellm/images/main.py index b3247ff938b..4aad5210028 100644 --- a/litellm/images/main.py +++ b/litellm/images/main.py @@ -68,6 +68,11 @@ from litellm.utils import ( ) _PROVIDER_ROUTING_HINTS: Final = TypeAdapter(ProviderRoutingHints) +_PROVIDERS_RESOLVING_THEIR_OWN_IMAGE_BASE: Final = ( + LlmProviders.DASHSCOPE, + LlmProviders.QWENCLOUD, + LlmProviders.QWEN_AI_PLATFORM, +) # Cache for ImageEditRequestUtils to avoid repeated __getattr__ calls _ImageEditRequestUtils_cache: type["ImageEditRequestUtils"] | None = None @@ -271,6 +276,7 @@ def image_generation( """ try: args: Final = locals() + requested_api_base: Final = api_base aimg_generation: Final = kwargs.get("aimg_generation", False) litellm_call_id: Final = kwargs.get("litellm_call_id", None) logger_fn: Final = kwargs.get("logger_fn", None) @@ -468,7 +474,11 @@ def image_generation( raise ValueError(f"image generation config is not supported for {custom_llm_provider}") # Resolve api_base from litellm.api_base if not explicitly provided - _api_base: Final = api_base or litellm.api_base + _api_base: Final = ( + requested_api_base + if custom_llm_provider in _PROVIDERS_RESOLVING_THEIR_OWN_IMAGE_BASE + else api_base or litellm.api_base + ) litellm_params_dict["api_base"] = _api_base return llm_http_handler.image_generation_handler( diff --git a/litellm/llms/dashscope/chat/transformation.py b/litellm/llms/dashscope/chat/transformation.py index 966d1a01b8d..dfff9cec298 100644 --- a/litellm/llms/dashscope/chat/transformation.py +++ b/litellm/llms/dashscope/chat/transformation.py @@ -10,6 +10,7 @@ from litellm.secret_managers.main import get_secret_str from litellm.types.llms.openai import AllMessageValues, ChatCompletionToolParam from ...openai.chat.gpt_transformation import OpenAIGPTConfig +from ..common_utils import resolve_dashscope_openai_api_base class DashScopeChatConfig(OpenAIGPTConfig): @@ -55,11 +56,9 @@ class DashScopeChatConfig(OpenAIGPTConfig): def _get_openai_compatible_provider_info( self, api_base: str | None, api_key: str | None ) -> tuple[str | None, str | None]: - api_base = ( - api_base or get_secret_str("DASHSCOPE_API_BASE") or "https://dashscope.aliyuncs.com/compatible-mode/v1" - ) + resolved_api_base: Final = resolve_dashscope_openai_api_base(api_base) dynamic_api_key: Final = api_key or get_secret_str("DASHSCOPE_API_KEY") - return api_base, dynamic_api_key + return resolved_api_base, dynamic_api_key def get_openai_compatible_provider_info( self, @@ -69,7 +68,7 @@ class DashScopeChatConfig(OpenAIGPTConfig): return self._get_openai_compatible_provider_info(api_base, api_key) def _resolve_chat_api_base(self, api_base: str | None) -> str: - return api_base or "https://dashscope.aliyuncs.com/compatible-mode/v1" + return resolve_dashscope_openai_api_base(api_base) def get_complete_url( self, diff --git a/litellm/llms/dashscope/common_utils.py b/litellm/llms/dashscope/common_utils.py index 3e2e718a469..5ad2aa17a3a 100644 --- a/litellm/llms/dashscope/common_utils.py +++ b/litellm/llms/dashscope/common_utils.py @@ -2,6 +2,7 @@ Common utilities for the DashScope LLM provider. """ +from dataclasses import dataclass from typing import TYPE_CHECKING, Final from urllib.parse import urlparse @@ -19,10 +20,123 @@ if TYPE_CHECKING: from litellm.llms.base_llm.videos.transformation import BaseVideoConfig DASHSCOPE_CHAT_COMPATIBLE_PATH: Final = "/compatible-mode/v1" -DASHSCOPE_RERANK_PATH: Final = "/compatible-api/v1/reranks" +DASHSCOPE_ANTHROPIC_PATH: Final = "/apps/anthropic" +DASHSCOPE_NATIVE_V1_PATH: Final = "/api/v1" +DASHSCOPE_COMPATIBLE_API_PATH: Final = "/compatible-api/v1" +DASHSCOPE_RERANK_PATH: Final = f"{DASHSCOPE_COMPATIBLE_API_PATH}/reranks" +DASHSCOPE_IMAGE_PATH: Final = "/api/v1/services/aigc/multimodal-generation/generation" +DASHSCOPE_DEFAULT_ORIGIN: Final = "https://dashscope.aliyuncs.com" + +_STANDARD_DASHSCOPE_HOSTS: Final = frozenset( + { + "dashscope.aliyuncs.com", + "dashscope-intl.aliyuncs.com", + "dashscope-us.aliyuncs.com", + } +) +_WORKSPACE_HOST_SUFFIX: Final = ".maas.aliyuncs.com" +_EXCLUDED_WORKSPACE_PREFIXES: Final = ("trial.", "token-plan.") +_PROTOCOL_PATH_PREFIXES: Final = ( + DASHSCOPE_CHAT_COMPATIBLE_PATH, + DASHSCOPE_ANTHROPIC_PATH, + DASHSCOPE_NATIVE_V1_PATH, + DASHSCOPE_COMPATIBLE_API_PATH, +) -def _rerank_base_for_chat_shaped_base(api_base: str | None) -> str | None: +@dataclass(frozen=True, slots=True) +class DashScopeEndpoints: + origin: str + openai: str + anthropic: str + native: str + + +def _is_dashscope_host(host: str) -> bool: + if host in _STANDARD_DASHSCOPE_HOSTS: + return True + return host.endswith(_WORKSPACE_HOST_SUFFIX) and not host.startswith(_EXCLUDED_WORKSPACE_PREFIXES) + + +def _dashscope_origin(api_base: str | None) -> str | None: + if not api_base: + return None + parsed: Final = urlparse(api_base) + if parsed.scheme not in ("http", "https") or not parsed.netloc: + return None + if parsed.query or parsed.fragment: + return None + host: Final = (parsed.hostname or "").lower() + if not _is_dashscope_host(host): + return None + path: Final = parsed.path.rstrip("/") + if path and not any(path == prefix or path.startswith(f"{prefix}/") for prefix in _PROTOCOL_PATH_PREFIXES): + return None + return f"{parsed.scheme}://{parsed.netloc}" + + +def resolve_dashscope_endpoints(api_base: str | None) -> DashScopeEndpoints | None: + """ + Derive DashScope's protocol-specific bases from any standard DashScope base. + + Accepts the domain root, an OpenAI-compatible base, an Anthropic Messages + base, or a native `/api/v1` base for `dashscope.aliyuncs.com`, + `dashscope-intl.aliyuncs.com`, `dashscope-us.aliyuncs.com`, and + `{workspace}.{region}.maas.aliyuncs.com`. Returns None for any other host + or path so callers keep the caller-supplied base untouched. + """ + origin: Final = _dashscope_origin(api_base) + if origin is None: + return None + return DashScopeEndpoints( + origin=origin, + openai=f"{origin}{DASHSCOPE_CHAT_COMPATIBLE_PATH}", + anthropic=f"{origin}{DASHSCOPE_ANTHROPIC_PATH}", + native=f"{origin}{DASHSCOPE_NATIVE_V1_PATH}", + ) + + +def is_dashscope_native_endpoint_url(api_base: str) -> bool: + return urlparse(api_base).path.rstrip("/").startswith(f"{DASHSCOPE_NATIVE_V1_PATH}/") + + +def resolve_dashscope_openai_api_base( + api_base: str | None, + env_var: str = "DASHSCOPE_API_BASE", + default: str = f"{DASHSCOPE_DEFAULT_ORIGIN}{DASHSCOPE_CHAT_COMPATIBLE_PATH}", +) -> str: + candidate: Final = api_base or get_secret_str(env_var) + if candidate and is_dashscope_native_endpoint_url(candidate): + return candidate + endpoints: Final = resolve_dashscope_endpoints(candidate) + if endpoints is not None: + return endpoints.openai + return candidate or default + + +def resolve_dashscope_anthropic_api_base( + api_base: str | None, + env_var: str = "DASHSCOPE_API_BASE", + default: str = f"{DASHSCOPE_DEFAULT_ORIGIN}{DASHSCOPE_ANTHROPIC_PATH}", +) -> str: + candidate: Final = api_base or get_secret_str(env_var) + endpoints: Final = resolve_dashscope_endpoints(candidate) + if endpoints is not None: + return endpoints.anthropic + return candidate or default + + +def dashscope_base_selects_anthropic_messages(api_base: str | None) -> bool: + candidate: Final = api_base or get_secret_str("DASHSCOPE_API_BASE") + if not candidate: + return False + path: Final = urlparse(candidate).path.rstrip("/") + if path.endswith(DASHSCOPE_ANTHROPIC_PATH) or f"{DASHSCOPE_ANTHROPIC_PATH}/" in path: + return True + return not path and resolve_dashscope_endpoints(candidate) is not None + + +def _chat_shaped_aliyun_origin(api_base: str | None) -> str | None: if api_base is None: return None parsed: Final = urlparse(api_base) @@ -30,14 +144,24 @@ def _rerank_base_for_chat_shaped_base(api_base: str | None) -> str | None: on_aliyun_host: Final = host == "aliyuncs.com" or host.endswith(".aliyuncs.com") if not on_aliyun_host or parsed.path.rstrip("/") != DASHSCOPE_CHAT_COMPATIBLE_PATH: return None - return f"{parsed.scheme}://{parsed.netloc}{DASHSCOPE_RERANK_PATH}" + return f"{parsed.scheme}://{parsed.netloc}" + + +def _is_compatible_api_base(api_base: str) -> bool: + path: Final = urlparse(api_base).path.rstrip("/") + return path == DASHSCOPE_COMPATIBLE_API_PATH or path.startswith(f"{DASHSCOPE_COMPATIBLE_API_PATH}/") def resolve_dashscope_family_rerank_api_base(api_base: str | None, env_var: str, default_rerank_base: str) -> str: - remapped: Final = _rerank_base_for_chat_shaped_base(api_base) - if api_base is not None and remapped is None: + if api_base is None: + return get_secret_str(env_var) or default_rerank_base + if _is_compatible_api_base(api_base): return api_base - return get_secret_str(env_var) or remapped or default_rerank_base + chat_origin: Final = _chat_shaped_aliyun_origin(api_base) + if chat_origin is not None: + return get_secret_str(env_var) or f"{chat_origin}{DASHSCOPE_RERANK_PATH}" + endpoints: Final = resolve_dashscope_endpoints(api_base) + return f"{endpoints.origin}{DASHSCOPE_RERANK_PATH}" if endpoints is not None else api_base def get_dashscope_family_embedding_config(custom_llm_provider: str) -> "BaseEmbeddingConfig": diff --git a/litellm/llms/dashscope/embed/transformation.py b/litellm/llms/dashscope/embed/transformation.py index 35132346261..0f01a4b3172 100644 --- a/litellm/llms/dashscope/embed/transformation.py +++ b/litellm/llms/dashscope/embed/transformation.py @@ -25,7 +25,7 @@ from litellm.secret_managers.main import get_secret_str from litellm.types.llms.openai import AllEmbeddingInputValues, AllMessageValues from litellm.types.utils import EmbeddingResponse, Usage -from ..common_utils import DashScopeError +from ..common_utils import DashScopeError, resolve_dashscope_openai_api_base DEFAULT_API_BASE: Final = "https://dashscope.aliyuncs.com/compatible-mode/v1" @@ -91,7 +91,7 @@ class DashScopeEmbeddingConfig(BaseEmbeddingConfig): return resolved_api_key def _resolve_embedding_api_base(self, api_base: str | None) -> str: - return api_base or get_secret_str("DASHSCOPE_API_BASE") or DEFAULT_API_BASE + return resolve_dashscope_openai_api_base(api_base, default=DEFAULT_API_BASE) def validate_environment( self, diff --git a/litellm/llms/dashscope/image_generation/transformation.py b/litellm/llms/dashscope/image_generation/transformation.py index d1faa6c6cd0..2f528b5a2c3 100644 --- a/litellm/llms/dashscope/image_generation/transformation.py +++ b/litellm/llms/dashscope/image_generation/transformation.py @@ -39,6 +39,13 @@ from litellm.types.llms.openai import ( ) from litellm.types.utils import ImageObject, ImageResponse +from ..common_utils import ( + DASHSCOPE_CHAT_COMPATIBLE_PATH, + DASHSCOPE_IMAGE_PATH, + is_dashscope_native_endpoint_url, + resolve_dashscope_endpoints, +) + if TYPE_CHECKING: from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj from litellm.litellm_core_utils.tokenizer import Encoding as Tokenizer @@ -53,8 +60,6 @@ _DECODED_JSON: Final = TypeAdapter(object) DEFAULT_API_BASE: Final = "https://dashscope-intl.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation" -CHAT_COMPATIBLE_MODE_PATH: Final = "/compatible-mode/v1" - # Maps OpenAI size strings (WxH) to DashScope size strings (W*H) OPENAI_TO_DASHSCOPE_SIZE: Final[dict] = { "256x256": "256*256", @@ -102,8 +107,14 @@ class DashScopeImageGenerationConfig(BaseImageGenerationConfig): raise ValueError("DASHSCOPE_API_KEY is not set") return resolved_api_key - def _resolve_image_api_base(self, image_api_base: str | None) -> str: - return image_api_base or get_secret_str("DASHSCOPE_API_BASE_IMAGE") or DEFAULT_API_BASE + def _image_api_base_env(self) -> str | None: + return get_secret_str("DASHSCOPE_API_BASE_IMAGE") + + def _default_image_api_base(self) -> str: + return DEFAULT_API_BASE + + def _configured_chat_api_base(self) -> str | None: + return get_secret_str("DASHSCOPE_API_BASE") def get_complete_url( self, @@ -114,10 +125,18 @@ class DashScopeImageGenerationConfig(BaseImageGenerationConfig): litellm_params: dict, stream: bool | None = None, ) -> str: - image_api_base: Final = ( - api_base if api_base and not api_base.rstrip("/").endswith(CHAT_COMPATIBLE_MODE_PATH) else None - ) - return self._resolve_image_api_base(image_api_base) + image_env_api_base: Final = self._image_api_base_env() + configured_api_base: Final = api_base or self._configured_chat_api_base() + if not configured_api_base or configured_api_base.rstrip("/").endswith(DASHSCOPE_CHAT_COMPATIBLE_PATH): + return image_env_api_base or self._default_image_api_base() + if is_dashscope_native_endpoint_url(configured_api_base): + return configured_api_base + endpoints: Final = resolve_dashscope_endpoints(configured_api_base) + if endpoints is None: + return configured_api_base + if not api_base and image_env_api_base: + return image_env_api_base + return f"{endpoints.origin}{DASHSCOPE_IMAGE_PATH}" def validate_environment( self, diff --git a/litellm/llms/dashscope/messages/__init__.py b/litellm/llms/dashscope/messages/__init__.py new file mode 100644 index 00000000000..08cd184404c --- /dev/null +++ b/litellm/llms/dashscope/messages/__init__.py @@ -0,0 +1,7 @@ +""" +DashScope Anthropic-compatible Messages Module +""" + +from .transformation import DashScopeAnthropicMessagesConfig + +__all__ = ["DashScopeAnthropicMessagesConfig"] diff --git a/litellm/llms/dashscope/messages/transformation.py b/litellm/llms/dashscope/messages/transformation.py new file mode 100644 index 00000000000..9d0d1db91c8 --- /dev/null +++ b/litellm/llms/dashscope/messages/transformation.py @@ -0,0 +1,96 @@ +""" +DashScope Anthropic-compatible Messages transformation config. +""" + +from typing import ClassVar, Final + +import litellm +from litellm.exceptions import AuthenticationError +from litellm.litellm_core_utils.url_utils import ensure_api_route +from litellm.llms.anthropic.messages.native.transformation import ( + AnthropicMessagesConfig, +) +from litellm.secret_managers.main import get_secret_str, normalize_nonempty_secret_str + +from ..common_utils import ( + dashscope_base_selects_anthropic_messages, + resolve_dashscope_anthropic_api_base, +) + + +class DashScopeAnthropicMessagesConfig(AnthropicMessagesConfig): + """ + DashScope exposes an Anthropic-compatible Messages API at + ``{origin}/apps/anthropic`` for every DashScope domain, including + ``dashscope.aliyuncs.com`` and ``{workspace}.{region}.maas.aliyuncs.com``. + """ + + _workload_identity_eligible: ClassVar[bool] = False + + @property + def custom_llm_provider(self) -> str | None: + return "dashscope" + + def should_strip_billing_metadata(self) -> bool: + return True + + def handles_web_search_natively(self) -> bool: + return False + + def serves_api_base(self, model: str, api_base: str | None) -> bool: + return dashscope_base_selects_anthropic_messages(api_base) + + @staticmethod + def get_api_key(api_key: str | None = None) -> str | None: + return ( + normalize_nonempty_secret_str(api_key) + or normalize_nonempty_secret_str(get_secret_str("DASHSCOPE_API_KEY")) + or normalize_nonempty_secret_str(litellm.api_key) + ) + + @staticmethod + def get_api_base(api_base: str | None = None) -> str: + return resolve_dashscope_anthropic_api_base(api_base) + + def validate_anthropic_messages_environment( + self, + headers: dict[str, str], # mutable-ok: mirrors the legacy base override signature + model: str, + messages: list[object], # mutable-ok: mirrors the legacy base override signature + optional_params: dict[str, object], # mutable-ok: mirrors the legacy base override signature + litellm_params: dict[str, object], # mutable-ok: mirrors the legacy base override signature + api_key: str | None = None, + api_base: str | None = None, + ) -> tuple[dict[str, str], str | None]: # mutable-ok: mirrors the base signature + resolved_api_key: Final = self.get_api_key(api_key=api_key) + if resolved_api_key is None: + present_headers: Final = {name.lower() for name in headers} + if "x-api-key" not in present_headers and "authorization" not in present_headers: + raise AuthenticationError( + message=( + "Missing DashScope API key. Set DASHSCOPE_API_KEY or pass api_key explicitly. " + "Anthropic credentials are never sent to a DashScope endpoint." + ), + llm_provider="dashscope", + model=model, + ) + return super().validate_anthropic_messages_environment( # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType, reportReturnType] # base signature uses an unparameterized dict + headers=headers, + model=model, + messages=messages, + optional_params=optional_params, + litellm_params=litellm_params, + api_key=resolved_api_key, + api_base=api_base, + ) + + def get_complete_url( + self, + api_base: str | None, + api_key: str | None, + model: str, + optional_params: dict[str, object], # mutable-ok: mirrors the base get_complete_url signature + litellm_params: dict[str, object], # mutable-ok: mirrors the base get_complete_url signature + stream: bool | None = None, + ) -> str: + return ensure_api_route(self.get_api_base(api_base=api_base), "v1/messages") diff --git a/litellm/llms/dashscope/qwen_ai_platform.py b/litellm/llms/dashscope/qwen_ai_platform.py index f7cc50cc82a..db01c8a0a15 100644 --- a/litellm/llms/dashscope/qwen_ai_platform.py +++ b/litellm/llms/dashscope/qwen_ai_platform.py @@ -3,7 +3,7 @@ from typing import Final from litellm.secret_managers.main import get_secret_str from .chat.transformation import DashScopeChatConfig -from .common_utils import resolve_dashscope_family_rerank_api_base +from .common_utils import resolve_dashscope_family_rerank_api_base, resolve_dashscope_openai_api_base from .embed.transformation import DashScopeEmbeddingConfig from .image_generation.transformation import DashScopeImageGenerationConfig from .rerank.transformation import DashScopeRerankConfig @@ -38,7 +38,7 @@ class QwenAIPlatformChatConfig(DashScopeChatConfig): return self._resolve_chat_api_base(api_base), _resolve_qwen_ai_platform_api_key(api_key) def _resolve_chat_api_base(self, api_base: str | None) -> str: - return api_base or get_secret_str("QWEN_AI_PLATFORM_API_BASE") or QWEN_AI_PLATFORM_API_BASE + return resolve_dashscope_openai_api_base(api_base, "QWEN_AI_PLATFORM_API_BASE", QWEN_AI_PLATFORM_API_BASE) class QwenAIPlatformEmbeddingConfig(DashScopeEmbeddingConfig): @@ -46,7 +46,7 @@ class QwenAIPlatformEmbeddingConfig(DashScopeEmbeddingConfig): return _require_qwen_ai_platform_api_key(api_key) def _resolve_embedding_api_base(self, api_base: str | None) -> str: - return api_base or get_secret_str("QWEN_AI_PLATFORM_API_BASE") or QWEN_AI_PLATFORM_API_BASE + return resolve_dashscope_openai_api_base(api_base, "QWEN_AI_PLATFORM_API_BASE", QWEN_AI_PLATFORM_API_BASE) class QwenAIPlatformRerankConfig(DashScopeRerankConfig): @@ -63,8 +63,14 @@ class QwenAIPlatformImageGenerationConfig(DashScopeImageGenerationConfig): def _resolve_api_key(self, api_key: str | None) -> str: return _require_qwen_ai_platform_api_key(api_key) - def _resolve_image_api_base(self, image_api_base: str | None) -> str: - return image_api_base or get_secret_str("QWEN_AI_PLATFORM_API_BASE_IMAGE") or QWEN_AI_PLATFORM_IMAGE_API_BASE + def _image_api_base_env(self) -> str | None: + return get_secret_str("QWEN_AI_PLATFORM_API_BASE_IMAGE") + + def _default_image_api_base(self) -> str: + return QWEN_AI_PLATFORM_IMAGE_API_BASE + + def _configured_chat_api_base(self) -> str | None: + return get_secret_str("QWEN_AI_PLATFORM_API_BASE") class QwenAIPlatformVideoConfig(DashScopeVideoConfig): diff --git a/litellm/llms/dashscope/qwencloud.py b/litellm/llms/dashscope/qwencloud.py index a12e21b0983..6ae42c62e42 100644 --- a/litellm/llms/dashscope/qwencloud.py +++ b/litellm/llms/dashscope/qwencloud.py @@ -3,7 +3,7 @@ from typing import Final from litellm.secret_managers.main import get_secret_str from .chat.transformation import DashScopeChatConfig -from .common_utils import resolve_dashscope_family_rerank_api_base +from .common_utils import resolve_dashscope_family_rerank_api_base, resolve_dashscope_openai_api_base from .embed.transformation import DashScopeEmbeddingConfig from .image_generation.transformation import DashScopeImageGenerationConfig from .rerank.transformation import DashScopeRerankConfig @@ -38,7 +38,7 @@ class QwenCloudChatConfig(DashScopeChatConfig): return self._resolve_chat_api_base(api_base), _resolve_qwencloud_api_key(api_key) def _resolve_chat_api_base(self, api_base: str | None) -> str: - return api_base or get_secret_str("QWENCLOUD_API_BASE") or QWENCLOUD_API_BASE + return resolve_dashscope_openai_api_base(api_base, "QWENCLOUD_API_BASE", QWENCLOUD_API_BASE) class QwenCloudEmbeddingConfig(DashScopeEmbeddingConfig): @@ -46,7 +46,7 @@ class QwenCloudEmbeddingConfig(DashScopeEmbeddingConfig): return _require_qwencloud_api_key(api_key) def _resolve_embedding_api_base(self, api_base: str | None) -> str: - return api_base or get_secret_str("QWENCLOUD_API_BASE") or QWENCLOUD_API_BASE + return resolve_dashscope_openai_api_base(api_base, "QWENCLOUD_API_BASE", QWENCLOUD_API_BASE) class QwenCloudRerankConfig(DashScopeRerankConfig): @@ -63,8 +63,14 @@ class QwenCloudImageGenerationConfig(DashScopeImageGenerationConfig): def _resolve_api_key(self, api_key: str | None) -> str: return _require_qwencloud_api_key(api_key) - def _resolve_image_api_base(self, image_api_base: str | None) -> str: - return image_api_base or get_secret_str("QWENCLOUD_API_BASE_IMAGE") or QWENCLOUD_IMAGE_API_BASE + def _image_api_base_env(self) -> str | None: + return get_secret_str("QWENCLOUD_API_BASE_IMAGE") + + def _default_image_api_base(self) -> str: + return QWENCLOUD_IMAGE_API_BASE + + def _configured_chat_api_base(self) -> str | None: + return get_secret_str("QWENCLOUD_API_BASE") class QwenCloudVideoConfig(DashScopeVideoConfig): diff --git a/litellm/llms/dashscope/videos/transformation.py b/litellm/llms/dashscope/videos/transformation.py index 5c738423a5a..59e42e7b926 100644 --- a/litellm/llms/dashscope/videos/transformation.py +++ b/litellm/llms/dashscope/videos/transformation.py @@ -38,6 +38,8 @@ from litellm.types.videos.utils import ( ) from litellm.utils import get_model_info +from ..common_utils import resolve_dashscope_endpoints + if TYPE_CHECKING: from litellm.litellm_core_utils.litellm_logging import Logging from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler @@ -206,6 +208,9 @@ def _parse_task_response(raw_response: httpx.Response) -> _TaskResponse: def _normalized_api_base(api_base: str) -> str: + endpoints: Final = resolve_dashscope_endpoints(api_base) + if endpoints is not None: + return endpoints.origin trimmed: Final = api_base.rstrip("/") if trimmed.endswith(DASHSCOPE_COMPATIBLE_MODE_PATH): return trimmed[: -len(DASHSCOPE_COMPATIBLE_MODE_PATH)] diff --git a/litellm/rerank_api/main.py b/litellm/rerank_api/main.py index 65d9b89cf79..e344ce46be7 100644 --- a/litellm/rerank_api/main.py +++ b/litellm/rerank_api/main.py @@ -19,6 +19,12 @@ from litellm.types.rerank import RerankResponse from litellm.types.router import * from litellm.utils import ProviderConfigManager, client, exception_type +_PROVIDERS_RESOLVING_THEIR_OWN_RERANK_BASE: Final = ( + litellm.LlmProviders.DASHSCOPE, + litellm.LlmProviders.QWENCLOUD, + litellm.LlmProviders.QWEN_AI_PLATFORM, +) + ####### ENVIRONMENT VARIABLES ################### # Initialize any necessary instances or variables here together_rerank: Final = TogetherAIRerank() @@ -543,7 +549,11 @@ def rerank( logging_obj=litellm_logging_obj, timeout=optional_params.timeout, api_key=dynamic_api_key or optional_params.api_key, - api_base=dynamic_api_base or optional_params.api_base, + api_base=( + optional_params.api_base or dynamic_api_base + if _custom_llm_provider in _PROVIDERS_RESOLVING_THEIR_OWN_RERANK_BASE + else dynamic_api_base or optional_params.api_base + ), _is_async=_is_async, headers=headers or litellm.headers or {}, client=client, diff --git a/litellm/utils.py b/litellm/utils.py index 8ea4f6ceda0..4cd2335dd0c 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -10048,6 +10048,12 @@ class ProviderConfigManager: ) return TencentAnthropicMessagesConfig() + elif litellm.LlmProviders.DASHSCOPE == provider: + from litellm.llms.dashscope.messages.transformation import ( + DashScopeAnthropicMessagesConfig, + ) + + return DashScopeAnthropicMessagesConfig() elif litellm.LlmProviders.GITHUB_COPILOT == provider: if "claude" in model_lower: from litellm.llms.github_copilot.messages.transformation import ( diff --git a/tests/unit/integrations/websearch_interception/test_websearch_short_circuit.py b/tests/unit/integrations/websearch_interception/test_websearch_short_circuit.py index 414bcc11602..ef050d96dc5 100644 --- a/tests/unit/integrations/websearch_interception/test_websearch_short_circuit.py +++ b/tests/unit/integrations/websearch_interception/test_websearch_short_circuit.py @@ -32,6 +32,7 @@ class TestTryShortCircuitSearch: ("github_copilot", "github_copilot/claude-sonnet-5.5"), ("databricks", "databricks/databricks-claude-sonnet-5-5"), ("databricks", "databricks/system.ai.claude-sonnet-5-5"), + ("dashscope", "dashscope/qwen3.7-plus"), ], ) async def test_short_circuits_single_web_search_tool(self, provider, model): diff --git a/tests/unit/llms/dashscope/test_dashscope_chat_transformation.py b/tests/unit/llms/dashscope/test_dashscope_chat_transformation.py index a3fd5933a83..99a8004b0a1 100644 --- a/tests/unit/llms/dashscope/test_dashscope_chat_transformation.py +++ b/tests/unit/llms/dashscope/test_dashscope_chat_transformation.py @@ -5,15 +5,50 @@ These tests validate the DashScopeConfig class which extends OpenAIGPTConfig. DashScope is an OpenAI-compatible provider with minor customizations. """ +from typing import Final +import pytest from litellm.types.llms.openai import AllMessageValues -import pytest import litellm from litellm import completion from litellm.llms.dashscope.chat.transformation import DashScopeChatConfig +WORKSPACE: Final = "https://llm-abc123.cn-beijing.maas.aliyuncs.com" + + +def test_get_complete_url_derives_the_compatible_base_from_any_standard_base(): + config: Final = DashScopeChatConfig() + + for api_base in ( + WORKSPACE, + f"{WORKSPACE}/compatible-mode/v1", + f"{WORKSPACE}/apps/anthropic", + f"{WORKSPACE}/api/v1", + ): + url: Final = config.get_complete_url( + api_base=api_base, + api_key="sk-test", + model="qwen3.7-plus", + optional_params={}, + litellm_params={}, + ) + + assert url == f"{WORKSPACE}/compatible-mode/v1/chat/completions" + + +def test_get_complete_url_keeps_unrecognized_hosts(): + url: Final = DashScopeChatConfig().get_complete_url( + api_base="https://gateway.example.com/v1", + api_key="sk-test", + model="qwen3.7-plus", + optional_params={}, + litellm_params={}, + ) + + assert url == "https://gateway.example.com/v1/chat/completions" + class TestDashScopeConfig: """Test class for DashScope functionality""" diff --git a/tests/unit/llms/dashscope/test_dashscope_common_utils.py b/tests/unit/llms/dashscope/test_dashscope_common_utils.py new file mode 100644 index 00000000000..2dc4d5e5a85 --- /dev/null +++ b/tests/unit/llms/dashscope/test_dashscope_common_utils.py @@ -0,0 +1,171 @@ +""" +Unit tests for DashScope's protocol-specific base URL derivation. +""" + +from typing import Final + +import pytest + +from litellm.llms.dashscope.common_utils import ( + DASHSCOPE_DEFAULT_ORIGIN, + resolve_dashscope_anthropic_api_base, + resolve_dashscope_endpoints, + resolve_dashscope_family_rerank_api_base, + resolve_dashscope_openai_api_base, +) + +WORKSPACE: Final = "https://llm-abc123.cn-beijing.maas.aliyuncs.com" + + +@pytest.mark.parametrize( + "api_base", + [ + WORKSPACE, + f"{WORKSPACE}/", + f"{WORKSPACE}/compatible-mode/v1", + f"{WORKSPACE}/compatible-mode/v1/chat/completions", + f"{WORKSPACE}/apps/anthropic", + f"{WORKSPACE}/apps/anthropic/v1/messages", + f"{WORKSPACE}/api/v1", + f"{WORKSPACE}/api/v1/services/aigc/video-generation/video-synthesis", + ], +) +def test_workspace_base_derives_every_protocol(api_base: str): + endpoints: Final = resolve_dashscope_endpoints(api_base) + + assert endpoints is not None + assert endpoints.origin == WORKSPACE + assert endpoints.openai == f"{WORKSPACE}/compatible-mode/v1" + assert endpoints.anthropic == f"{WORKSPACE}/apps/anthropic" + assert endpoints.native == f"{WORKSPACE}/api/v1" + + +@pytest.mark.parametrize( + "origin", + [ + "https://dashscope.aliyuncs.com", + "https://dashscope-intl.aliyuncs.com", + "https://dashscope-us.aliyuncs.com", + ], +) +@pytest.mark.parametrize("path", ["", "/", "/compatible-mode/v1"]) +def test_standard_hosts_derive_every_protocol(origin: str, path: str): + endpoints: Final = resolve_dashscope_endpoints(f"{origin}{path}") + + assert endpoints is not None + assert endpoints.openai == f"{origin}/compatible-mode/v1" + assert endpoints.anthropic == f"{origin}/apps/anthropic" + assert endpoints.native == f"{origin}/api/v1" + + +@pytest.mark.parametrize( + "api_base", + [ + None, + "", + "not-a-url", + "https://gateway.example.com", + "https://gateway.example.com/compatible-mode/v1", + f"{WORKSPACE}/custom-path", + f"{WORKSPACE}/compatible-mode/v1?api-version=1", + "https://trial.cn-beijing.maas.aliyuncs.com", + "https://token-plan.cn-beijing.maas.aliyuncs.com", + ], +) +def test_unrecognized_bases_are_not_rewritten(api_base: str | None): + assert resolve_dashscope_endpoints(api_base) is None + + +def test_openai_base_derives_from_the_domain(monkeypatch: pytest.MonkeyPatch): + monkeypatch.delenv("DASHSCOPE_API_BASE", raising=False) + + assert resolve_dashscope_openai_api_base(WORKSPACE) == f"{WORKSPACE}/compatible-mode/v1" + assert resolve_dashscope_openai_api_base(f"{WORKSPACE}/apps/anthropic") == f"{WORKSPACE}/compatible-mode/v1" + assert ( + resolve_dashscope_openai_api_base(f"{WORKSPACE}/compatible-api/v1/reranks") == f"{WORKSPACE}/compatible-mode/v1" + ) + + +@pytest.mark.parametrize( + "endpoint_url", + [ + "https://dashscope-intl.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation", + f"{WORKSPACE}/api/v1/services/aigc/video-generation/video-synthesis", + ], +) +def test_openai_base_keeps_a_native_endpoint_url(monkeypatch: pytest.MonkeyPatch, endpoint_url: str): + monkeypatch.setenv("DASHSCOPE_API_BASE", endpoint_url) + + assert resolve_dashscope_openai_api_base(None) == endpoint_url + assert resolve_dashscope_openai_api_base(endpoint_url) == endpoint_url + assert resolve_dashscope_openai_api_base(f"{WORKSPACE}/api/v1") == f"{WORKSPACE}/compatible-mode/v1" + + +def test_openai_base_keeps_unrecognized_hosts(monkeypatch: pytest.MonkeyPatch): + monkeypatch.delenv("DASHSCOPE_API_BASE", raising=False) + + assert resolve_dashscope_openai_api_base("https://gateway.example.com/v1") == "https://gateway.example.com/v1" + + +def test_openai_base_defaults_to_dashscope(monkeypatch: pytest.MonkeyPatch): + monkeypatch.delenv("DASHSCOPE_API_BASE", raising=False) + + assert resolve_dashscope_openai_api_base(None) == f"{DASHSCOPE_DEFAULT_ORIGIN}/compatible-mode/v1" + + +def test_anthropic_base_derives_from_the_domain(monkeypatch: pytest.MonkeyPatch): + monkeypatch.delenv("DASHSCOPE_API_BASE", raising=False) + + assert resolve_dashscope_anthropic_api_base(WORKSPACE) == f"{WORKSPACE}/apps/anthropic" + assert resolve_dashscope_anthropic_api_base(f"{WORKSPACE}/compatible-mode/v1") == f"{WORKSPACE}/apps/anthropic" + + +def test_anthropic_base_keeps_unrecognized_hosts(monkeypatch: pytest.MonkeyPatch): + monkeypatch.delenv("DASHSCOPE_API_BASE", raising=False) + + assert ( + resolve_dashscope_anthropic_api_base("https://gateway.example.com/anthropic") + == "https://gateway.example.com/anthropic" + ) + + +def test_rerank_base_derives_from_the_domain(monkeypatch: pytest.MonkeyPatch): + monkeypatch.delenv("DASHSCOPE_API_BASE_RERANK", raising=False) + + assert ( + resolve_dashscope_family_rerank_api_base(WORKSPACE, "DASHSCOPE_API_BASE_RERANK", "https://default/reranks") + == f"{WORKSPACE}/compatible-api/v1/reranks" + ) + + +def test_rerank_base_keeps_unrecognized_hosts(): + gateway: Final = "https://gateway.example.com/v1" + + assert ( + resolve_dashscope_family_rerank_api_base(gateway, "DASHSCOPE_API_BASE_RERANK", "https://default/reranks") + == gateway + ) + + +def test_rerank_base_uses_the_configured_default(monkeypatch: pytest.MonkeyPatch): + monkeypatch.delenv("DASHSCOPE_API_BASE_RERANK", raising=False) + + assert ( + resolve_dashscope_family_rerank_api_base(None, "DASHSCOPE_API_BASE_RERANK", "https://default/reranks") + == "https://default/reranks" + ) + + +@pytest.mark.parametrize( + "origin", + ["https://cn-hongkong.dashscope.aliyuncs.com", "https://token-plan.cn-beijing.maas.aliyuncs.com"], +) +def test_rerank_keeps_remapping_chat_bases_on_other_aliyun_hosts(monkeypatch: pytest.MonkeyPatch, origin: str): + monkeypatch.delenv("DASHSCOPE_API_BASE_RERANK", raising=False) + + assert ( + resolve_dashscope_family_rerank_api_base( + f"{origin}/compatible-mode/v1", "DASHSCOPE_API_BASE_RERANK", "https://default/reranks" + ) + == f"{origin}/compatible-api/v1/reranks" + ) diff --git a/tests/unit/llms/dashscope/test_dashscope_embedding_transformation.py b/tests/unit/llms/dashscope/test_dashscope_embedding_transformation.py index 1b6eea0e4c8..27afbee3173 100644 --- a/tests/unit/llms/dashscope/test_dashscope_embedding_transformation.py +++ b/tests/unit/llms/dashscope/test_dashscope_embedding_transformation.py @@ -3,6 +3,7 @@ Unit tests for DashScope embedding transformation. """ import json +from typing import Final from unittest.mock import MagicMock import httpx @@ -16,6 +17,40 @@ from litellm.llms.dashscope.embed.transformation import ( ) from litellm.types.utils import EmbeddingResponse +WORKSPACE: Final = "https://llm-abc123.cn-beijing.maas.aliyuncs.com" + + +def test_get_complete_url_derives_the_compatible_base_from_any_standard_base(): + config: Final = DashScopeEmbeddingConfig() + + for api_base in ( + WORKSPACE, + f"{WORKSPACE}/compatible-mode/v1", + f"{WORKSPACE}/apps/anthropic", + f"{WORKSPACE}/api/v1", + ): + url: Final = config.get_complete_url( + api_base=api_base, + api_key="sk-test", + model="text-embedding-v4", + optional_params={}, + litellm_params={}, + ) + + assert url == f"{WORKSPACE}/compatible-mode/v1/embeddings" + + +def test_get_complete_url_keeps_unrecognized_hosts(): + url: Final = DashScopeEmbeddingConfig().get_complete_url( + api_base="https://gateway.example.com/v1", + api_key="sk-test", + model="text-embedding-v4", + optional_params={}, + litellm_params={}, + ) + + assert url == "https://gateway.example.com/v1/embeddings" + def test_validate_environment_and_url(): config = DashScopeEmbeddingConfig() diff --git a/tests/unit/llms/dashscope/test_dashscope_messages_transformation.py b/tests/unit/llms/dashscope/test_dashscope_messages_transformation.py new file mode 100644 index 00000000000..decc12afd87 --- /dev/null +++ b/tests/unit/llms/dashscope/test_dashscope_messages_transformation.py @@ -0,0 +1,215 @@ +""" +Unit tests for DashScope's Anthropic-compatible Messages endpoint. +""" + +from typing import Final + +import pytest +import respx + +import litellm +from litellm.exceptions import AuthenticationError +from litellm.llms.dashscope.messages.transformation import ( + DashScopeAnthropicMessagesConfig, +) + +WORKSPACE: Final = "https://llm-abc123.cn-beijing.maas.aliyuncs.com" +INTL: Final = "https://dashscope-intl.aliyuncs.com" +CODING_PLAN: Final = "https://coding-intl.dashscope.aliyuncs.com" +CHAT_COMPLETION: Final = { + "id": "chatcmpl-1", + "object": "chat.completion", + "created": 1792227137, + "model": "qwen3.7-plus", + "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}], + "usage": {"prompt_tokens": 9, "completion_tokens": 1, "total_tokens": 10}, +} +MESSAGE: Final = { + "id": "msg_1", + "type": "message", + "role": "assistant", + "model": "qwen3.7-plus", + "content": [{"type": "text", "text": "ok"}], + "stop_reason": "end_turn", + "stop_sequence": None, + "usage": {"input_tokens": 9, "output_tokens": 1}, +} + + +@pytest.mark.parametrize( + ("api_base", "expected"), + [ + ( + None, + "https://dashscope.aliyuncs.com/apps/anthropic/v1/messages", + ), + ( + "https://dashscope.aliyuncs.com", + "https://dashscope.aliyuncs.com/apps/anthropic/v1/messages", + ), + ( + "https://dashscope.aliyuncs.com/compatible-mode/v1", + "https://dashscope.aliyuncs.com/apps/anthropic/v1/messages", + ), + ( + "https://dashscope.aliyuncs.com/apps/anthropic", + "https://dashscope.aliyuncs.com/apps/anthropic/v1/messages", + ), + ( + "https://dashscope.aliyuncs.com/apps/anthropic/v1/messages", + "https://dashscope.aliyuncs.com/apps/anthropic/v1/messages", + ), + ( + WORKSPACE, + f"{WORKSPACE}/apps/anthropic/v1/messages", + ), + ( + f"{WORKSPACE}/compatible-mode/v1", + f"{WORKSPACE}/apps/anthropic/v1/messages", + ), + ( + "https://gateway.example.com/anthropic", + "https://gateway.example.com/anthropic/v1/messages", + ), + ], +) +def test_get_complete_url(api_base: str | None, expected: str): + config: Final = DashScopeAnthropicMessagesConfig() + + url: Final = config.get_complete_url( + api_base=api_base, + api_key="sk-test", + model="qwen3.7-plus", + optional_params={}, + litellm_params={}, + ) + + assert url == expected + + +@pytest.mark.parametrize("dashscope_api_key", [None, " "]) +def test_missing_dashscope_key_does_not_fall_back_to_anthropic( + monkeypatch: pytest.MonkeyPatch, dashscope_api_key: str | None +): + monkeypatch.delenv("DASHSCOPE_API_KEY", raising=False) + if dashscope_api_key is not None: + monkeypatch.setenv("DASHSCOPE_API_KEY", dashscope_api_key) + monkeypatch.delenv("ANTHROPIC_AUTH_TOKEN", raising=False) + monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-ant-secret") + monkeypatch.setattr(litellm, "api_key", None) + headers: dict[str, str] = {} + + with pytest.raises(AuthenticationError) as error: + DashScopeAnthropicMessagesConfig().validate_anthropic_messages_environment( + headers=headers, + model="qwen3.7-plus", + messages=[], + optional_params={}, + litellm_params={}, + api_key=None, + api_base=WORKSPACE, + ) + + assert "DashScope" in str(error.value) + assert "sk-ant-secret" not in str(error.value) + assert {name.lower() for name in headers}.isdisjoint({"x-api-key", "authorization"}) + + +def test_caller_auth_header_is_honored_without_a_dashscope_key(monkeypatch: pytest.MonkeyPatch): + monkeypatch.delenv("DASHSCOPE_API_KEY", raising=False) + monkeypatch.setattr(litellm, "api_key", None) + headers: Final = {"x-api-key": "caller-key"} + + resolved_headers, _ = DashScopeAnthropicMessagesConfig().validate_anthropic_messages_environment( + headers=headers, + model="qwen3.7-plus", + messages=[], + optional_params={}, + litellm_params={}, + api_key=None, + api_base=WORKSPACE, + ) + + assert resolved_headers["x-api-key"] == "caller-key" + + +def test_caller_authorization_header_is_honored_without_a_dashscope_key(monkeypatch: pytest.MonkeyPatch): + monkeypatch.delenv("DASHSCOPE_API_KEY", raising=False) + monkeypatch.setattr(litellm, "api_key", None) + headers: Final = {"Authorization": "Bearer caller-token"} + + resolved_headers, _ = DashScopeAnthropicMessagesConfig().validate_anthropic_messages_environment( + headers=headers, + model="qwen3.7-plus", + messages=[], + optional_params={}, + litellm_params={}, + api_key=None, + api_base=WORKSPACE, + ) + + assert resolved_headers["Authorization"] == "Bearer caller-token" + + +def test_dashscope_key_wins_over_an_anthropic_environment_key(monkeypatch: pytest.MonkeyPatch): + monkeypatch.setenv("DASHSCOPE_API_KEY", "sk-dashscope") + monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-ant-secret") + + resolved_headers, _ = DashScopeAnthropicMessagesConfig().validate_anthropic_messages_environment( + headers={}, + model="qwen3.7-plus", + messages=[], + optional_params={}, + litellm_params={}, + api_key=None, + api_base=WORKSPACE, + ) + + header_values: Final = " ".join(resolved_headers.values()) + assert "sk-dashscope" in header_values + assert "sk-ant-secret" not in header_values + + +def test_dashscope_prefix_reaches_the_provider(): + model, provider, _, _ = litellm.get_llm_provider( + "dashscope/qwen3.7-plus", + api_base=WORKSPACE, + ) + + assert provider == "dashscope" + assert model == "qwen3.7-plus" + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + ("api_base", "expected_url"), + [ + (INTL, f"{INTL}/apps/anthropic/v1/messages"), + (f"{INTL}/apps/anthropic", f"{INTL}/apps/anthropic/v1/messages"), + (WORKSPACE, f"{WORKSPACE}/apps/anthropic/v1/messages"), + (f"{CODING_PLAN}/apps/anthropic", f"{CODING_PLAN}/apps/anthropic/v1/messages"), + (None, "https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions"), + (f"{INTL}/compatible-mode/v1", f"{INTL}/compatible-mode/v1/chat/completions"), + (f"{CODING_PLAN}/v1", f"{CODING_PLAN}/v1/chat/completions"), + ("https://gateway.example.com/anthropic", "https://gateway.example.com/anthropic/chat/completions"), + ], +) +async def test_only_a_domain_root_or_anthropic_base_leaves_the_chat_bridge( + monkeypatch: pytest.MonkeyPatch, api_base: str | None, expected_url: str +): + monkeypatch.delenv("DASHSCOPE_API_BASE", raising=False) + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + + with respx.mock(assert_all_called=False) as upstream: + upstream.route(path__regex=r".*/chat/completions$").respond(200, json=CHAT_COMPLETION) + upstream.route(path__regex=r".*/v1/messages$").respond(200, json=MESSAGE) + + await litellm.anthropic.messages.acreate( + model="dashscope/qwen3.7-plus", + messages=[{"role": "user", "content": "say ok"}], + max_tokens=16, + api_key="sk-dashscope", + api_base=api_base, + ) + + assert [str(call.request.url) for call in upstream.calls] == [expected_url] diff --git a/tests/unit/llms/dashscope/test_dashscope_rerank_transformation.py b/tests/unit/llms/dashscope/test_dashscope_rerank_transformation.py index a913c79af31..f0dbdb74b68 100644 --- a/tests/unit/llms/dashscope/test_dashscope_rerank_transformation.py +++ b/tests/unit/llms/dashscope/test_dashscope_rerank_transformation.py @@ -3,12 +3,15 @@ Unit tests for DashScope rerank transformation. """ import json +from typing import Final from unittest.mock import MagicMock import httpx import pytest +import respx from pydantic import ValidationError +import litellm from litellm.llms.dashscope.common_utils import DashScopeError from litellm.llms.dashscope.rerank.transformation import ( @@ -26,6 +29,21 @@ class TestDashScopeRerankURL: url = self.config.get_complete_url(api_base=None, model="qwen3-rerank") assert url == DEFAULT_RERANK_URL + def test_domain_root_remaps_to_rerank_route(self): + url: Final = self.config.get_complete_url( + api_base="https://llm-abc123.cn-beijing.maas.aliyuncs.com", + model="qwen3-rerank", + ) + assert url == "https://llm-abc123.cn-beijing.maas.aliyuncs.com/compatible-api/v1/reranks" + + def test_explicit_rerank_url_wins_over_the_rerank_env(self, monkeypatch: pytest.MonkeyPatch): + monkeypatch.setenv("DASHSCOPE_API_BASE_RERANK", DEFAULT_RERANK_URL) + explicit_url: Final = "https://dashscope-intl.aliyuncs.com/compatible-api/v1/reranks" + + url: Final = self.config.get_complete_url(api_base=explicit_url, model="qwen3-rerank") + + assert url == explicit_url + def test_chat_shaped_base_remaps_to_rerank_route(self): url = self.config.get_complete_url( api_base="https://dashscope.aliyuncs.com/compatible-mode/v1", @@ -434,3 +452,37 @@ def test_transform_rerank_response_shape_errors_do_not_echo_the_payload(): _transform({"results": [["leaked document text"]]}) assert "leaked document text" not in str(exc_info.value) + + +@pytest.mark.parametrize( + ("provider", "rerank_env"), + [ + ("dashscope", "DASHSCOPE_API_BASE_RERANK"), + ("qwencloud", "QWENCLOUD_API_BASE_RERANK"), + ("qwen_ai_platform", "QWEN_AI_PLATFORM_API_BASE_RERANK"), + ], +) +@pytest.mark.parametrize( + "api_base", + [ + "https://dashscope-intl.aliyuncs.com/compatible-api/v1/reranks", + "https://dashscope-intl.aliyuncs.com/compatible-api/v1", + "https://dashscope-intl.aliyuncs.com", + "https://dashscope-intl.aliyuncs.com/api/v1", + ], +) +def test_rerank_sends_an_explicit_base_over_the_rerank_env( + monkeypatch: pytest.MonkeyPatch, provider: str, rerank_env: str, api_base: str +): + monkeypatch.setenv("DASHSCOPE_API_BASE", "https://dashscope-intl.aliyuncs.com") + monkeypatch.setenv(rerank_env, DEFAULT_RERANK_URL) + + with respx.mock(assert_all_called=True) as upstream: + upstream.post("https://dashscope-intl.aliyuncs.com/compatible-api/v1/reranks").respond( + 200, json={"id": "r1", "results": [{"index": 0, "relevance_score": 0.9}], "usage": {"total_tokens": 5}} + ) + response: Final = litellm.rerank( + model=f"{provider}/qwen3-rerank", query="q", documents=["a"], api_base=api_base, api_key="sk-test" + ) + + assert response.results == [{"index": 0, "relevance_score": 0.9}] diff --git a/tests/unit/llms/dashscope/test_dashscope_video_transformation.py b/tests/unit/llms/dashscope/test_dashscope_video_transformation.py index 15b77aaee3d..c1b4157c8ab 100644 --- a/tests/unit/llms/dashscope/test_dashscope_video_transformation.py +++ b/tests/unit/llms/dashscope/test_dashscope_video_transformation.py @@ -1119,6 +1119,20 @@ class TestDashScopeVideoEnvironment: assert url == "https://ws-123.cn-beijing.maas.aliyuncs.com" + def test_every_protocol_base_normalizes_to_the_host(self): + workspace: Final = "https://ws-123.cn-beijing.maas.aliyuncs.com" + + for api_base in ( + f"{workspace}/compatible-mode/v1", + f"{workspace}/apps/anthropic", + f"{workspace}/api/v1", + ): + url: Final = DashScopeVideoConfig().get_complete_url( + model="wan3.0-video", api_base=api_base, litellm_params={} + ) + + assert url == workspace + def test_default_host_is_beijing(self, monkeypatch): monkeypatch.delenv("DASHSCOPE_API_BASE_VIDEO", raising=False) monkeypatch.delenv("DASHSCOPE_API_BASE", raising=False) diff --git a/tests/unit/llms/dashscope/test_qwen_brand_aliases.py b/tests/unit/llms/dashscope/test_qwen_brand_aliases.py index 69ae66d3818..43ef8f90b39 100644 --- a/tests/unit/llms/dashscope/test_qwen_brand_aliases.py +++ b/tests/unit/llms/dashscope/test_qwen_brand_aliases.py @@ -2,8 +2,10 @@ import json import math import re from pathlib import Path +from typing import Final, Literal import pytest +from typing_extensions import ReadOnly, TypedDict import litellm from litellm import completion, get_llm_provider @@ -53,6 +55,23 @@ DASHSCOPE_FAMILY_ENV_VARS = [ "QWEN_AI_PLATFORM_API_BASE_IMAGE", ] + + +class BrandCase(TypedDict): + provider: ReadOnly[str] + display_name: ReadOnly[str] + enum: ReadOnly[LlmProviders] + key_env: ReadOnly[str] + base_env: ReadOnly[str] + default_base: ReadOnly[str] + default_rerank_base: ReadOnly[str] + default_image_base: ReadOnly[str] + chat_config: ReadOnly[type[DashScopeChatConfig]] + embedding_config: ReadOnly[type[DashScopeEmbeddingConfig]] + rerank_config: ReadOnly[type[DashScopeRerankConfig]] + image_config: ReadOnly[type[DashScopeImageGenerationConfig]] + + BRAND_CASES = [ pytest.param( { @@ -211,6 +230,20 @@ class TestQwenBrandDefaultUrls: ) assert url == f"{brand['default_base']}/embeddings" + @pytest.mark.parametrize("brand", BRAND_CASES) + @pytest.mark.parametrize(("config_key", "route"), [("chat_config", "chat/completions"), ("embedding_config", "embeddings")]) + def test_openai_routes_derive_from_a_domain_root( + self, brand: BrandCase, config_key: Literal["chat_config", "embedding_config"], route: str + ) -> None: + url: Final = brand[config_key]().get_complete_url( + api_base="https://dashscope-intl.aliyuncs.com", + api_key="sk-test", + model="qwen-max", + optional_params={}, + litellm_params={}, + ) + assert url == f"https://dashscope-intl.aliyuncs.com/compatible-mode/v1/{route}" + @pytest.mark.parametrize("brand", BRAND_CASES) def test_rerank_complete_url(self, brand): url = brand["rerank_config"]().get_complete_url(api_base=None, model="gte-rerank-v2") @@ -249,6 +282,19 @@ class TestQwenBrandDefaultUrls: ) assert url == brand["default_image_base"] + @pytest.mark.parametrize("brand", BRAND_CASES) + def test_image_generation_uses_its_own_configured_base(self, monkeypatch: pytest.MonkeyPatch, brand: BrandCase) -> None: + monkeypatch.setenv("DASHSCOPE_API_BASE", "https://legacy.example.com/images") + monkeypatch.setenv(brand["base_env"], "https://gateway.example.com/qwen/images") + url: Final = brand["image_config"]().get_complete_url( + api_base=None, + api_key="sk-test", + model="qwen-image", + optional_params={}, + litellm_params={}, + ) + assert url == "https://gateway.example.com/qwen/images" + @pytest.mark.parametrize("brand", BRAND_CASES) def test_validate_environment_requires_key(self, brand): with pytest.raises(ValueError, match="DASHSCOPE_API_KEY"): diff --git a/tests/unit/test_dashscope_image_generation.py b/tests/unit/test_dashscope_image_generation.py index 960807a033b..9c65399350c 100644 --- a/tests/unit/test_dashscope_image_generation.py +++ b/tests/unit/test_dashscope_image_generation.py @@ -5,21 +5,26 @@ qwen-image-3.0, qwen-image-3.0-pro). Run in docker: pytest tests/unit/test_dashscope_image_generation.py -v """ +from typing import Final from unittest.mock import MagicMock, patch import httpx import pytest +import respx from pydantic import ValidationError import litellm +from litellm.llms.base_llm.chat.transformation import BaseLLMException +from litellm.llms.dashscope.common_utils import ( + DASHSCOPE_DEFAULT_ORIGIN, + DASHSCOPE_IMAGE_PATH, +) from litellm.llms.dashscope.image_generation.transformation import ( - DashScopeImageGenerationConfig, DEFAULT_API_BASE, + DashScopeImageGenerationConfig, ) from litellm.types.utils import ImageResponse from litellm.utils import get_llm_provider -from litellm.llms.base_llm.chat.transformation import BaseLLMException - # --------------------------------------------------------------------------- # 1. Provider detection @@ -77,6 +82,47 @@ class TestDashScopeImageGenerationConfig: url = self.cfg.get_complete_url(chat_api_base, None, "qwen-image-3.0", {}, {}) assert url == DEFAULT_API_BASE + @pytest.mark.parametrize( + ("api_base", "expected"), + [ + ( + "https://dashscope.aliyuncs.com", + "https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation", + ), + ( + "https://llm-abc123.cn-beijing.maas.aliyuncs.com", + "https://llm-abc123.cn-beijing.maas.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation", + ), + ( + "https://llm-abc123.cn-beijing.maas.aliyuncs.com/apps/anthropic", + "https://llm-abc123.cn-beijing.maas.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation", + ), + ( + "https://dashscope-us.aliyuncs.com/api/v1/", + "https://dashscope-us.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation", + ), + ], + ) + def test_get_complete_url_derives_from_the_domain( + self, api_base: str, expected: str + ): + url: Final = self.cfg.get_complete_url(api_base, None, "qwen-image-3.0", {}, {}) + assert url == expected + + @pytest.mark.parametrize( + "endpoint_url", + [ + "https://dashscope.aliyuncs.com/api/v1/services/aigc/image-generation/generation", + "https://dashscope-intl.aliyuncs.com/api/v1/services/aigc/text2image/image-synthesis", + ], + ) + def test_get_complete_url_keeps_a_full_dashscope_endpoint_url( + self, monkeypatch: pytest.MonkeyPatch, endpoint_url: str + ): + monkeypatch.setenv("DASHSCOPE_API_BASE_IMAGE", "https://images.example.com/generate") + url: Final = self.cfg.get_complete_url(endpoint_url, None, "qwen-image-3.0", {}, {}) + assert url == endpoint_url + def test_validate_environment_sets_auth_header(self): headers = self.cfg.validate_environment( headers={}, @@ -458,6 +504,81 @@ def test_litellm_image_generation_dashscope_end_to_end(model: str): assert body["parameters"]["size"] == "1024*1024" + +@pytest.mark.parametrize( + ("api_base", "chat_base_env", "image_base_env", "expected_url"), + [ + (None, "", "", DEFAULT_API_BASE), + ("https://dashscope.aliyuncs.com", "", "", f"{DASHSCOPE_DEFAULT_ORIGIN}{DASHSCOPE_IMAGE_PATH}"), + (None, "", "https://images.example.com/generate", "https://images.example.com/generate"), + ( + "https://dashscope.aliyuncs.com/compatible-mode/v1", + "", + "https://images.example.com/generate", + "https://images.example.com/generate", + ), + (None, "https://gateway.example.com/dashscope/images", "", "https://gateway.example.com/dashscope/images"), + (None, "https://dashscope-us.aliyuncs.com", "https://images.example.com/generate", "https://images.example.com/generate"), + ( + None, + f"https://dashscope-us.aliyuncs.com{DASHSCOPE_IMAGE_PATH}", + "https://images.example.com/generate", + f"https://dashscope-us.aliyuncs.com{DASHSCOPE_IMAGE_PATH}", + ), + ( + None, + "https://gateway.example.com/dashscope/images", + "https://images.example.com/generate", + "https://gateway.example.com/dashscope/images", + ), + ( + None, + "https://dashscope-us.aliyuncs.com/compatible-mode/v1", + "https://images.example.com/generate", + "https://images.example.com/generate", + ), + ( + "https://dashscope-us.aliyuncs.com", + "", + "https://images.example.com/generate", + f"https://dashscope-us.aliyuncs.com{DASHSCOPE_IMAGE_PATH}", + ), + ( + None, + "https://dashscope-us.aliyuncs.com", + "", + f"https://dashscope-us.aliyuncs.com{DASHSCOPE_IMAGE_PATH}", + ), + ( + "https://dashscope-intl.aliyuncs.com", + "https://gateway.example.com/dashscope/images", + "", + f"https://dashscope-intl.aliyuncs.com{DASHSCOPE_IMAGE_PATH}", + ), + ], +) +def test_image_generation_resolves_from_the_configured_base_not_the_chat_default( + monkeypatch: pytest.MonkeyPatch, + api_base: str | None, + chat_base_env: str, + image_base_env: str, + expected_url: str, +): + monkeypatch.setenv("DASHSCOPE_API_BASE", chat_base_env) + monkeypatch.setenv("DASHSCOPE_API_BASE_IMAGE", image_base_env) + image_url: Final = "https://dashscope-result.oss.aliyuncs.com/t.png" + + with respx.mock(assert_all_called=True) as upstream: + upstream.post(expected_url).respond( + 200, json={"output": {"choices": [{"message": {"content": [{"image": image_url}]}}]}} + ) + response: Final = litellm.image_generation( + model="dashscope/qwen-image-3.0", prompt="a puppy", api_key="sk-test-key", api_base=api_base + ) + + assert response.data[0].url == image_url + + def _transform_response(payload: object) -> ImageResponse: return DashScopeImageGenerationConfig().transform_image_generation_response( model="qwen-image-2.0", From c45b26402428804cdbd7504e1ea244df5c4d328e Mon Sep 17 00:00:00 2001 From: "devin-ai-integration[bot]" <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Thu, 29 Oct 2026 09:32:18 -0700 Subject: [PATCH 5/7] fix(proxy): persist tag spend and drain success callbacks within the shutdown budget (#48988) * fix: drain shutdown callbacks and persist pending tag spend * fix: catch shutdown timeouts on Python 3.10 * fix: share shutdown callback budget with HTTP draining * fix: type shutdown cancellation tasks and document its clock * test: retain callbacks after bounded worker cancellation * test(proxy): inspect the lifespan body after the SIGTERM wrapper split Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(utils): keep the success callback enqueue on its own task The synchronous enqueue spawned the logging worker loop inside the awaited call and ran callbacks that per-call loops used to drop, which hung the SDK callback tests on the autouse worker flush. The shutdown worker flush covers the queued callback on its own, so the one-tick task is restored Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(logging_worker): keep the queue's unfinished count exact when a full queue rejects a callback The aggressive clear ran the rejected callback and marked it done against the queue it was never put in, so flush() could return while a real callback was still running and the shutdown drain cancelled it early. The rejected callback is now re-enqueued once the clear made room, and only runs uncounted when the queue is still full Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * refactor(logging_worker): build the aggressive-clear run set immutably Final tuple instead of an appended list, with the re-enqueue in a _requeue helper, and Final annotations on the overflow regression test Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(logging): track overflow callbacks through shutdown * chore(logging): explain shared callback ownership * test(logging): isolate deferred callback completion --------- Co-authored-by: Joshua Valluru <326636767+joshua-berri@users.noreply.github.com> Co-authored-by: joshua Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- helm/litellm-helm/README.md | 6 + litellm/litellm_core_utils/logging_worker.py | 125 +++++---- litellm/proxy/db/db_spend_update_writer.py | 2 - litellm/proxy/proxy_server.py | 40 ++- .../shutdown/graceful_shutdown_manager.py | 58 ++++- litellm/responses/streaming_iterator.py | 2 +- litellm/utils.py | 8 +- .../integration/spend/test_shutdown_flush.py | 224 +++++++++++++++- .../litellm_core_utils/test_logging_worker.py | 246 ++++++++++++++++++ .../test_deferred_guardrail_logging.py | 103 ++------ .../unit/proxy/proxy_server/test_lifecycle.py | 5 +- .../test_graceful_shutdown_manager.py | 104 ++++++++ tests/unit/proxy/test_proxy_server.py | 199 ++++++++++++++ tests/unit/test_utils.py | 32 +++ 14 files changed, 1011 insertions(+), 143 deletions(-) diff --git a/helm/litellm-helm/README.md b/helm/litellm-helm/README.md index e059d45e037..a3488d6114c 100644 --- a/helm/litellm-helm/README.md +++ b/helm/litellm-helm/README.md @@ -259,3 +259,9 @@ itself. `lensWorker.mode` selects `disabled`, `bundled` or `external`. Bundled mode generates separate service and identity-signing credentials. External mode requires `lensWorker.gateway.secretName` and `lensWorker.serviceTokenSecret.name`; their keys default to `gateway-secret` and `service-token`. Provision distinct values matching the external Lens deployment Follow the [Lens Helm connection guide](https://github.com/BerriAI/lens/blob/main/helm/lens/README.md#connect-a-gateway) for complete values, verification, GitOps credential requirements and existing-data migration. Source-chart installation requires a built Lens image until the first signed Lens release is published + +## Shutdown accounting + +`GRACEFUL_SHUTDOWN_TIMEOUT` defaults to 30 seconds. Pending success callbacks share the time remaining since SIGTERM or the first `/health/drain` call. Callback flushing reserves the existing 10-second teardown headroom for cancellation, tracked accounting tasks and persistence. Exhausting this budget can discard callbacks, which is logged + +Set `terminationGracePeriodSeconds` to at least `GRACEFUL_SHUTDOWN_TIMEOUT + 10`, allowing additional time for preStop hooks and database writes. This is not a hard shutdown bound or a guarantee of persistence. Database retries retain their own limits, and cancellation-resistant callbacks can remain alive. The existing preStop hook can also be followed by Uvicorn's separate connection drain, as described in `values.yaml` diff --git a/litellm/litellm_core_utils/logging_worker.py b/litellm/litellm_core_utils/logging_worker.py index ff649953e69..8f217b1ea5e 100644 --- a/litellm/litellm_core_utils/logging_worker.py +++ b/litellm/litellm_core_utils/logging_worker.py @@ -6,7 +6,7 @@ import atexit import contextvars import inspect import logging -from collections.abc import Coroutine, Iterator +from collections.abc import Awaitable, Callable, Coroutine, Iterator from typing import Final from typing_extensions import TypedDict @@ -66,14 +66,19 @@ class LoggingWorker: max_queue_size: int = LOGGING_WORKER_MAX_QUEUE_SIZE, concurrency: int = LOGGING_WORKER_CONCURRENCY, timeout_summary_window: float = LOGGING_WORKER_TIMEOUT_SUMMARY_WINDOW_SECONDS, + *, + retry_sleep: Callable[[float], Awaitable[None]] = asyncio.sleep, ): self.timeout = timeout self.max_queue_size = max_queue_size self.concurrency = concurrency self.timeout_summary_window = timeout_summary_window self._queue: asyncio.Queue[LoggingTask] | None = None - self._worker_task: asyncio.Task | None = None - self._running_tasks: set[asyncio.Task] = set() + self._worker_task: asyncio.Task[None] | None = None + self._running_tasks: set[asyncio.Task[None]] = set() + self._overflow_tasks: set[asyncio.Task[None]] = set() # mutable-ok: done callbacks discard refs + self._pending_enqueues: dict[int, LoggingTask] = {} # mutable-ok: retries and shutdown share callback ownership + self._retry_sleep = retry_sleep self._dequeued_tasks: dict[int, LoggingTask] = {} # mutable-ok: refs so flush can rescue never-started tasks self._sem: asyncio.Semaphore | None = None self._bound_loop: asyncio.AbstractEventLoop | None = None @@ -168,6 +173,7 @@ class LoggingWorker: self._sem = None self._worker_task = None self._running_tasks.clear() + self._overflow_tasks.clear() self._timeout_summary_task = None self._timeout_burst_count = 0 self._queue = new_queue @@ -185,6 +191,10 @@ class LoggingWorker: self._sem = asyncio.Semaphore(self.concurrency) if self._worker_task is None or self._worker_task.done(): self._worker_task = asyncio.create_task(self._worker_loop()) + if not self._overflow_tasks: + self._aggressive_clear_in_progress = False + for pending in tuple(self._pending_enqueues.values()): + self._schedule_delayed_enqueue_retry(pending) async def _process_log_task( self, task: LoggingTask, sem: asyncio.Semaphore, queue: "asyncio.Queue[LoggingTask]" @@ -278,12 +288,24 @@ class LoggingWorker: # Capture the current context when enqueueing task: Final = LoggingTask(coroutine=coroutine, context=contextvars.copy_context()) + self._enqueue_task(task) + + def _enqueue_task(self, task: LoggingTask) -> None: + if self._queue is None: + return try: self._queue.put_nowait(task) except asyncio.QueueFull: - # Queue is full - handle it appropriately verbose_logger.exception("LoggingWorker queue is full") + self._pending_enqueues[id(task)] = task self._handle_queue_full(task) + else: + self._pending_enqueues.pop(id(task), None) + + def _start_overflow_task(self, coroutine: Coroutine[object, object, None]) -> None: + task: Final = asyncio.create_task(coroutine) + self._overflow_tasks.add(task) + task.add_done_callback(self._overflow_tasks.discard) def _should_start_aggressive_clear(self) -> bool: """ @@ -326,7 +348,7 @@ class LoggingWorker: if self._should_start_aggressive_clear(): self._mark_aggressive_clear_started() # Schedule clearing as async task so enqueue returns immediately (non-blocking) - asyncio.create_task(self._aggressively_clear_queue_async(task)) + self._start_overflow_task(self._aggressively_clear_queue_async(task)) else: # Cooldown active or clear in progress, schedule a delayed retry self._schedule_delayed_enqueue_retry(task) @@ -363,7 +385,7 @@ class LoggingWorker: delay: Final = self._calculate_retry_delay() # Schedule the retry as a background task - asyncio.create_task(self._retry_enqueue_task(task, delay)) + self._start_overflow_task(self._retry_enqueue_task(task, delay)) except RuntimeError: # No event loop, drop the task as we can't schedule a retry pass @@ -373,17 +395,8 @@ class LoggingWorker: Retry enqueueing the task after delay, preserving original context. This is called as a background task from _schedule_delayed_enqueue_retry. """ - await asyncio.sleep(delay) - - # Try to enqueue the task directly, preserving its original context - if self._queue is None: - return - - try: - self._queue.put_nowait(task) - except asyncio.QueueFull: - # Still full - handle it appropriately (clear or retry again) - self._handle_queue_full(task) + await self._retry_sleep(delay) + self._enqueue_task(task) def _extract_tasks_from_queue(self) -> list[LoggingTask]: """ @@ -423,15 +436,10 @@ class LoggingWorker: return extracted_tasks: Final = self._extract_tasks_from_queue() - - # Add new task to extracted tasks to process directly if new_task is not None: - self._track_dequeued(new_task) - extracted_tasks.append(new_task) - - # Process extracted tasks directly + self._enqueue_task(new_task) if extracted_tasks: - await self._process_extracted_tasks(extracted_tasks) + await asyncio.gather(*(self._process_single_task(task) for task in extracted_tasks)) except Exception as e: verbose_logger.exception("LoggingWorker error during aggressive clear: %s", e) finally: @@ -439,7 +447,7 @@ class LoggingWorker: self._aggressive_clear_in_progress = False async def _process_single_task(self, task: LoggingTask) -> None: - """Process a single task and mark it done.""" + """Process a task extracted from the queue and mark it done.""" if self._queue is None: return @@ -455,17 +463,6 @@ class LoggingWorker: self._untrack_dequeued(task) self._queue.task_done() - async def _process_extracted_tasks(self, tasks: list[LoggingTask]) -> None: - """ - Process tasks that were extracted from the queue to make room. - Processes them concurrently without semaphore limits for maximum speed. - """ - if not tasks or self._queue is None: - return - - # Process all tasks concurrently for maximum speed - await asyncio.gather(*[self._process_single_task(task) for task in tasks]) - def ensure_initialized_and_enqueue(self, async_coroutine: Coroutine[object, None, None]) -> None: """ Ensure the logging worker is initialized and enqueue the coroutine. @@ -473,32 +470,63 @@ class LoggingWorker: self.start() self.enqueue(async_coroutine) - async def stop(self) -> None: + @property + def pending_count(self) -> int: + return ( + (self._queue.qsize() if self._queue is not None else 0) + + len(self._dequeued_tasks) + + len(self._pending_enqueues) + ) + + async def stop(self, *, discard_pending: bool = False, timeout: float | None = None) -> None: """Stop the logging worker and clean up resources.""" + if discard_pending and self._queue is not None: + discarded: Final = self._drain_pending(self._queue) + for task in discarded: + task["coroutine"].close() + self._queue.task_done() + if discarded: + verbose_logger.warning("LoggingWorker shutdown: %d queued callbacks discarded", len(discarded)) if self._timeout_summary_task is not None: self._timeout_summary_task.cancel() self._timeout_summary_task = None self._emit_timeout_summary() - if self._worker_task is None and not self._running_tasks: - # No worker launched and no in-flight tasks to drain. - return + if discard_pending: + rejected: Final = tuple(self._pending_enqueues.values()) + self._pending_enqueues.clear() + for task in rejected: + task["coroutine"].close() + if rejected: + verbose_logger.warning("LoggingWorker shutdown: %d overflow callbacks discarded", len(rejected)) - tasks_to_cancel: Final[list[asyncio.Task]] = list(self._running_tasks) - if self._worker_task: - # Include the main worker loop so it stops fetching work. - tasks_to_cancel.append(self._worker_task) + tasks_to_cancel: Final = ( + *self._running_tasks, + *self._overflow_tasks, + *((self._worker_task,) if self._worker_task is not None else ()), + ) + if not tasks_to_cancel: + return for task in tasks_to_cancel: # Propagate cancellation to every pending task. task.cancel() # Wait for cancellation to settle; ignore errors raised during shutdown. - await asyncio.gather(*tasks_to_cancel, return_exceptions=True) + done, pending = await asyncio.wait(tasks_to_cancel, timeout=timeout) + await asyncio.gather(*done, return_exceptions=True) + if pending: + verbose_logger.warning("LoggingWorker shutdown: %d tasks still alive after cancellation", len(pending)) - self._worker_task = None + if discard_pending and self._queue is not None: + for task in self._unstarted_dequeued_tasks(): + task["coroutine"].close() + self._untrack_dequeued(task) + self._queue.task_done() + if self._worker_task is not None and self._worker_task.done(): + self._worker_task = None # Drop references to completed tasks so we can restart cleanly. - self._running_tasks.clear() + self._running_tasks.difference_update(done) async def flush(self) -> None: """Flush the logging queue. @@ -516,6 +544,9 @@ class LoggingWorker: return self.start() await self._queue.join() + while self._overflow_tasks: + await asyncio.wait(tuple(self._overflow_tasks)) + await self._queue.join() async def clear_queue(self): """ diff --git a/litellm/proxy/db/db_spend_update_writer.py b/litellm/proxy/db/db_spend_update_writer.py index 8ac2597c99b..16dbe2b4131 100644 --- a/litellm/proxy/db/db_spend_update_writer.py +++ b/litellm/proxy/db/db_spend_update_writer.py @@ -1900,8 +1900,6 @@ class DBSpendUpdateWriter: proxy_logging_obj=proxy_logging_obj, ) - # NOTE: Daily tag spend is committed by a separate scheduler job. - ################## Daily End-User Spend Update Transactions ################## # Aggregate all in memory daily end-user spend transactions and commit to db await self._flush_daily_spend_queue( diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index a75ef593b7a..65de7a3c472 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -79,6 +79,7 @@ from litellm.constants import ( DEFAULT_SHARED_HEALTH_CHECK_LOCK_TTL, DEFAULT_SHARED_HEALTH_CHECK_TTL, DEFAULT_SLACK_ALERTING_THRESHOLD, + GRACEFUL_SHUTDOWN_TEARDOWN_HEADROOM_SECONDS, INTERNAL_CALLBACK_NAMES, LANGFUSE_SHUTDOWN_FLUSH_TIMEOUT_MILLIS, LITELLM_EMBEDDING_PROVIDERS_SUPPORTING_INPUT_ARRAY_OF_TOKENS, @@ -98,6 +99,7 @@ from litellm.litellm_core_utils.litellm_logging import ( _init_custom_logger_compatible_class, initialize_sentry, ) +from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER, LoggingWorker from litellm.litellm_core_utils.safe_json_dumps import safe_dumps from litellm.litellm_core_utils.safe_json_loads import safe_json_loads from litellm.litellm_core_utils.token_counter import ( @@ -1274,15 +1276,44 @@ async def _flush_spend_logs_queue_on_shutdown() -> None: async def _flush_spend_on_shutdown() -> None: + from litellm.proxy.utils import update_daily_tag_spend + await flush_spend_counters_on_shutdown() + if prisma_client is not None: + await update_daily_tag_spend(prisma_client=prisma_client, proxy_logging_obj=proxy_logging_obj) await _flush_spend_logs_queue_on_shutdown() async def _flush_spend_around_logging_drain_on_shutdown( flush_spend: Callable[[], Awaitable[None]] = _flush_spend_on_shutdown, + *, + worker: LoggingWorker = GLOBAL_LOGGING_WORKER, + clock: Callable[[], float] = time.monotonic, + drain_tasks: Callable[[float], Awaitable[None]] = drain_logging_tasks, ) -> None: await flush_spend() - await drain_logging_tasks() + remaining: Final = GracefulShutdownManager.remaining(clock=clock) + callback_timeout: Final = max(0.0, remaining - GRACEFUL_SHUTDOWN_TEARDOWN_HEADROOM_SECONDS) + verbose_proxy_logger.info( + "Draining success callbacks before the final spend flush remaining_s=%.3f callback_timeout_s=%.3f", + remaining, + callback_timeout, + ) + + async def drain_callbacks() -> None: + await drain_tasks(callback_timeout) + await worker.flush() + + try: + await asyncio.wait_for(drain_callbacks(), timeout=callback_timeout) + except asyncio.TimeoutError: + verbose_proxy_logger.warning( + "Success callback shutdown budget exhausted with %d unfinished callbacks; accounting may be incomplete", + worker.pending_count, + ) + finally: + await worker.stop(discard_pending=True, timeout=GracefulShutdownManager.remaining(2.0, clock=clock)) + await drain_tasks(GracefulShutdownManager.remaining(5.0, clock=clock)) await flush_spend() @@ -1445,6 +1476,13 @@ async def _connect_to_count_stored_values() -> SupportsRawQueries: @asynccontextmanager async def proxy_startup_event(app: FastAPI) -> AsyncGenerator[ProxyLifespanState, None]: + with GracefulShutdownManager.capture_sigterm(): + async with _proxy_startup_event(app) as state: + yield state + + +@asynccontextmanager +async def _proxy_startup_event(app: FastAPI) -> AsyncGenerator[ProxyLifespanState, None]: global \ prisma_client, \ master_key, \ diff --git a/litellm/proxy/shutdown/graceful_shutdown_manager.py b/litellm/proxy/shutdown/graceful_shutdown_manager.py index d7a07ec34e5..9aeee0453ff 100644 --- a/litellm/proxy/shutdown/graceful_shutdown_manager.py +++ b/litellm/proxy/shutdown/graceful_shutdown_manager.py @@ -5,7 +5,8 @@ Kubernetes terminates a pod by sending ``SIGTERM`` and, after ``terminationGracePeriodSeconds``, ``SIGKILL``. On ``SIGTERM`` uvicorn stops accepting connections and waits up to ``GRACEFUL_SHUTDOWN_TIMEOUT`` for open ones, cancelling whatever is still running, before the lifespan teardown that -runs this drain. +runs this drain. The lifespan registers a SIGTERM wrapper before serving +requests so callback draining can subtract the preceding connection drain. A fixed ``preStop`` sleep can not solve this: it has to be sized for the *worst-case* request, so it either wastes time on every routine shutdown or is @@ -20,8 +21,12 @@ granularity of ``InFlightRequestsMiddleware``. import asyncio import os +import signal +import threading import time -from collections.abc import Callable +from collections.abc import Callable, Generator +from contextlib import contextmanager +from types import FrameType from typing import Final from litellm._logging import verbose_proxy_logger @@ -44,6 +49,7 @@ class GracefulShutdownManager: _is_shutting_down: bool = False _shutdown_started_at: float | None = None _drain_performed: bool = False + _sigterm_handler: Callable[[int, FrameType | None], None] | None = None @classmethod def is_shutting_down(cls) -> bool: @@ -71,7 +77,7 @@ class GracefulShutdownManager: return DEFAULT_GRACEFUL_SHUTDOWN_TIMEOUT @classmethod - def start_shutdown(cls) -> None: + def start_shutdown(cls, clock: Callable[[], float] = time.monotonic) -> None: """ Mark the worker as draining. Idempotent โ€” repeated calls (e.g. SIGTERM followed by a preStop hit on /health/drain) do not reset the clock. @@ -79,12 +85,42 @@ class GracefulShutdownManager: if cls._is_shutting_down: return cls._is_shutting_down = True - cls._shutdown_started_at = time.monotonic() + cls._shutdown_started_at = clock() verbose_proxy_logger.info( "graceful_shutdown_started in_flight_requests=%s", get_in_flight_requests(), ) + @classmethod + def remaining(cls, cap: float | None = None, *, clock: Callable[[], float] = time.monotonic) -> float: + elapsed: Final = 0.0 if cls._shutdown_started_at is None else clock() - cls._shutdown_started_at + remaining: Final = max(0.0, cls.get_timeout() - elapsed) + return remaining if cap is None else min(max(0.0, cap), remaining) + + @classmethod + @contextmanager + def capture_sigterm(cls, clock: Callable[[], float] = time.monotonic) -> Generator[None, None, None]: + if threading.current_thread() is not threading.main_thread(): + yield + return + previous: Final = signal.getsignal(signal.SIGTERM) + if not callable(previous) or previous is cls._sigterm_handler: + yield + return + + def handle_sigterm(signum: int, frame: FrameType | None) -> None: + cls.start_shutdown(clock) + previous(signum, frame) + + cls._sigterm_handler = handle_sigterm + signal.signal(signal.SIGTERM, handle_sigterm) + try: + yield + finally: + if signal.getsignal(signal.SIGTERM) is handle_sigterm: + signal.signal(signal.SIGTERM, previous) + cls._sigterm_handler = None + @classmethod async def wait_for_drain( cls, @@ -93,13 +129,14 @@ class GracefulShutdownManager: count_fn: Callable[[], int] | None = None, poll_interval: float = _DRAIN_POLL_INTERVAL, log_interval: float = _DRAIN_LOG_INTERVAL, + clock: Callable[[], float] = time.monotonic, ) -> int: """ Poll the in-flight request counter until it reaches the drain target or ``timeout`` seconds elapse. Args: - timeout: Max seconds to wait. Defaults to ``get_timeout()``. + timeout: Max seconds to wait. Defaults to the remaining shutdown budget. exclude_self: When the caller is itself an in-flight HTTP request (the /health/drain endpoint), set this so the caller's own request is not counted as outstanding work. @@ -107,6 +144,7 @@ class GracefulShutdownManager: live ``InFlightRequestsMiddleware`` counter; injectable for tests. poll_interval: Seconds between counter polls. log_interval: Minimum seconds between ``drain_waiting`` log lines. + clock: Monotonic clock shared with the shutdown timestamp. Returns: Number of requests that drained while waiting (>= 0). @@ -120,7 +158,7 @@ class GracefulShutdownManager: cls._drain_performed = True if timeout is None: - timeout = cls.get_timeout() + timeout = cls.remaining(clock=clock) if count_fn is None: count_fn = get_in_flight_requests @@ -128,7 +166,7 @@ class GracefulShutdownManager: # and so counts itself; treat <=1 as "drained" in that case. target: Final = 1 if exclude_self else 0 - start: Final = time.monotonic() + start: Final = clock() initial: Final = count_fn() last_log = start @@ -142,11 +180,11 @@ class GracefulShutdownManager: verbose_proxy_logger.info( "graceful_shutdown_complete drained_requests=%s elapsed_s=%.2f", drained, - time.monotonic() - start, + clock() - start, ) return drained - elapsed = time.monotonic() - start + elapsed = clock() - start if elapsed >= timeout: verbose_proxy_logger.warning( "graceful_shutdown_timeout in_flight_requests=%s elapsed_s=%.2f " @@ -157,7 +195,7 @@ class GracefulShutdownManager: ) return max(0, initial - current) - now = time.monotonic() + now = clock() if now - last_log >= log_interval: verbose_proxy_logger.info( "drain_waiting in_flight_requests=%s elapsed_s=%.2f", diff --git a/litellm/responses/streaming_iterator.py b/litellm/responses/streaming_iterator.py index 8959c32fff2..6a82304b76a 100644 --- a/litellm/responses/streaming_iterator.py +++ b/litellm/responses/streaming_iterator.py @@ -227,7 +227,7 @@ async def drain_logging_tasks(timeout: float = 5.0) -> None: return _, pending = await asyncio.wait(tuple(_PENDING_LOGGING_TASKS), timeout=timeout) if pending: - verbose_logger.warning("Timed out draining %d logging tasks on shutdown, dropping them", len(pending)) + verbose_logger.warning("Timed out draining logging tasks on shutdown: %d tasks still alive", len(pending)) _ERROR_CODE_HTTP_STATUS: Final[Mapping[str, int]] = MappingProxyType( diff --git a/litellm/utils.py b/litellm/utils.py index 4cd2335dd0c..4a0e4cafb22 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -1632,15 +1632,19 @@ def _schedule_async_success_logging( return def _enqueue_async_logging() -> None: + from litellm.responses.streaming_iterator import spawn_logging_task + logging_obj.mark_async_success_log_enqueued() - asyncio.create_task( + spawn_logging_task( + asyncio.get_running_loop(), _client_async_logging_helper( logging_obj=logging_obj, result=result, start_time=start_time, end_time=end_time, is_completion_with_fallbacks=is_completion_with_fallbacks, - ) + ), + task_name="success callback enqueue", ) if not logging_obj.defer_async_logging: diff --git a/tests/integration/spend/test_shutdown_flush.py b/tests/integration/spend/test_shutdown_flush.py index e96511511d9..5c2e9cf98da 100644 --- a/tests/integration/spend/test_shutdown_flush.py +++ b/tests/integration/spend/test_shutdown_flush.py @@ -1,11 +1,14 @@ +import hashlib import json import os +import re import threading +import time import uuid from collections.abc import Callable, Iterator from concurrent.futures import ThreadPoolExecutor from contextlib import contextmanager -from dataclasses import dataclass +from dataclasses import dataclass, replace from pathlib import Path from typing import Final @@ -26,6 +29,225 @@ MODEL_INSERT_ARRIVED_LOG_LINE: Final = "path=/model/new" COMMIT_DELAY_SECONDS: Final = 15 BURST_REQUESTS: Final = 30 +_GATED_LOGGER: Final = """ +import os +import re +import httpx +from litellm.integrations.custom_logger import CustomLogger + +class ShutdownLogger(CustomLogger): + async def async_log_success_event(self, kwargs, response_obj, start_time, end_time): + async with httpx.AsyncClient(timeout=25, trust_env=False) as client: + await client.post(os.environ["SHUTDOWN_CALLBACK_URL"]) + +logger = ShutdownLogger() +""" + + +def _tag_shutdown_config(directory: Path, upstream: str, redis_buffer: bool, *, gated: bool = False) -> Path: + path: Final = directory / "tag-shutdown.yaml" + if gated: + (directory / "shutdown_callbacks.py").write_text(_GATED_LOGGER) + path.write_text( + yaml.safe_dump( + { + "model_list": [ + { + "model_name": "shutdown-model", + "litellm_params": { + "model": "openai/gpt-4o-mini", + "api_base": upstream + "/v1", + "api_key": "synthetic-key", + "input_cost_per_token": 0.01, + "output_cost_per_token": 0.02, + }, + } + ], + "general_settings": { + "master_key": "os.environ/LITELLM_MASTER_KEY", + "database_url": "os.environ/DATABASE_URL", + "proxy_batch_write_at": 3600, + "use_redis_transaction_buffer": redis_buffer, + "disable_model_info_refresh": True, + }, + "litellm_settings": { + "cache": True, + "cache_params": {"type": "redis", "host": "os.environ/REDIS_HOST", "port": "os.environ/REDIS_PORT"}, + "callbacks": ["shutdown_callbacks.logger"] if gated else [], + }, + } + ) + ) + return path + + +def _tagged_chat(gateway: Gateway, key: str, tags: tuple[str, str]) -> None: + response: Final = gateway.request( + "POST", + "/v1/chat/completions", + { + "model": "shutdown-model", + "messages": [{"role": "user", "content": uuid.uuid4().hex}], + "metadata": {"tags": list(tags)}, + }, + key=key, + ) + assert response.status_code == 200, response.text + + +def _assert_shutdown_accounting(key: str, tags: tuple[str, str], requests: int) -> None: + hashed: Final = hashlib.sha256(key.encode()).hexdigest() + for table, count in (("LiteLLM_SpendLogs", "count(*)"), ("LiteLLM_DailyUserSpend", "sum(api_requests)")): + rows: Final = read_rows( + f'SELECT ({count})::int AS requests, sum(spend) AS spend FROM "{table}" WHERE api_key=%s', (hashed,) + ) + assert rows[0]["requests"] == requests, (table, rows) + assert rows[0]["spend"] == pytest.approx(requests * 0.05, abs=1e-9), (table, rows) + key_rows: Final = read_rows('SELECT spend FROM "LiteLLM_VerificationToken" WHERE token=%s', (hashed,)) + assert key_rows[0]["spend"] == pytest.approx(requests * 0.05, abs=1e-9) + for tag in tags: + tag_rows: Final = read_rows( + 'SELECT sum(api_requests)::int AS requests, sum(spend) AS spend FROM "LiteLLM_DailyTagSpend" WHERE api_key=%s AND tag=%s', + (hashed, tag), + ) + assert tag_rows[0]["requests"] == requests, (tag, tag_rows) + assert tag_rows[0]["spend"] == pytest.approx(requests * 0.05, abs=1e-9), (tag, tag_rows) + + +@pytest.mark.parametrize("redis_buffer", [False, True]) +def test_tagged_request_survives_shutdown_before_the_tag_timer( + gateway: Gateway, tmp_path: Path, redis_buffer: bool +) -> None: + tags: Final = (f"shutdown-a-{uuid.uuid4().hex}", f"shutdown-b-{uuid.uuid4().hex}") + overrides: Final = {"DATABASE_URL": os.environ["DATABASE_URL"], "LITELLM_LOG": "DEBUG"} + with wire_server(_provider) as wire: + config: Final = _tag_shutdown_config(tmp_path, wire.url, redis_buffer) + with owned_proxy_process(gateway, tmp_path, overrides, config=config) as owned: + key: Final = string_value( + owned.gateway.post("/key/generate", {"user_id": f"shutdown-{uuid.uuid4().hex}"})["key"] + ) + _tagged_chat(owned.gateway, key, tags) + owned.begin_shutdown() + owned.process.wait(timeout=90) + with owned_proxy_process(gateway, tmp_path, overrides, config=config) as restarted: + _assert_shutdown_accounting(key, tags, 1) + restarted.begin_shutdown() + restarted.process.wait(timeout=90) + _assert_shutdown_accounting(key, tags, 1) + + +@pytest.mark.parametrize("redis_buffer", [False, True]) +@pytest.mark.parametrize("overflow", [False, True]) +def test_callback_backlog_is_accounted_before_shutdown_disconnects( + gateway: Gateway, tmp_path: Path, redis_buffer: bool, overflow: bool +) -> None: + release: Final = threading.Event() + requests: Final = 6 if overflow else 3 + tags: Final = (f"backlog-a-{uuid.uuid4().hex}", f"backlog-b-{uuid.uuid4().hex}") + with ( + wire_server(_provider) as wire, + wire_server(lambda _: Reply(chunks=(b"{}",), gate_before_first=release)) as logger, + ): + config: Final = _tag_shutdown_config(tmp_path, wire.url, redis_buffer, gated=True) + with owned_proxy_process( + gateway, + tmp_path, + { + "DATABASE_URL": os.environ["DATABASE_URL"], + "LOGGING_WORKER_CONCURRENCY": "2", + "LOGGING_WORKER_MAX_QUEUE_SIZE": "2" if overflow else "50000", + "SHUTDOWN_CALLBACK_URL": logger.url, + "GRACEFUL_SHUTDOWN_TIMEOUT": "15", + "LITELLM_LOG": "DEBUG", + }, + config=config, + ) as owned: + key: Final = string_value( + owned.gateway.post("/key/generate", {"user_id": f"backlog-{uuid.uuid4().hex}"})["key"] + ) + try: + for _ in range(requests): + _tagged_chat(owned.gateway, key, tags) + eventually(logger.received.qsize, lambda count: count >= 2, seconds=10) + owned.begin_shutdown() + eventually( + lambda: ( + owned.process.poll() is not None + or "Draining success callbacks before the final spend flush" in owned.log.read_text() + ), + lambda ready: ready, + seconds=30, + ) + release.set() + owned.process.wait(timeout=90) + _assert_shutdown_accounting(key, tags, requests) + finally: + release.set() + + +@pytest.mark.parametrize("gunicorn", [False, True]) +def test_sigterm_http_drain_consumes_callback_budget(gateway: Gateway, tmp_path: Path, gunicorn: bool) -> None: + release_request: Final = threading.Event() + release_callback: Final = threading.Event() + tags: Final = (f"deadline-a-{uuid.uuid4().hex}", f"deadline-b-{uuid.uuid4().hex}") + + def held_provider(request: Request) -> Reply: + response: Final = _provider(request) + return replace(response, chunks=(response.body,), gate_before_first=release_request) + + with ( + wire_server(held_provider) as wire, + wire_server(lambda _: Reply(chunks=(b"{}",), gate_before_first=release_callback)) as logger, + ): + config: Final = _tag_shutdown_config(tmp_path, wire.url, False, gated=True) + with ( + owned_proxy_process( + gateway, + tmp_path, + { + "DATABASE_URL": os.environ["DATABASE_URL"], + "SHUTDOWN_CALLBACK_URL": logger.url, + "GRACEFUL_SHUTDOWN_TIMEOUT": "30", + "LITELLM_LOG": "DEBUG", + }, + config=config, + extra_arguments=("--run_gunicorn",) if gunicorn else (), + ) as owned, + ThreadPoolExecutor(max_workers=1) as executor, + ): + key: Final = string_value( + owned.gateway.post("/key/generate", {"user_id": f"deadline-{uuid.uuid4().hex}"})["key"] + ) + request: Final = executor.submit(_tagged_chat, owned.gateway, key, tags) + try: + eventually(wire.received.qsize, lambda count: count == 1, seconds=10) + owned.begin_shutdown() + eventually( + lambda: "graceful_shutdown_started" in owned.log.read_text(), lambda stamped: stamped, seconds=5 + ) + assert not request.done(), "The shutdown clock must start before the HTTP request drains" + stamped: Final = time.monotonic() + eventually(time.monotonic, lambda now: now - stamped >= 2, seconds=5) + held_seconds: Final = time.monotonic() - stamped + release_request.set() + request.result(timeout=10) + budget: Final = eventually( + lambda: re.search(r"remaining_s=([\d.]+) callback_timeout_s=([\d.]+)", owned.log.read_text()), + lambda match: match is not None, + seconds=15, + ) + assert budget is not None + remaining: Final = float(budget[1]) + allowance: Final = float(budget[2]) + assert remaining <= 30 - held_seconds + 0.05 + assert allowance == pytest.approx(max(0, remaining - 10), abs=0.002) + release_callback.set() + owned.process.wait(timeout=45) + _assert_shutdown_accounting(key, tags, 1) + finally: + release_request.set() + release_callback.set() + def _api_requests(table: str, column: str, identity: str) -> int: rows: Final = read_rows( diff --git a/tests/unit/litellm_core_utils/test_logging_worker.py b/tests/unit/litellm_core_utils/test_logging_worker.py index f48495d1062..fccc11617bb 100644 --- a/tests/unit/litellm_core_utils/test_logging_worker.py +++ b/tests/unit/litellm_core_utils/test_logging_worker.py @@ -82,6 +82,252 @@ class _RecordCollector(logging.Handler): self.records.append(record) +@pytest.mark.asyncio +@pytest.mark.parametrize("running", [False, True]) +async def test_stop_discards_waiting_callbacks_without_executing_them(running: bool) -> None: + worker: Final = LoggingWorker(concurrency=1) + started: Final = asyncio.Event() + cancelled: Final = asyncio.Event() + executed: Final = asyncio.Event() + + async def blocked() -> None: + started.set() + try: + await asyncio.Event().wait() + finally: + cancelled.set() + + async def waiting() -> None: + executed.set() + + if running: + worker.ensure_initialized_and_enqueue(blocked()) + await started.wait() + else: + worker._ensure_queue() + worker.enqueue(waiting()) + await worker.stop(discard_pending=True) + assert not executed.is_set() + assert cancelled.is_set() == running + assert worker._queue is not None + assert worker._queue.empty() + await asyncio.wait_for(worker._queue.join(), timeout=1) + + +@pytest.mark.asyncio +async def test_stop_does_not_replay_a_callback_dequeued_before_its_processor_started() -> None: + worker: Final = LoggingWorker(concurrency=1) + executed: Final = asyncio.Event() + + async def callback() -> None: + executed.set() + + worker.ensure_initialized_and_enqueue(callback()) + await asyncio.sleep(0) + await worker.stop(discard_pending=True) + assert not executed.is_set() + worker._flush_on_exit() + assert not executed.is_set() + assert worker._queue is not None + await asyncio.wait_for(worker._queue.join(), timeout=1) + + +@pytest.mark.asyncio +async def test_stop_timeout_keeps_callbacks_observable_until_queue_cleanup_finishes() -> None: + worker: Final = LoggingWorker(concurrency=1) + started: Final = asyncio.Event() + release: Final = asyncio.Event() + completed: Final = asyncio.Event() + + async def running() -> None: + started.set() + await asyncio.Event().wait() + + async def queued() -> None: + await release.wait() + completed.set() + + worker.ensure_initialized_and_enqueue(running()) + await started.wait() + worker.enqueue(queued()) + try: + await worker.stop(timeout=0) + assert not completed.is_set() + release.set() + await worker.flush() + assert completed.is_set(), "Queue cleanup must survive a bounded cancellation wait" + finally: + release.set() + await worker.stop() + + +@pytest.mark.asyncio +async def test_overflow_callback_does_not_undercount_the_unfinished_queue() -> None: + """A callback rejected by a full queue is run by the aggressive clear. It must not consume a + queue.task_done() it never contributed, otherwise flush() returns while a real callback is still + running and the shutdown drain cancels it with budget to spare. + """ + worker: Final = LoggingWorker(timeout=5.0, max_queue_size=2, concurrency=1) + gates: Final = {name: asyncio.Event() for name in ("first", "second", "third")} + completed: Final[list[str]] = [] + + async def gated(name: str) -> None: + await gates[name].wait() + completed.append(name) + + async def instant() -> None: + completed.append("overflow") + + worker.start() + worker.enqueue(gated("first")) + await asyncio.sleep(0) + worker.enqueue(gated("second")) + worker.enqueue(gated("third")) + assert worker._queue is not None and worker._queue.full(), "precondition: the queue is full" + worker.enqueue(instant()) + for _ in range(5): + await asyncio.sleep(0) + + gates["first"].set() + gates["second"].set() + asyncio.get_running_loop().call_later(0.2, gates["third"].set) + await asyncio.wait_for(worker.flush(), timeout=2.0) + assert "third" in completed, "flush returned while a queued callback was still unfinished" + assert sorted(completed) == ["first", "overflow", "second", "third"] + await worker.stop() + + +@pytest.mark.asyncio +@pytest.mark.parametrize("queue_size", (1, 2)) +async def test_flush_waits_for_delayed_overflow_and_preserves_callback_context(queue_size: int) -> None: + retry_started: Final = asyncio.Event() + release_retry: Final = asyncio.Event() + completed: Final[asyncio.Queue[tuple[int, str]]] = asyncio.Queue() + identity: Final = contextvars.ContextVar("overflow_identity", default="outside") + + async def retry_sleep(delay: float) -> None: + retry_started.set() + await release_retry.wait() + + worker: Final = LoggingWorker(max_queue_size=queue_size, concurrency=1, retry_sleep=retry_sleep) + + async def callback(index: int) -> None: + completed.put_nowait((index, identity.get())) + + worker.start() + token: Final = identity.set("request") + for index in range(4): + worker.enqueue(callback(index)) + identity.reset(token) + flushing: Final = asyncio.create_task(worker.flush()) + try: + await asyncio.wait_for(retry_started.wait(), timeout=2) + assert worker._queue is not None + await asyncio.wait_for(worker._queue.join(), timeout=2) + await asyncio.sleep(0) + assert not flushing.done(), "Queue completion must not hide delayed overflow callbacks" + assert worker.pending_count > 0 + release_retry.set() + await asyncio.wait_for(flushing, timeout=2) + assert sorted(completed.get_nowait() for _ in range(completed.qsize())) == [ + (index, "request") for index in range(4) + ] + assert worker.pending_count == 0 + finally: + release_retry.set() + await worker.stop(discard_pending=True) + await asyncio.gather(flushing, return_exceptions=True) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("start_retry", (False, True)) +async def test_shutdown_discards_overflow_without_late_enqueue_or_replay(start_retry: bool) -> None: + import inspect + + retry_started: Final = asyncio.Event() + release_retry: Final = asyncio.Event() + executed: Final[asyncio.Queue[int]] = asyncio.Queue() + + async def retry_sleep(delay: float) -> None: + retry_started.set() + await release_retry.wait() + + worker: Final = LoggingWorker(max_queue_size=2, concurrency=1, retry_sleep=retry_sleep) + + async def callback(index: int) -> None: + executed.put_nowait(index) + + callbacks: Final = tuple(callback(index) for index in range(4)) + worker._ensure_queue() + for callback_coro in callbacks: + worker.enqueue(callback_coro) + if start_retry: + await asyncio.wait_for(retry_started.wait(), timeout=2) + await worker.stop(discard_pending=True, timeout=1) + completed_at_stop: Final = executed.qsize() + release_retry.set() + await worker.flush() + try: + assert executed.qsize() == completed_at_stop, "Discarded overflow must not run when the worker restarts" + assert worker.pending_count == 0 + assert all(inspect.getcoroutinestate(coro) == inspect.CORO_CLOSED for coro in callbacks) + finally: + await worker.stop(discard_pending=True) + + +@pytest.mark.asyncio +async def test_shutdown_cancels_aggressive_clear_callbacks_before_returning() -> None: + started: Final = asyncio.Event() + cancelled: Final = asyncio.Event() + worker: Final = LoggingWorker(max_queue_size=2, concurrency=1) + + async def blocked() -> None: + started.set() + try: + await asyncio.Event().wait() + finally: + cancelled.set() + + async def noop() -> None: + pass + + worker._ensure_queue() + worker.enqueue(blocked()) + worker.enqueue(noop()) + worker.enqueue(noop()) + await asyncio.wait_for(started.wait(), timeout=2) + await worker.stop(discard_pending=True, timeout=1) + assert cancelled.is_set(), "Overflow execution must participate in shutdown cancellation" + assert worker.pending_count == 0 + assert worker._queue is not None + await asyncio.wait_for(worker._queue.join(), timeout=1) + + +@pytest.mark.asyncio +async def test_stopped_overflow_can_be_flushed_when_not_discarded() -> None: + release_retry: Final = asyncio.Event() + completed: Final[asyncio.Queue[int]] = asyncio.Queue() + + async def retry_sleep(delay: float) -> None: + await release_retry.wait() + + worker: Final = LoggingWorker(max_queue_size=2, concurrency=1, retry_sleep=retry_sleep) + + async def callback(index: int) -> None: + completed.put_nowait(index) + + worker._ensure_queue() + for index in range(4): + worker.enqueue(callback(index)) + await worker.stop() + release_retry.set() + try: + await asyncio.wait_for(worker.flush(), timeout=2) + assert sorted(completed.get_nowait() for _ in range(completed.qsize())) == list(range(4)) + finally: + await worker.stop(discard_pending=True) + + class TestLoggingWorker: """Test cases for LoggingWorker functionality.""" diff --git a/tests/unit/proxy/guardrails/test_deferred_guardrail_logging.py b/tests/unit/proxy/guardrails/test_deferred_guardrail_logging.py index bd6190fee5f..6d1e5c72982 100644 --- a/tests/unit/proxy/guardrails/test_deferred_guardrail_logging.py +++ b/tests/unit/proxy/guardrails/test_deferred_guardrail_logging.py @@ -75,11 +75,13 @@ class _RecordingLogger(CustomLogger): def __init__(self) -> None: super().__init__() self.standard_logging_object: StandardLoggingPayload | None = None + self.logged: asyncio.Event = asyncio.Event() async def async_log_success_event( self, kwargs: Mapping[str, object], response_obj: object, start_time: datetime, end_time: datetime ) -> None: self.standard_logging_object = cast(StandardLoggingPayload, kwargs["standard_logging_object"]) + self.logged.set() class PostCallGuardrail(CustomGuardrail): @@ -239,52 +241,33 @@ class TestHasPostCallGuardrailsForPassthrough: @pytest.mark.asyncio -async def test_deferred_flag_stores_and_executes_closure(): - """ - When defer_async_logging is True on logging_obj: - 1. wrapper_async stores a callable closure instead of calling create_task - 2. Calling the closure fires create_task - 3. Sync callbacks fire immediately (not deferred) - """ - mock_logging_obj = MagicMock() - mock_logging_obj.defer_async_logging = True - mock_logging_obj.enqueue_deferred_logging = None - - await litellm.acompletion( - model="gpt-3.5-turbo", +@pytest.mark.parametrize("deferred", (False, True)) +async def test_success_logging_waits_for_deferred_closure_when_requested(deferred: bool) -> None: + recorder: Final = _RecordingLogger() + logging_obj: Final = Logging( + model="gpt-4o-mini", + messages=[{"role": "user", "content": "hi"}], + stream=False, + call_type="acompletion", + start_time=datetime(2026, 1, 1), + litellm_call_id="deferred-success", + function_id="deferred-success", + dynamic_async_success_callbacks=[recorder], + ) + logging_obj.defer_async_logging = deferred + response: Final = await litellm.acompletion( + model="gpt-4o-mini", messages=[{"role": "user", "content": "hi"}], mock_response="Hello!", - litellm_logging_obj=mock_logging_obj, + litellm_logging_obj=logging_obj, ) - - # Closure was stored - enqueue_fn = mock_logging_obj.enqueue_deferred_logging - assert callable(enqueue_fn), "Closure should be stored on logging_obj" - - # Sync callbacks fired immediately - mock_logging_obj.handle_sync_success_callbacks_for_async_calls.assert_called_once() - - # Calling the closure fires create_task - created_tasks = [] - real_create_task = asyncio.create_task - - def tracking_create_task(coro): - task = real_create_task(coro) - created_tasks.append(task) - return task - - with patch("asyncio.create_task", side_effect=tracking_create_task): - enqueue_fn() - - assert len(created_tasks) >= 1, "Closure should fire asyncio.create_task" - - for task in created_tasks: - if not task.done(): - task.cancel() - try: - await task - except (asyncio.CancelledError, Exception): - pass + if deferred: + assert recorder.standard_logging_object is None + assert logging_obj.enqueue_deferred_logging is not None + logging_obj.enqueue_deferred_logging() + await asyncio.wait_for(recorder.logged.wait(), timeout=5) + assert recorder.standard_logging_object is not None + assert recorder.standard_logging_object["id"] == response.id @pytest.mark.asyncio @@ -401,40 +384,6 @@ async def test_deferred_anthropic_messages_bridged_to_the_responses_api_logs_the assert recorder.standard_logging_object["response_cost"] == pytest.approx(3 * 2e-7 + 7333 * 2e-8 + 23 * 1.25e-6) -# --------------------------------------------------------------------------- -# 3. Non-streaming regression: without flag, create_task fires normally -# --------------------------------------------------------------------------- - - -@pytest.mark.asyncio -async def test_no_flag_fires_create_task_normally(): - """Without defer_async_logging, wrapper_async calls create_task as before.""" - created_tasks = [] - real_create_task = asyncio.create_task - - def tracking_create_task(coro): - task = real_create_task(coro) - created_tasks.append(task) - return task - - with patch("asyncio.create_task", side_effect=tracking_create_task): - await litellm.acompletion( - model="gpt-3.5-turbo", - messages=[{"role": "user", "content": "hi"}], - mock_response="Hello!", - ) - - assert len(created_tasks) >= 1 - - for task in created_tasks: - if not task.done(): - task.cancel() - try: - await task - except (asyncio.CancelledError, Exception): - pass - - # --------------------------------------------------------------------------- # 4. Non-streaming: deferred success log is suppressed when guardrail raises # --------------------------------------------------------------------------- diff --git a/tests/unit/proxy/proxy_server/test_lifecycle.py b/tests/unit/proxy/proxy_server/test_lifecycle.py index 8b1d7c1222a..a8706c5bccd 100644 --- a/tests/unit/proxy/proxy_server/test_lifecycle.py +++ b/tests/unit/proxy/proxy_server/test_lifecycle.py @@ -35,6 +35,7 @@ from typing_extensions import TypedDict import litellm.proxy.proxy_server as ps from litellm.proxy.management_helpers.audit_logs import track_audit_task from litellm.proxy.proxy_server import ( + _proxy_startup_event, ProxyStartupEvent, _resolve_pydantic_type, _resolve_typed_dict_type, @@ -1059,7 +1060,7 @@ def test_otel_global_provider_published_after_callback_init(): span was orphaned. The publish (``publish_global_otel_v2_provider``) must therefore appear after ``_initialize_startup_logging`` in the lifespan source. """ - wrapped = getattr(proxy_startup_event, "__wrapped__", proxy_startup_event) + wrapped = getattr(_proxy_startup_event, "__wrapped__", _proxy_startup_event) source = inspect.getsource(wrapped) init_pos = source.find("ProxyStartupEvent.initialize_startup_logging(") publish_pos = source.find("publish_global_otel_v2_provider(") @@ -1145,7 +1146,7 @@ def test_proxy_startup_event_warns_for_global_budget_without_database(): The call must follow Prisma setup so DB-backed deployments do not false-positive. Direct ``_warn_budget_without_db`` tests cover the warning behavior itself. """ - wrapped = getattr(proxy_startup_event, "__wrapped__", proxy_startup_event) + wrapped = getattr(_proxy_startup_event, "__wrapped__", _proxy_startup_event) source = inspect.getsource(wrapped) budget_check_pos = source.find("if prisma_client is not None and litellm.max_budget > 0:") warn_pos = source.find("warn_budget_without_db(") diff --git a/tests/unit/proxy/shutdown/test_graceful_shutdown_manager.py b/tests/unit/proxy/shutdown/test_graceful_shutdown_manager.py index d38852617b9..fdf1adda1f4 100644 --- a/tests/unit/proxy/shutdown/test_graceful_shutdown_manager.py +++ b/tests/unit/proxy/shutdown/test_graceful_shutdown_manager.py @@ -179,3 +179,107 @@ async def test_emits_periodic_drain_waiting_log_while_waiting(): timeout=10, count_fn=count_fn, poll_interval=0, log_interval=0 ) assert drained == 3 + + +@pytest.mark.parametrize("elapsed, cap, expected", [(4, None, 26), (25, 2, 2), (29, 5, 1), (31, None, 0)]) +def test_remaining_budget_includes_time_before_lifespan( + monkeypatch: pytest.MonkeyPatch, elapsed: float, cap: float | None, expected: float +) -> None: + monkeypatch.setenv("GRACEFUL_SHUTDOWN_TIMEOUT", "30") + GracefulShutdownManager.start_shutdown(clock=lambda: 100.0) + GracefulShutdownManager.start_shutdown(clock=lambda: 100.0 + elapsed) + assert GracefulShutdownManager.remaining(cap, clock=lambda: 100.0 + elapsed) == expected + + +def test_remaining_budget_before_shutdown(monkeypatch: pytest.MonkeyPatch) -> None: + monkeypatch.setenv("GRACEFUL_SHUTDOWN_TIMEOUT", "30") + assert GracefulShutdownManager.remaining(clock=lambda: 100.0) == 30.0 + assert GracefulShutdownManager.remaining(-1, clock=lambda: 100.0) == 0.0 + + +@pytest.mark.parametrize("prestop", [False, True]) +def test_sigterm_stamps_before_chaining_and_restores_server_handler( + monkeypatch: pytest.MonkeyPatch, prestop: bool +) -> None: + import signal + from queue import SimpleQueue + from types import FrameType + from typing import Final + + monkeypatch.setenv("GRACEFUL_SHUTDOWN_TIMEOUT", "30") + calls: Final[SimpleQueue[float]] = SimpleQueue() + + def server_handler(signum: int, frame: FrameType | None) -> None: + assert signum == signal.SIGTERM + calls.put(GracefulShutdownManager.remaining(clock=lambda: 105.0)) + + original: Final = signal.signal(signal.SIGTERM, server_handler) + try: + if prestop: + GracefulShutdownManager.start_shutdown(clock=lambda: 95.0) + with GracefulShutdownManager.capture_sigterm(clock=lambda: 100.0): + installed: Final = signal.getsignal(signal.SIGTERM) + with GracefulShutdownManager.capture_sigterm(clock=lambda: 102.0): + assert signal.getsignal(signal.SIGTERM) is installed + signal.raise_signal(signal.SIGTERM) + signal.raise_signal(signal.SIGTERM) + assert calls.get_nowait() == (20.0 if prestop else 25.0) + assert calls.get_nowait() == (20.0 if prestop else 25.0) + assert calls.empty() + assert signal.getsignal(signal.SIGTERM) is server_handler + finally: + signal.signal(signal.SIGTERM, original) + + +def test_sigterm_capture_preserves_a_replacement_handler() -> None: + import signal + from types import FrameType + from typing import Final + + def server_handler(signum: int, frame: FrameType | None) -> None: + pass + + original: Final = signal.signal(signal.SIGTERM, server_handler) + try: + with GracefulShutdownManager.capture_sigterm(): + signal.signal(signal.SIGTERM, signal.SIG_IGN) + assert signal.getsignal(signal.SIGTERM) is signal.SIG_IGN + finally: + signal.signal(signal.SIGTERM, original) + + +@pytest.mark.parametrize("handler", ["SIG_DFL", "SIG_IGN"]) +def test_sigterm_capture_preserves_default_and_ignored_dispositions(handler: str) -> None: + import signal + from typing import Final + + disposition: Final = signal.SIG_DFL if handler == "SIG_DFL" else signal.SIG_IGN + original: Final = signal.signal(signal.SIGTERM, disposition) + try: + with GracefulShutdownManager.capture_sigterm(): + assert signal.getsignal(signal.SIGTERM) is disposition + finally: + signal.signal(signal.SIGTERM, original) + + +def test_sigterm_capture_in_worker_thread_leaves_signal_ownership_unchanged() -> None: + import signal + from concurrent.futures import ThreadPoolExecutor + from typing import Final + + original: Final = signal.getsignal(signal.SIGTERM) + + def run_lifespan() -> None: + with GracefulShutdownManager.capture_sigterm(): + assert signal.getsignal(signal.SIGTERM) is original + + with ThreadPoolExecutor(max_workers=1) as executor: + executor.submit(run_lifespan).result() + assert signal.getsignal(signal.SIGTERM) is original + + +@pytest.mark.asyncio +async def test_lifespan_request_drain_does_not_restart_an_exhausted_budget(monkeypatch: pytest.MonkeyPatch) -> None: + monkeypatch.setenv("GRACEFUL_SHUTDOWN_TIMEOUT", "30") + GracefulShutdownManager.start_shutdown(clock=lambda: 100.0) + assert await GracefulShutdownManager.wait_for_drain(count_fn=lambda: 2, clock=lambda: 130.0) == 2 diff --git a/tests/unit/proxy/test_proxy_server.py b/tests/unit/proxy/test_proxy_server.py index 22d6316c448..c835e1e27ea 100644 --- a/tests/unit/proxy/test_proxy_server.py +++ b/tests/unit/proxy/test_proxy_server.py @@ -263,6 +263,174 @@ async def test_shutdown_commits_spend_a_logging_task_queued_while_the_first_comm token = MASTER_KEY + +@pytest.mark.asyncio +async def test_shutdown_persists_accounting_from_worker_callbacks_before_returning() -> None: + from litellm.litellm_core_utils.logging_worker import LoggingWorker + from litellm.responses.streaming_iterator import spawn_logging_task + + worker: Final = LoggingWorker() + started: Final = asyncio.Event() + release: Final = asyncio.Event() + first_flush: Final = asyncio.Event() + pending: Final[asyncio.Queue[str]] = asyncio.Queue() + persisted: Final[asyncio.Queue[str]] = asyncio.Queue() + + async def enqueue_spend() -> None: + pending.put_nowait("accounted request") + + async def callback() -> None: + started.set() + await release.wait() + spawn_logging_task(asyncio.get_running_loop(), enqueue_spend(), task_name="accounting") + + async def flush_spend() -> None: + while not pending.empty(): + persisted.put_nowait(pending.get_nowait()) + first_flush.set() + + worker.ensure_initialized_and_enqueue(callback()) + await started.wait() + shutdown: Final = asyncio.create_task( + litellm.proxy.proxy_server._flush_spend_around_logging_drain_on_shutdown(flush_spend, worker=worker) + ) + try: + await first_flush.wait() + waited_for_callback: Final = not shutdown.done() + release.set() + await shutdown + assert waited_for_callback, "Shutdown returned while accounting callbacks were still running" + assert persisted.get_nowait() == "accounted request" + assert pending.empty() + finally: + release.set() + await worker.flush() + await worker.stop() + + +@pytest.mark.asyncio +async def test_shutdown_reports_worker_timeout_and_flushes_completed_accounting( + monkeypatch: pytest.MonkeyPatch, caplog: pytest.LogCaptureFixture +) -> None: + from litellm.litellm_core_utils.logging_worker import LoggingWorker + from litellm.proxy.shutdown.graceful_shutdown_manager import GracefulShutdownManager + + monkeypatch.setenv("GRACEFUL_SHUTDOWN_TIMEOUT", "30") + worker: Final = LoggingWorker() + started: Final = asyncio.Event() + cancelled: Final = asyncio.Event() + flushed_after_cancellation: Final = asyncio.Event() + GracefulShutdownManager.start_shutdown(clock=lambda: 100.0) + + async def callback() -> None: + started.set() + try: + await asyncio.Event().wait() + finally: + cancelled.set() + + async def flush_spend() -> None: + if cancelled.is_set(): + flushed_after_cancellation.set() + + worker.ensure_initialized_and_enqueue(callback()) + await started.wait() + try: + with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): + await litellm.proxy.proxy_server._flush_spend_around_logging_drain_on_shutdown( + flush_spend, worker=worker, clock=lambda: 120.0 + ) + assert cancelled.is_set() + assert flushed_after_cancellation.is_set() + assert "budget exhausted with 1 unfinished callbacks" in caplog.text + finally: + GracefulShutdownManager.reset() + await worker.stop() + + +@pytest.mark.asyncio +async def test_shutdown_bounds_callback_cancellation_and_flushes_accounting( + monkeypatch: pytest.MonkeyPatch, caplog: pytest.LogCaptureFixture +) -> None: + from litellm.litellm_core_utils.logging_worker import LoggingWorker + from litellm.proxy.shutdown.graceful_shutdown_manager import GracefulShutdownManager + + monkeypatch.setenv("GRACEFUL_SHUTDOWN_TIMEOUT", "30") + worker: Final = LoggingWorker() + started: Final = asyncio.Event() + cancelling: Final = asyncio.Event() + release: Final = asyncio.Event() + finished: Final = asyncio.Event() + flushed: Final = asyncio.Event() + GracefulShutdownManager.start_shutdown(clock=lambda: 100.0) + + async def callback() -> None: + started.set() + try: + await release.wait() + except asyncio.CancelledError: + cancelling.set() + await release.wait() + finally: + finished.set() + + async def drain_tasks(timeout: float) -> None: + assert timeout == 0.0 + await cancelling.wait() + + async def flush_spend() -> None: + if cancelling.is_set(): + flushed.set() + + worker.ensure_initialized_and_enqueue(callback()) + await started.wait() + try: + with caplog.at_level(logging.WARNING, logger="LiteLLM"): + await litellm.proxy.proxy_server._flush_spend_around_logging_drain_on_shutdown( + flush_spend, worker=worker, clock=lambda: 130.0, drain_tasks=drain_tasks + ) + assert not finished.is_set(), "The cancellation-resistant callback must remain observable" + assert flushed.is_set() + assert "tasks still alive after cancellation" in caplog.text + finally: + release.set() + await worker.flush() + await worker.stop() + GracefulShutdownManager.reset() + + +@pytest.mark.asyncio +async def test_shutdown_preserves_persistence_reserve_after_partial_http_drain( + monkeypatch: pytest.MonkeyPatch, caplog: pytest.LogCaptureFixture +) -> None: + from litellm.litellm_core_utils.logging_worker import LoggingWorker + from litellm.proxy.shutdown.graceful_shutdown_manager import GracefulShutdownManager + + monkeypatch.setenv("GRACEFUL_SHUTDOWN_TIMEOUT", "30") + worker: Final = LoggingWorker() + persisted: Final = asyncio.Event() + drain_budgets: Final[asyncio.Queue[float]] = asyncio.Queue() + GracefulShutdownManager.start_shutdown(clock=lambda: 100.0) + + async def flush_spend() -> None: + persisted.set() + + async def drain_tasks(timeout: float) -> None: + drain_budgets.put_nowait(timeout) + + try: + with caplog.at_level(logging.INFO, logger="LiteLLM Proxy"): + await litellm.proxy.proxy_server._flush_spend_around_logging_drain_on_shutdown( + flush_spend, worker=worker, clock=lambda: 106.0, drain_tasks=drain_tasks + ) + assert "remaining_s=24.000 callback_timeout_s=14.000" in caplog.text + assert persisted.is_set() + assert tuple(drain_budgets.get_nowait() for _ in range(drain_budgets.qsize())) == (14.0, 5.0) + finally: + await worker.stop() + GracefulShutdownManager.reset() + + headers = {"Authorization": f"Bearer {token}"} example_completion_result = { @@ -4059,3 +4227,34 @@ def test_detailed_debug_logs_request_body_only_as_received_and_as_sent( assert len(body_dumps) == 2, [message[:120] for message in body_dumps] assert body_dumps[0].startswith("Request received by LiteLLM:") assert "POST Request Sent from LiteLLM" in body_dumps[1] + + +@pytest.mark.asyncio +async def test_shutdown_waits_for_enqueue_producers_before_stopping_worker() -> None: + from litellm.litellm_core_utils.logging_worker import LoggingWorker + from litellm.responses.streaming_iterator import spawn_logging_task + + worker: Final = LoggingWorker() + persisted: Final[asyncio.Queue[str]] = asyncio.Queue() + accounting: Final[asyncio.Queue[str]] = asyncio.Queue() + + async def callback() -> None: + accounting.put_nowait("completed request") + + async def producer() -> None: + await asyncio.sleep(0) + await asyncio.sleep(0) + worker.ensure_initialized_and_enqueue(callback()) + + async def flush_spend() -> None: + while not accounting.empty(): + persisted.put_nowait(accounting.get_nowait()) + + spawn_logging_task(asyncio.get_running_loop(), producer(), task_name="success enqueue") + try: + await litellm.proxy.proxy_server._flush_spend_around_logging_drain_on_shutdown(flush_spend, worker=worker) + assert persisted.qsize() == 1, "An enqueue producer must finish before callback draining stops" + assert persisted.get_nowait() == "completed request" + assert worker.pending_count == 0 + finally: + await worker.stop(discard_pending=True) diff --git a/tests/unit/test_utils.py b/tests/unit/test_utils.py index cd56b6e1b46..f3b470044ae 100644 --- a/tests/unit/test_utils.py +++ b/tests/unit/test_utils.py @@ -11294,6 +11294,7 @@ def test_get_optional_params_forwards_a_tool_config_cache_point_to_a_downstream_ assert optional_params["extra_body"] == {"kept": True, "cache_control_injection_points": [point]} + @pytest.mark.parametrize( ("provider", "expected_warnings"), (("streamlake", 1), ("veniceai", 1), ("my-own-llm", 0)), @@ -11312,3 +11313,34 @@ def test_custom_provider_named_like_a_built_in_warns_once( warnings: Final = [r for r in caplog.records if f"provider '{provider}'" in r.getMessage()] assert len(warnings) == expected_warnings assert provider in litellm._custom_providers + + +@pytest.mark.asyncio +@pytest.mark.parametrize("deferred", (False, True)) +async def test_success_enqueue_is_tracked_without_starting_callbacks_synchronously(deferred: bool) -> None: + from litellm.litellm_core_utils.litellm_logging import Logging + from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER + from litellm.responses.streaming_iterator import drain_logging_tasks + from litellm.utils import _schedule_async_success_logging + + now: Final = datetime(2026, 1, 1, tzinfo=timezone.utc) + logging_obj: Final = Logging( + model="gpt-4o-mini", + messages=[], + stream=False, + call_type="acompletion", + start_time=now, + litellm_call_id="shutdown-enqueue", + function_id="shutdown-enqueue", + ) + logging_obj.defer_async_logging = deferred + _schedule_async_success_logging(logging_obj, ModelResponse(), now, now, False) + if deferred: + assert not logging_obj.async_success_log_enqueued + assert logging_obj.enqueue_deferred_logging is not None + logging_obj.enqueue_deferred_logging() + assert "has_logged_async_success" not in logging_obj.model_call_details + await drain_logging_tasks() + assert GLOBAL_LOGGING_WORKER.pending_count > 0, "Tracked enqueue must finish before worker draining starts" + await GLOBAL_LOGGING_WORKER.flush() + assert logging_obj.model_call_details["has_logged_async_success"] is True From 4bc157e281eab779cb76f78044e291c21604b211 Mon Sep 17 00:00:00 2001 From: Kerry Lu Date: Thu, 29 Oct 2026 09:39:26 -0700 Subject: [PATCH 6/7] feat(proxy): replace the pre-migration row check with a rolled-back dry run of the pending migrations (#48961) * feat(proxy): replace the pre-migration row check with a rolled-back dry run of the pending migrations * feat(proxy): roll back each migration of the dry run on its own and baseline-check databases without a ledger * test(proxy): cover the dry run CLI and PrismaManager wrapper without a database * fix(proxy): refuse the no-ledger dry run for a custom migration folder like boot does, and probe the rolled-back indexes directly * ci(codecov): stop carrying forward the roi-postgres and redis-compat flags Their carried-forward sessions were measured against older proxy_cli.py revisions (633 and 689 line maps against 710 today) and mark docstring lines of the merged file as missed, which failed codecov/patch on this PR at 70% while every unit shard reports the lines hit --------- Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- codecov.yaml | 11 +- .../migration_data_checks.py | 252 +----------------- .../litellm_proxy_extras/migration_dry_run.py | 184 +++++++++++++ .../migration_recovery.py | 21 +- .../litellm_proxy_extras/utils.py | 107 ++++---- litellm/proxy/db/prisma_client.py | 12 +- litellm/proxy/proxy_cli.py | 18 +- .../migrations/test_request_log_indexes.py | 248 ++++++++++++----- .../test_migration_data_checks.py | 96 +------ .../test_migration_dry_run.py | 80 ++++++ tests/unit/proxy/db/test_prisma_client.py | 16 ++ tests/unit/proxy/test_proxy_cli.py | 17 ++ 12 files changed, 591 insertions(+), 471 deletions(-) create mode 100644 litellm-proxy-extras/litellm_proxy_extras/migration_dry_run.py create mode 100644 tests/unit/litellm_proxy_extras/test_migration_dry_run.py diff --git a/codecov.yaml b/codecov.yaml index 5b30e78a8fb..521088868bd 100644 --- a/codecov.yaml +++ b/codecov.yaml @@ -15,9 +15,10 @@ ignore: flag_management: default_rules: carryforward: true - # Dead flags no CI job uploads anymore: their carried-forward sessions were - # measured against old revisions, and the stale line maps mark comment lines - # of since-edited files as missed, sinking patch coverage on unrelated PRs. + # Dead flags no CI job uploads anymore, and flags whose job is path-gated on + # PRs: their carried-forward sessions were measured against old revisions, and + # the stale line maps mark comment lines of since-edited files as missed, + # sinking patch coverage on unrelated PRs. individual_flags: - name: proxy-mgmt-behavior carryforward: false @@ -99,6 +100,10 @@ flag_management: carryforward: false - name: proxy-db-endpoints-and-responses carryforward: false + - name: roi-postgres + carryforward: false + - name: redis-compat + carryforward: false component_management: individual_components: diff --git a/litellm-proxy-extras/litellm_proxy_extras/migration_data_checks.py b/litellm-proxy-extras/litellm_proxy_extras/migration_data_checks.py index e22639daaef..75c5bcf4e1e 100644 --- a/litellm-proxy-extras/litellm_proxy_extras/migration_data_checks.py +++ b/litellm-proxy-extras/litellm_proxy_extras/migration_data_checks.py @@ -14,11 +14,9 @@ either, so the migration keeps today's behavior for it. """ import os -import re -from collections.abc import Callable, Iterator +from collections.abc import Callable from dataclasses import dataclass from functools import reduce -from itertools import accumulate from pathlib import Path from types import MappingProxyType from typing import TYPE_CHECKING, Final, Literal @@ -43,7 +41,7 @@ MIGRATION_DATA_CHECK_TIMEOUT_ENV_VAR: Final = "LITELLM_MIGRATION_DATA_CHECK_TIME DEFAULT_MIGRATION_DATA_CHECK_TIMEOUT_SECONDS: Final = 600.0 REFUSAL: Final = "Existing rows would make a pending schema change fail:" -Kind = Literal["too_long", "duplicates", "nulls", "any_row", "orphans", "check_violations"] +Kind = Literal["too_long"] @dataclass(frozen=True, slots=True) @@ -54,9 +52,6 @@ class Check: statement: str source: str index: str | None = None - referenced_table: str | None = None - referenced_columns: tuple[str, ...] = () - expression: str | None = None @dataclass(frozen=True, slots=True) @@ -100,33 +95,6 @@ class ColumnType: ColumnTypeLookup = Callable[[str, str], ColumnType | None] -_IDENT: Final = r'"?(\w+)"?' -_QUALIFIED: Final = r'(?:"?\w+"?\s*\.\s*)?"?(\w+)"?' -_CREATE_INDEX: Final = re.compile( - rf"^\s*CREATE\s+(UNIQUE\s+)?INDEX\s+(?:CONCURRENTLY\s+)?(?:IF\s+NOT\s+EXISTS\s+)?{_IDENT}\s+ON\s+(?:ONLY\s+)?" - rf"{_QUALIFIED}\s*(?:USING\s+(\w+)\s*)?\((.*?)\)", - re.IGNORECASE | re.DOTALL, -) -_ALTER_TABLE: Final = re.compile( - rf"^\s*ALTER\s+TABLE\s+(?:IF\s+EXISTS\s+)?(?:ONLY\s+)?{_QUALIFIED}\s+(.*)$", re.IGNORECASE | re.DOTALL -) -_SET_NOT_NULL: Final = re.compile(rf"^ALTER\s+(?:COLUMN\s+)?{_IDENT}\s+SET\s+NOT\s+NULL$", re.IGNORECASE) -_ADD_COLUMN: Final = re.compile( - rf"^ADD\s+(?:COLUMN\s+)?(?:IF\s+NOT\s+EXISTS\s+)?(?!(?:CONSTRAINT|FOREIGN|UNIQUE|CHECK|PRIMARY|EXCLUDE)\b){_IDENT}\s+(.*)$", - re.IGNORECASE | re.DOTALL, -) -_CONSTRAINT: Final = rf"^ADD\s+(?:CONSTRAINT\s+{_IDENT}\s+)?" -_ADD_FOREIGN_KEY: Final = re.compile( - rf"{_CONSTRAINT}FOREIGN\s+KEY\s*\(([^)]*)\)\s*REFERENCES\s+{_QUALIFIED}\s*\(([^)]*)\)(.*)$", - re.IGNORECASE | re.DOTALL, -) -_ADD_UNIQUE: Final = re.compile(rf"{_CONSTRAINT}UNIQUE\s*\(([^)]*)\)", re.IGNORECASE | re.DOTALL) -_ADD_CHECK: Final = re.compile(rf"{_CONSTRAINT}CHECK\s*\((.*)\)\s*(NOT\s+VALID)?\s*$", re.IGNORECASE | re.DOTALL) -_PLAIN_COLUMN: Final = re.compile(rf"^{_IDENT}(?:\s+(?:ASC|DESC|NULLS\s+(?:FIRST|LAST)))*$", re.IGNORECASE) -_COMMENT: Final = re.compile(r"--[^\n]*") -_NOT_NULL: Final = re.compile(r"\bNOT\s+NULL\b", re.IGNORECASE) -_HAS_DEFAULT: Final = re.compile(r"\bDEFAULT\b|\bGENERATED\b", re.IGNORECASE) - def migration_data_check_timeout_seconds() -> float: """Seconds one count query may run before the preflight gives up on verifying it.""" @@ -145,117 +113,6 @@ def migration_data_check_timeout_seconds() -> float: return DEFAULT_MIGRATION_DATA_CHECK_TIMEOUT_SECONDS -def statements(migration_sql: str) -> tuple[str, ...]: - """The statements of one migration file, comments stripped, empty ones dropped.""" - parts: Final = _COMMENT.sub("", migration_sql).split(";") - return tuple(part.strip() for part in parts if part.strip()) - - -def _split_top_level(text: str) -> tuple[str, ...]: - """Split on the commas outside parentheses.""" - depth: Final = accumulate((1 if char == "(" else -1 if char == ")" else 0) for char in text) - cuts: Final = tuple( - position for position, (char, level) in enumerate(zip(text, depth, strict=True)) if char == "," and level == 0 - ) - starts: Final = (0, *(cut + 1 for cut in cuts)) - stops: Final = (*cuts, len(text)) - pieces: Final = tuple(text[start:stop].strip() for start, stop in zip(starts, stops, strict=True)) - return tuple(piece for piece in pieces if piece) - - -def _plain_columns(column_list: str) -> tuple[str, ...] | None: - """The column names of an index or constraint column list, None when one entry is an - expression the derivation cannot turn into a count.""" - matches: Final = tuple(_PLAIN_COLUMN.match(entry) for entry in _split_top_level(column_list)) - if any(match is None for match in matches): - return None - return tuple(match.group(1) for match in matches if match is not None) - - -def _index_checks(match: "re.Match[str]", statement_text: str, source: str) -> Iterator[Check]: - unique, name, table, method, column_list = match.groups() - columns: Final = _plain_columns(column_list) - if columns is None: - return - label: Final = f'CREATE {"UNIQUE " if unique else ""}INDEX "{name}"' - if method is None or method.lower() == "btree": - yield Check("too_long", table, columns, label, source, index=name) - if unique: - yield Check("duplicates", table, columns, label, source, index=name) - - -def _alter_table_checks(table: str, action: str, source: str) -> Iterator[Check]: - if (set_not_null := _SET_NOT_NULL.match(action)) is not None: - yield Check("nulls", table, (set_not_null.group(1),), f'ALTER TABLE "{table}" {action}', source) - return - if (add_column := _ADD_COLUMN.match(action)) is not None: - column, body = add_column.groups() - if _NOT_NULL.search(body) is not None and _HAS_DEFAULT.search(body) is None: - yield Check( - "any_row", table, (column,), f'ALTER TABLE "{table}" ADD COLUMN "{column}" ... NOT NULL', source - ) - return - if (foreign_key := _ADD_FOREIGN_KEY.match(action)) is not None: - constraint, columns, referenced_table, referenced_columns, tail = foreign_key.groups() - own: Final = _plain_columns(columns) - theirs: Final = _plain_columns(referenced_columns) - if own is None or theirs is None or len(own) != len(theirs) or "NOT VALID" in tail.upper(): - return - yield Check( - "orphans", - table, - own, - _constraint_label(table, constraint, "FOREIGN KEY"), - source, - referenced_table=referenced_table, - referenced_columns=theirs, - ) - return - if (unique := _ADD_UNIQUE.match(action)) is not None: - constraint, columns = unique.groups() - unique_columns: Final = _plain_columns(columns) - if unique_columns is not None: - yield Check("duplicates", table, unique_columns, _constraint_label(table, constraint, "UNIQUE"), source) - return - if (check := _ADD_CHECK.match(action)) is not None: - constraint, expression, not_valid = check.groups() - if not_valid is None: - yield Check( - "check_violations", - table, - (), - _constraint_label(table, constraint, "CHECK"), - source, - expression=expression, - ) - - -def _constraint_label(table: str, constraint: str | None, kind: str) -> str: - named: Final = f'CONSTRAINT "{constraint}" ' if constraint else "" - return f'ALTER TABLE "{table}" ADD {named}{kind}' - - -def checks_for_statement(statement_text: str, source: str) -> tuple[Check, ...]: - """The count queries one SQL statement needs before it can be applied, empty for a - statement that cannot fail on existing rows or that the derivation cannot read.""" - if (index := _CREATE_INDEX.match(statement_text)) is not None: - return tuple(_index_checks(index, statement_text, source)) - if (alter := _ALTER_TABLE.match(statement_text)) is not None: - table, actions = alter.groups() - return tuple( - check for action in _split_top_level(actions) for check in _alter_table_checks(table, action, source) - ) # comprehension-ok: flattening the per-action checks of one statement - return () - - -def checks_for_migration(name: str, migration_sql: str) -> tuple[Check, ...]: - return tuple( - check - for statement_text in statements(migration_sql) - for check in checks_for_statement(statement_text, f"migration {name}") - ) # comprehension-ok: flattening the per-statement checks of one migration - - def checks_for_index(table: str, name: str, columns: tuple[str, ...]) -> tuple[Check, ...]: """The entry-size check a request-log index needs before `CREATE INDEX CONCURRENTLY`.""" return (Check("too_long", table, columns, f'CREATE INDEX "{name}"', "request-log index", index=name),) @@ -424,42 +281,7 @@ def _entry_size_query(check: Check, schema: str, types: tuple[ColumnType, ...], def _count_query(check: Check, schema: str, types: tuple[ColumnType, ...], limit_bytes: int) -> "sql.Composed": - from psycopg import sql - - table: Final = sql.Identifier(schema, check.table) - columns: Final = sql.SQL(", ").join(sql.Identifier(column) for column in check.columns) - match check.kind: - case "too_long": - return _entry_size_query(check, schema, types, limit_bytes) - case "duplicates": - not_null: Final = sql.SQL(" AND ").join( - sql.SQL("{} IS NOT NULL").format(sql.Identifier(column)) for column in check.columns - ) - return sql.SQL( - "SELECT count(*) FROM (SELECT 1 FROM {t} WHERE {nn} GROUP BY {cols} HAVING count(*) > 1) AS d" - ).format(t=table, nn=not_null, cols=columns) - case "nulls": - return sql.SQL("SELECT count(*) FROM {t} WHERE {col} IS NULL").format( - t=table, col=sql.Identifier(check.columns[0]) - ) - case "any_row": - return sql.SQL("SELECT count(*) FROM {t}").format(t=table) - case "orphans": - pairs: Final = sql.SQL(" AND ").join( - sql.SQL("parent.{} = child.{}").format(sql.Identifier(theirs), sql.Identifier(own)) - for own, theirs in zip(check.columns, check.referenced_columns, strict=True) - ) - present: Final = sql.SQL(" AND ").join( - sql.SQL("child.{} IS NOT NULL").format(sql.Identifier(column)) for column in check.columns - ) - return sql.SQL( - "SELECT count(*) FROM {t} AS child WHERE {present} AND NOT EXISTS (SELECT 1 FROM {ref} AS parent WHERE {pairs})" - ).format(t=table, present=present, ref=sql.Identifier(schema, check.referenced_table or ""), pairs=pairs) - case "check_violations": - return sql.SQL("SELECT count(*) FROM {t} WHERE NOT ({expr})").format( - t=table, - expr=sql.SQL(check.expression or "true"), # pyright: ignore[reportArgumentType] # copied from our own migration file - ) + return _entry_size_query(check, schema, types, limit_bytes) def _applies( @@ -474,15 +296,7 @@ def _applies( return False if check.index is not None and _index_is_valid(connection, schema, check.index): return False - if check.kind == "any_row": - return types is None - if types is None: - return False - if check.kind == "too_long": - return any(column.can_overflow_btree for column in types) - if check.kind == "orphans": - return check.referenced_table is not None and _table_exists(connection, schema, check.referenced_table) - return True + return types is not None and any(column.can_overflow_btree for column in types) def _too_long_detail(check: Check, types: tuple[ColumnType, ...], values: tuple[object, ...], limit_bytes: int) -> str: @@ -515,21 +329,7 @@ def _run_check( rows: Final = _int(values[0], 0) if values else 0 if rows == 0: return None - match check.kind: - case "too_long": - return BlockedChange(check, rows, _too_long_detail(check, types, values, limit_bytes)) - case "duplicates": - return BlockedChange(check, rows, "duplicated value groups") - case "nulls": - return BlockedChange(check, rows, "NULL") - case "any_row": - return BlockedChange( - check, rows, "rows, and a NOT NULL column with no DEFAULT cannot be added to a table with rows" - ) - case "orphans": - return BlockedChange(check, rows, f'with no matching row in "{check.referenced_table}"') - case "check_violations": - return BlockedChange(check, rows, f"failing CHECK ({check.expression})") + return BlockedChange(check, rows, _too_long_detail(check, types, values, limit_bytes)) def find_blockers( @@ -571,19 +371,7 @@ def find_blockers( def _remedy(check: Check) -> str: joined: Final = ", ".join(check.columns) - match check.kind: - case "too_long": - return f'Shorten or hash the long "{check.table}" ({joined}) values so every row\'s index entry fits.' - case "duplicates": - return f'Remove or merge the duplicate "{check.table}" ({joined}) values.' - case "nulls": - return f'Fill the NULL "{check.table}"."{joined}" values, or give the column a default first.' - case "any_row": - return "Give the new column a DEFAULT in the migration, or add it nullable and fill it afterwards." - case "orphans": - return f'Delete or repoint the "{check.table}" ({joined}) values with no "{check.referenced_table}" row.' - case "check_violations": - return f'Fix the "{check.table}" rows that fail the expression before adding the constraint.' + return f'Shorten or hash the long "{check.table}" ({joined}) values so every row\'s index entry fits.' def _describe(finding: Finding) -> str: @@ -607,30 +395,4 @@ def _describe(finding: Finding) -> str: def report(findings: tuple[Finding, ...]) -> str: """The operator-facing refusal: one block per finding, then what to do next.""" lines: Final = "\n".join(_describe(finding) for finding in findings) - return ( - f"{REFUSAL}\n{lines}\n" - "No schema change was made. Fix the rows and start again, or run `litellm --check_migrations_against_existing_data` with the new image to check again." - ) - - -def pending_change_checks( - connection: "psycopg.Connection[tuple[object, ...]]", - schema: str, - migrations_dir: Path | None, - indexes: tuple[tuple[str, str, tuple[str, ...]], ...], -) -> tuple[Check, ...]: - """Every check the pending migrations and the request-log indexes need. `migrations_dir` - is None when the schema is pushed rather than migrated, so only the indexes are checked.""" - migration_checks: Final = ( - tuple( - check - for name, migration_sql in pending_migrations(connection, schema, migrations_dir) - for check in checks_for_migration(name, migration_sql) - ) # comprehension-ok: flattening the per-migration checks - if migrations_dir is not None - else () - ) - index_checks: Final = tuple( - check for table, name, columns in indexes for check in checks_for_index(table, name, columns) - ) # comprehension-ok: flattening the per-index checks - return migration_checks + index_checks + return f"{REFUSAL}\n{lines}\nNo index was built. Fix the rows and the next index build picks it up." diff --git a/litellm-proxy-extras/litellm_proxy_extras/migration_dry_run.py b/litellm-proxy-extras/litellm_proxy_extras/migration_dry_run.py new file mode 100644 index 00000000000..c24742ce145 --- /dev/null +++ b/litellm-proxy-extras/litellm_proxy_extras/migration_dry_run.py @@ -0,0 +1,184 @@ +"""`litellm --check_migrations_against_existing_data`: run the pending migrations for real, each +inside its own transaction that is always rolled back, so Postgres itself says whether this +database's rows and schema take them. Each migration holds the locks the real upgrade will, +and releases them before the next one runs.""" + +from __future__ import annotations + +import re +import time +from collections.abc import Callable +from dataclasses import dataclass +from functools import reduce +from pathlib import Path +from typing import TYPE_CHECKING, Final + +from litellm_proxy_extras.migration_data_checks import pending_migrations + +if TYPE_CHECKING: + import psycopg + +_CONCURRENTLY: Final = re.compile( + r"\b((?:CREATE\s+(?:UNIQUE\s+)?INDEX|DROP\s+INDEX|REINDEX\s+\w+)\s+)CONCURRENTLY\s+", re.IGNORECASE +) +UNDEFINED_SQLSTATES: Final = frozenset({"42703", "42P01", "42704", "42883"}) + +Migration = tuple[str, str] + + +@dataclass(frozen=True, slots=True) +class Applied: + name: str + seconds: float + replayed: tuple[str, ...] = () + + +@dataclass(frozen=True, slots=True) +class Rejected: + name: str + error: str + seconds: float + + +@dataclass(frozen=True, slots=True) +class Baseline: + """A database with tables but no `_prisma_migrations`: the boot baselines it instead of replaying + history, so the dry run asks the same question the boot does. `refusal` is None when the boot + will baseline, otherwise its reason (a custom migration folder, or the schema diff).""" + + refusal: str | None + + @property + def matches(self) -> bool: + return self.refusal is None + + +@dataclass(frozen=True, slots=True) +class DryRun: + applied: tuple[Applied, ...] + rejected: Rejected | None + baseline: Baseline | None = None + + @property + def ok(self) -> bool: + return self.rejected is None and (self.baseline is None or self.baseline.matches) + + +def transactional(migration_sql: str) -> str: + """The migration as it can run inside a transaction: `CONCURRENTLY` index builds become plain ones.""" + return _CONCURRENTLY.sub(r"\1", migration_sql) + + +def _attempt( + connection: psycopg.Connection[tuple[object, ...]], + schema: str, + chain: tuple[Migration, ...], + clock: Callable[[], float], +) -> tuple[float, psycopg.Error | None]: + import psycopg + from psycopg import sql + + started: Final = clock() + try: + with connection.transaction(force_rollback=True): + connection.execute(sql.SQL("SET LOCAL search_path TO {}").format(sql.Identifier(schema))) + for _, migration_sql in chain: + connection.execute(transactional(migration_sql)) + except psycopg.Error as exc: + return clock() - started, exc + return clock() - started, None + + +def _run( + connection: psycopg.Connection[tuple[object, ...]], + schema: str, + pending: tuple[Migration, ...], + position: int, + done: DryRun, + clock: Callable[[], float], +) -> DryRun: + """One migration in its own rolled-back transaction. A migration that fails because something + an earlier pending migration creates does not exist yet is rerun with those replayed in front.""" + if done.rejected is not None: + return done + name: Final = pending[position][0] + seconds, error = _attempt(connection, schema, (pending[position],), clock) + if error is None: + return DryRun(done.applied + (Applied(name, seconds),), None, done.baseline) + if error.sqlstate not in UNDEFINED_SQLSTATES or position == 0: + return DryRun(done.applied, Rejected(name, str(error).strip(), seconds), done.baseline) + earlier: Final = pending[:position] + chained_seconds, chained_error = _attempt(connection, schema, earlier + (pending[position],), clock) + if chained_error is None: + replayed: Final = tuple(earlier_name for earlier_name, _ in earlier) + return DryRun(done.applied + (Applied(name, chained_seconds, replayed),), None, done.baseline) + return DryRun(done.applied, Rejected(name, str(chained_error).strip(), chained_seconds), done.baseline) + + +def _has_ledger(connection: psycopg.Connection[tuple[object, ...]], schema: str) -> bool: + from psycopg import sql + + row: Final = connection.execute( + "SELECT to_regclass(%s)", (sql.Identifier(schema, "_prisma_migrations").as_string(connection),) + ).fetchone() + return row is not None and row[0] is not None + + +def _has_litellm_tables(connection: psycopg.Connection[tuple[object, ...]], schema: str) -> bool: + row: Final = connection.execute( + "SELECT 1 FROM information_schema.tables WHERE table_schema = %s AND table_name LIKE 'LiteLLM\\_%%' LIMIT 1", + (schema,), + ).fetchone() + return row is not None + + +def dry_run( + connection: psycopg.Connection[tuple[object, ...]], + schema: str, + migrations_dir: Path, + baseline_check: Callable[[], str | None] | None = None, + clock: Callable[[], float] = time.monotonic, +) -> DryRun: + """Apply every pending migration in order, each in its own transaction that is rolled back + before the next one starts. The first migration Postgres rejects ends the run with its error. + `baseline_check` (None when the boot would baseline, otherwise why it refuses) stands in for + the replay on a database that has tables but no migration history, which the boot baselines.""" + if baseline_check is not None and not _has_ledger(connection, schema) and _has_litellm_tables(connection, schema): + return DryRun((), None, Baseline(baseline_check())) + pending: Final = pending_migrations(connection, schema, migrations_dir) + return reduce( + lambda done, position: _run(connection, schema, pending, position, done, clock), + range(len(pending)), + DryRun((), None), + ) + + +def _timing(applied: Applied) -> str: + replay: Final = f" (depends on {', '.join(applied.replayed)}, replayed in front of it)" if applied.replayed else "" + return f"\n {applied.name}: {applied.seconds:.1f} s{replay}" + + +def report(result: DryRun) -> str: + """What the operator reads: the per-migration time the real upgrade will hold its locks, + or the migration Postgres rejected with Postgres's own words. Nothing was applied either way.""" + if result.baseline is not None: + if result.baseline.matches: + return ( + "No migration history and the schema matches this build, " + "the proxy baselines it at boot, nothing to dry-run." + ) + return f"No migration history and the proxy cannot baseline it at boot: {result.baseline.refusal}" + timings: Final = "".join(_timing(applied) for applied in result.applied) + if result.rejected is None: + count: Final = len(result.applied) + if count == 0: + return "No pending migrations, this database is on this version's schema." + return ( + f"{count} pending migration(s) applied one at a time in a dry run, each rolled back before the next " + f"so its locks are released, nothing changed:{timings}" + ) + return ( + f"Migration {result.rejected.name} fails on this database after {result.rejected.seconds:.1f} s, " + f"nothing was applied.\n{result.rejected.error}" + + (f"\nApplied before it in the dry run:{timings}" if timings else "") + ) diff --git a/litellm-proxy-extras/litellm_proxy_extras/migration_recovery.py b/litellm-proxy-extras/litellm_proxy_extras/migration_recovery.py index 5a55b35b255..1c16eb5e6b7 100644 --- a/litellm-proxy-extras/litellm_proxy_extras/migration_recovery.py +++ b/litellm-proxy-extras/litellm_proxy_extras/migration_recovery.py @@ -86,6 +86,18 @@ def migration_files(directory: Path) -> tuple[tuple[str, str], ...]: ) +def has_packaged_history(migrations_dir: Path) -> bool: + """Whether `migrations_dir` carries exactly the migration files and schema.prisma this package ships, + the only history the boot baselines an existing database onto.""" + packaged_dir: Final = Path(__file__).parent + migrations: Final = migration_files(migrations_dir) + return ( + bool(migrations) + and migrations == migration_files(packaged_dir) + and (migrations_dir / "schema.prisma").read_bytes() == (packaged_dir / "schema.prisma").read_bytes() + ) + + def baseline_current_schema( coordinator: MigrationCoordinator, schema: str, @@ -95,14 +107,9 @@ def baseline_current_schema( ) -> None: from psycopg import sql - packaged_dir: Final = Path(__file__).parent - migrations: Final = migration_files(migrations_dir) - if ( - not migrations - or migrations != migration_files(packaged_dir) - or (migrations_dir / "schema.prisma").read_bytes() != (packaged_dir / "schema.prisma").read_bytes() - ): + if not has_packaged_history(migrations_dir): raise RuntimeError("Cannot automatically baseline an existing database with custom migration history.") + migrations: Final = migration_files(migrations_dir) coordinator.acquire_prisma_lock() existing: Final = coordinator.connection.execute( diff --git a/litellm-proxy-extras/litellm_proxy_extras/utils.py b/litellm-proxy-extras/litellm_proxy_extras/utils.py index 528bf7def20..8b19576d86e 100644 --- a/litellm-proxy-extras/litellm_proxy_extras/utils.py +++ b/litellm-proxy-extras/litellm_proxy_extras/utils.py @@ -11,18 +11,13 @@ from collections.abc import Callable from contextlib import AbstractContextManager, nullcontext from dataclasses import dataclass, replace from pathlib import Path -from typing import TYPE_CHECKING, Final, Optional, Protocol, Union +from typing import TYPE_CHECKING, Final, Protocol from litellm_proxy_extras import prisma_toolchain from litellm_proxy_extras._logging import logger from litellm_proxy_extras._redaction import _redact_credentials -from litellm_proxy_extras.migration_data_checks import ( - Finding, - find_blockers, - migration_data_check_timeout_seconds, - pending_change_checks, - report, -) +from litellm_proxy_extras.migration_dry_run import dry_run +from litellm_proxy_extras.migration_dry_run import report as dry_run_report from litellm_proxy_extras.migration_lock import held_migration_lock from litellm_proxy_extras.prisma_toolchain import ( PRISMA_COMMAND_TIMEOUT_ENV_VAR, @@ -36,7 +31,6 @@ from litellm_proxy_extras.replica_identity import ( apply_replica_identity_full, ) from litellm_proxy_extras.request_log_indexes import ( - REQUEST_LOG_INDEXES, ensure_request_log_indexes, filter_request_log_index_diff, ) @@ -52,7 +46,7 @@ class LensCheckConnect(Protocol): ) -> "psycopg.Connection[tuple[object, ...]]": ... -def str_to_bool(value: Optional[str]) -> bool: +def str_to_bool(value: str | None) -> bool: if value is None: return False return value.lower() in ("true", "1", "t", "y", "yes") @@ -169,7 +163,7 @@ def _without_sql_comments(statement: str) -> str: ).strip() -def _without_spend_logs_pk_clauses(statement: str) -> Optional[str]: +def _without_spend_logs_pk_clauses(statement: str) -> str | None: prefix_match = _SPEND_LOGS_ALTER_RE.match(statement) if not prefix_match: return statement @@ -217,7 +211,7 @@ def _max_migration_timestamp(names) -> int: return max(_migration_timestamp(n) for n in names) -def _redacted_command(command: object) -> Union[str, tuple[str, ...], list[str]]: +def _redacted_command(command: object) -> str | tuple[str, ...] | list[str]: if isinstance(command, tuple): return tuple(_redact_credentials(str(argument)) for argument in command) if isinstance(command, list): @@ -404,7 +398,7 @@ class ProxyExtrasDBManager: return tuple(row) if row is not None else () @staticmethod - def _failed_migration_logs(migration_name: str, started_at: str) -> Optional[str]: + def _failed_migration_logs(migration_name: str, started_at: str) -> str | None: row: Final = ProxyExtrasDBManager._read_migration_ledger( "SELECT logs FROM {} WHERE migration_name = %s AND started_at = %s::timestamptz " "AND finished_at IS NULL AND rolled_back_at IS NULL", @@ -721,10 +715,10 @@ class ProxyExtrasDBManager: return row is not None @staticmethod - def _prisma_schema_param(url: str) -> Optional[str]: + def _prisma_schema_param(url: str) -> str | None: """The `schema` query param Prisma uses to pick its target schema, or None when the URL does not set one.""" - from urllib.parse import urlparse, parse_qsl + from urllib.parse import parse_qsl, urlparse return next( (v for k, v in parse_qsl(urlparse(url).query) if k == "schema"), @@ -1424,7 +1418,6 @@ class ProxyExtrasDBManager: bool: True if setup was successful, False otherwise """ ensure_prisma_toolchain(prisma_command=_get_prisma_command(), prisma_env=_get_prisma_env()) - ProxyExtrasDBManager.raise_if_rows_block_pending_changes(use_migrate=use_migrate) migrated = ProxyExtrasDBManager._run_migrations(use_migrate=use_migrate, use_v2_resolver=use_v2_resolver) if not migrated: return False @@ -1433,46 +1426,69 @@ class ProxyExtrasDBManager: return True @staticmethod - def migration_data_check_findings(use_migrate: bool = True) -> tuple[Finding, ...]: - """What in the existing rows would make a pending migration or request-log index - fail, read with count queries only: no write, no table lock. Empty when nothing - blocks, when no database is configured, or when it cannot be reached, so the - migration step that follows reports the connection problem the way it does today.""" + def _baseline_refusal(prisma_dir: Path) -> str | None: + """Why `baseline_current_schema` would refuse this database at boot, None when it would baseline: + the migrations folder must be the packaged one, then `prisma migrate diff` must find the live + database and this build's schema.prisma identical.""" + from litellm_proxy_extras.migration_lock import migration_environment + from litellm_proxy_extras.migration_recovery import has_packaged_history + + if not has_packaged_history(prisma_dir): + return ( + f"{prisma_dir} is a custom migration history (LITELLM_MIGRATION_DIR), it differs from the " + "migrations this build ships, and the proxy only baselines onto the packaged history." + ) + try: + prisma_toolchain.run_prisma( + ( + _get_prisma_command(), + "migrate", + "diff", + "--from-schema-datasource", + str(prisma_dir / "schema.prisma"), + "--to-schema-datamodel", + str(prisma_dir / "schema.prisma"), + "--exit-code", + ), + timeout=prisma_toolchain.prisma_command_timeout(), + env=migration_environment(_get_prisma_env()), + ) + except subprocess.CalledProcessError as exc: + diff: Final = str(exc.stdout or exc.stderr or exc).strip() + return f"the schema differs from this build. prisma migrate diff:\n{diff}" + except subprocess.TimeoutExpired as exc: + return f"prisma migrate diff did not finish within {exc.timeout} s" + return None + + @staticmethod + def migration_dry_run_summary(use_migrate: bool = True) -> tuple[bool, str]: + """`--check_migrations_against_existing_data`: the pending migrations applied for real on + DIRECT_URL or DATABASE_URL inside one transaction that is rolled back, so Postgres reports + what the upgrade would hit. (True, the per-migration timings) when they all apply, + (False, Postgres's own error) when one does not.""" import psycopg + if not use_migrate: + return (True, "The schema is pushed rather than migrated, there is nothing to dry-run.") database_url: Final = os.environ.get("DATABASE_URL") if not database_url: - return () + return (False, "DATABASE_URL is not set.") direct_url: Final = ProxyExtrasDBManager._strip_prisma_query_params( os.environ.get("DIRECT_URL") or database_url ) schema: Final = ProxyExtrasDBManager._prisma_schema_param(database_url) or "public" - migrations_dir: Final = Path(ProxyExtrasDBManager._get_prisma_dir()) / "migrations" if use_migrate else None - indexes: Final = tuple((index.table, index.name, index.columns) for index in REQUEST_LOG_INDEXES) + prisma_dir: Final = Path(ProxyExtrasDBManager._get_prisma_dir()) try: with psycopg.connect(direct_url, connect_timeout=10, autocommit=True) as connection: - checks: Final = pending_change_checks(connection, schema, migrations_dir, indexes) - return find_blockers(connection, schema, checks, migration_data_check_timeout_seconds()) + result: Final = dry_run( + connection, + schema, + prisma_dir / "migrations", + baseline_check=lambda: ProxyExtrasDBManager._baseline_refusal(prisma_dir), + ) except psycopg.OperationalError as exc: - logger.warning( - "Could not reach the database for the pre-migration row check: %s", _redact_credentials(str(exc)) - ) - return () - - @staticmethod - def migration_data_check_report(use_migrate: bool = True) -> Optional[str]: - """The operator-facing refusal for `migration_data_check_findings`, None when nothing blocks.""" - findings: Final = ProxyExtrasDBManager.migration_data_check_findings(use_migrate=use_migrate) - return report(findings) if findings else None - - @staticmethod - def raise_if_rows_block_pending_changes(use_migrate: bool) -> None: - """Refuse to run migrations that the existing rows would make fail, before any is - applied. The RuntimeError carries the report and reaches the operator through the - proxy's "Database migration cannot proceed" exit.""" - refusal: Final = ProxyExtrasDBManager.migration_data_check_report(use_migrate=use_migrate) - if refusal is not None: - raise RuntimeError(refusal) + return (False, f"Could not reach the database: {_redact_credentials(str(exc))}") + return (result.ok, dry_run_report(result)) @staticmethod def build_request_log_indexes(build: Callable[[str, str], bool] = ensure_request_log_indexes) -> bool: @@ -1724,5 +1740,4 @@ class ProxyExtrasDBManager: time.sleep(random.randrange(5, 15)) finally: os.chdir(original_dir) - pass return False diff --git a/litellm/proxy/db/prisma_client.py b/litellm/proxy/db/prisma_client.py index 3d783d8aba1..be5de040fdf 100644 --- a/litellm/proxy/db/prisma_client.py +++ b/litellm/proxy/db/prisma_client.py @@ -1127,16 +1127,16 @@ class PrismaManager: return ProxyExtrasDBManager.build_request_log_indexes() @staticmethod - def migration_data_check_report(use_migrate: bool = True) -> str | None: - """The read-only pre-migration row check on its own, for `--check_migrations_against_existing_data`: the - refusal the migration step would print, or None when the pending schema changes - can be applied to the rows the database has.""" + def migration_dry_run_summary(use_migrate: bool = True) -> tuple[bool, str]: + """`--check_migrations_against_existing_data` on its own: the pending migrations applied inside a + transaction that is rolled back, as (whether they all applied, what the operator reads).""" from litellm_proxy_extras.utils import ( ProxyExtrasDBManager, # pyright: ignore[reportUnknownVariableType] # litellm-proxy-extras is an optional install ) + from pydantic import TypeAdapter - refusal: Final[object] = ProxyExtrasDBManager.migration_data_check_report(use_migrate=use_migrate) # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType] # litellm-proxy-extras is an optional install - return refusal if isinstance(refusal, str) else None + summary: Final[object] = ProxyExtrasDBManager.migration_dry_run_summary(use_migrate=use_migrate) # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType] # litellm-proxy-extras is an optional install + return TypeAdapter(tuple[bool, str]).validate_python(summary) @staticmethod def start_request_log_index_build() -> None: diff --git a/litellm/proxy/proxy_cli.py b/litellm/proxy/proxy_cli.py index 0d8ab98a0b1..f71d1960e24 100644 --- a/litellm/proxy/proxy_cli.py +++ b/litellm/proxy/proxy_cli.py @@ -276,18 +276,19 @@ class ProxyInitializationHelpers: @staticmethod def run_migration_data_check(use_migrate: bool) -> int: - """`--check_migrations_against_existing_data`: the pre-migration row check against DATABASE_URL, nothing applied. - 0 when the pending schema changes fit the rows, 2 with the refusal when they do not.""" + """`--check_migrations_against_existing_data`: the pending migrations applied to DATABASE_URL inside one + transaction that is rolled back, so Postgres reports what the upgrade would hit. 0 when they all apply, + 2 with Postgres's error when one does not.""" if os.getenv("DATABASE_URL") is None: raise click.UsageError("--check_migrations_against_existing_data requires DATABASE_URL") from litellm.proxy.db.prisma_client import PrismaManager - refusal: Final[str | None] = PrismaManager.migration_data_check_report(use_migrate=use_migrate) # pyright: ignore[reportUnknownVariableType] # litellm-proxy-extras is an optional install - if refusal is None: - print("LiteLLM Proxy: the pending migrations and indexes can be applied to the rows this database has.") + ok, summary = PrismaManager.migration_dry_run_summary(use_migrate=use_migrate) + if ok: + print(f"LiteLLM Proxy: {summary}") return 0 print( - f"\033[1;31mLiteLLM Proxy: Database migration cannot proceed. {refusal}\033[0m", file=sys.stderr, flush=True + f"\033[1;31mLiteLLM Proxy: Database migration cannot proceed. {summary}\033[0m", file=sys.stderr, flush=True ) return 2 @@ -1059,8 +1060,9 @@ class ProxyInitializationHelpers: "--check_migrations_against_existing_data", is_flag=True, default=False, - help="Check, read-only, that the rows in DATABASE_URL can take this version's pending migrations and indexes, " - "print what would block them, then exit without applying anything. Exit code 2 when something blocks.", + help="Dry-run this version's pending migrations on DATABASE_URL: the real SQL inside one transaction that is " + "rolled back, so Postgres reports the row or schema that would make the upgrade fail, then exit. Holds the same " + "locks the upgrade will. Exit code 2 when a migration fails.", ) @click.option( "--keepalive_timeout", diff --git a/tests/integration/database/migrations/test_request_log_indexes.py b/tests/integration/database/migrations/test_request_log_indexes.py index f558c355c84..d06f0b13cff 100644 --- a/tests/integration/database/migrations/test_request_log_indexes.py +++ b/tests/integration/database/migrations/test_request_log_indexes.py @@ -16,7 +16,7 @@ from typing import Final import psycopg import pytest -from litellm_proxy_extras import migration_data_checks, request_log_indexes +from litellm_proxy_extras import migration_data_checks, migration_dry_run, request_log_indexes from litellm_proxy_extras.migration_lock import MIGRATION_LOCK_KEY, migration_lock from litellm_proxy_extras.migration_recovery import roll_back_failed_inert_migration from litellm_proxy_extras.request_log_indexes import ( @@ -967,33 +967,25 @@ _HASH_LEGACY_KEYS: Final = ( @requires_db @RESOLVERS @pytest.mark.parametrize("release", (API_KEY_INDEX_MIGRATION,), ids=("v1.102.1",)) -def test_the_migration_job_refuses_before_the_first_migration_when_spend_log_keys_are_too_long_for_the_index( +def test_the_migration_job_applies_every_migration_and_only_the_index_build_waits_for_too_long_keys( scratch_database: str, use_v2_resolver: bool ) -> None: + """Rows never stop an upgrade. The migrations go through, and the one request-log index + whose keys do not fit is left for the next build with the rows named in the log.""" with psycopg.connect(scratch_database, autocommit=True) as conn: _insert_spend_log_with_key(conn, "legacy-1", _legacy_bearer_token()) _insert_spend_log_with_key(conn, "legacy-2", _legacy_bearer_token()) _insert_spend_log_with_key(conn, "hashed-1", "a" * 64) - _insert_spend_log_with_key(conn, "repetitive-1", "z" * 3000) ledger_before: Final = _ledger(scratch_database) - with pytest.raises(RuntimeError) as refused: - ProxyExtrasDBManager.run_migration_job(use_migrate=True, use_v2_resolver=use_v2_resolver) - - refusal: Final = str(refused.value) - assert ( - '"LiteLLM_SpendLogs".api_key, startTime: 2 row(s) with an index entry over 2704 bytes, largest 3232 bytes ' - "(api_key up to 3207 bytes)" - ) in refusal, refusal - assert f'blocks: CREATE INDEX "{API_KEY_INDEX}"' in refusal, refusal - assert _ledger(scratch_database) == ledger_before, "a refused job must not apply a single migration" + assert ProxyExtrasDBManager.run_migration_job(use_migrate=True, use_v2_resolver=use_v2_resolver) is False + assert len(_ledger(scratch_database)) > len(ledger_before), "the rows must not hold back a single migration" assert API_KEY_INDEX not in _index_validity(scratch_database, "api_key_startTime_idx") with psycopg.connect(scratch_database, autocommit=True) as conn: conn.execute(_HASH_LEGACY_KEYS) assert ProxyExtrasDBManager.run_migration_job(use_migrate=True, use_v2_resolver=use_v2_resolver) is True assert _index_validity(scratch_database, "api_key_startTime_idx") == {API_KEY_INDEX: True} - assert len(_ledger(scratch_database)) > len(ledger_before) @requires_db @@ -1017,49 +1009,6 @@ def test_a_serving_proxy_index_build_that_meets_too_long_keys_reports_the_rows_i assert API_KEY_INDEX not in _index_validity(scratch_database, "api_key_startTime_idx") -_PROBE_STATEMENTS: Final = ( - 'CREATE UNIQUE INDEX "probe_b_key" ON "preflight_probe" ("b")', - 'ALTER TABLE "preflight_probe" ALTER COLUMN "b" SET NOT NULL', - 'ALTER TABLE "preflight_probe" ADD COLUMN "c" INTEGER NOT NULL', - 'ALTER TABLE "preflight_probe" ADD CONSTRAINT "probe_r_fkey" FOREIGN KEY ("r") REFERENCES "preflight_parent"("id")', - 'ALTER TABLE "preflight_probe" ADD CONSTRAINT "probe_a_check" CHECK ("a" <> \'bad\')', - 'CREATE INDEX "probe_a_idx" ON "preflight_probe" ("a")', -) - - -@requires_db -def test_each_check_kind_counts_exactly_the_rows_postgres_would_reject(scratch_database: str) -> None: - with psycopg.connect(scratch_database, autocommit=True) as conn: - conn.execute('CREATE TABLE "preflight_parent" ("id" INTEGER PRIMARY KEY)') - conn.execute('CREATE TABLE "preflight_probe" ("a" TEXT, "b" INTEGER, "r" INTEGER)') - conn.execute('INSERT INTO "preflight_parent" VALUES (1)') - conn.execute( - 'INSERT INTO "preflight_probe" VALUES (%s, 1, 1), (%s, 1, 1), (%s, NULL, 2), (%s, 2, NULL), (%s, 3, 1)', - ("ok", "ok", "bad", "bad", secrets.token_hex(1400)), - ) - checks: Final = tuple( - check - for statement in _PROBE_STATEMENTS - for check in migration_data_checks.checks_for_statement(statement, "probe") - ) - findings: Final = migration_data_checks.find_blockers(conn, "public", checks) - for statement in _PROBE_STATEMENTS: - with conn.transaction(force_rollback=True), pytest.raises(psycopg.Error): - conn.execute(statement) # pyright: ignore[reportArgumentType] # test SQL - - counted: Final = { - (f.check.kind, f.check.columns): f.rows for f in findings if isinstance(f, migration_data_checks.BlockedChange) - } - assert counted == { - ("duplicates", ("b",)): 1, - ("nulls", ("b",)): 1, - ("any_row", ("c",)): 5, - ("orphans", ("r",)): 1, - ("check_violations", ()): 2, - ("too_long", ("a",)): 1, - } - - @requires_db @pytest.mark.parametrize( ("index_columns", "s_value", "t_value", "longest_fitting_key", "described_as"), @@ -1098,7 +1047,7 @@ def test_the_entry_size_check_draws_the_line_exactly_where_postgres_does( ), ) findings: Final = migration_data_checks.find_blockers( - conn, "public", migration_data_checks.checks_for_statement(create_index.as_string(conn), "probe") + conn, "public", migration_data_checks.checks_for_index("entry_probe", "entry_probe_idx", index_columns) ) with pytest.raises(psycopg.errors.ProgramLimitExceeded, match="exceeds btree version 4 maximum 2704"): conn.execute(create_index) @@ -1122,7 +1071,7 @@ def test_a_value_postgres_compresses_under_the_limit_is_not_counted_while_random conn.execute('CREATE TABLE "compress_probe" ("k" TEXT)') conn.execute('INSERT INTO "compress_probe" ("k") VALUES (%s), (%s)', ("a" * 3000, secrets.token_hex(1500))) findings: Final = migration_data_checks.find_blockers( - conn, "public", migration_data_checks.checks_for_statement(create_index.as_string(conn), "probe") + conn, "public", migration_data_checks.checks_for_index("compress_probe", "compress_probe_idx", ("k",)) ) with pytest.raises(psycopg.errors.ProgramLimitExceeded, match="index row size 3016 exceeds"): conn.execute(create_index) @@ -1154,7 +1103,6 @@ def test_the_check_flags_exactly_the_compressed_keys_postgres_rejects_at_the_ent check and through Postgres one by one. The check must flag a key exactly when Postgres rejects it.""" noise: Final = "".join(hashlib.sha256(str(i).encode()).hexdigest() for i in range(40)) keys: Final = tuple(noise[:n] + "a" * 1500 for n in range(2400, 2500)) - create_index: Final = sql.SQL('CREATE INDEX "stored_probe_idx" ON "stored_probe" ("k")') with psycopg.connect(scratch_database, autocommit=True) as conn: conn.execute('CREATE TABLE "stored_probe" ("k" TEXT)') conn.execute('INSERT INTO "stored_probe" ("k") SELECT unnest(%s::text[])', (list(keys),)) @@ -1162,7 +1110,7 @@ def test_the_check_flags_exactly_the_compressed_keys_postgres_rejects_at_the_ent 'SELECT count(*) FROM "stored_probe" WHERE pg_column_size("k") < octet_length("k") AND octet_length("k") > 2704' ).fetchone() findings: Final = migration_data_checks.find_blockers( - conn, "public", migration_data_checks.checks_for_statement(create_index.as_string(conn), "probe") + conn, "public", migration_data_checks.checks_for_index("stored_probe", "stored_probe_idx", ("k",)) ) rejected: Final = tuple(key for key in keys if _postgres_rejects_an_index_on(conn, key)) @@ -1176,14 +1124,176 @@ def test_the_check_flags_exactly_the_compressed_keys_postgres_rejects_at_the_ent @requires_db def test_a_scan_that_cannot_finish_in_time_is_reported_as_unverified_rather_than_passed(scratch_database: str) -> None: with psycopg.connect(scratch_database, autocommit=True) as conn: - conn.execute('CREATE TABLE "preflight_slow" ("a" INTEGER)') - conn.execute('INSERT INTO "preflight_slow" SELECT generate_series(1, 50)') - (check,) = migration_data_checks.checks_for_statement( - 'ALTER TABLE "preflight_slow" ADD CONSTRAINT "slow" CHECK (pg_sleep(0.1) IS NULL)', "probe" + conn.execute('CREATE TABLE "preflight_slow" ("k" TEXT)') + conn.execute( + "INSERT INTO \"preflight_slow\" SELECT md5(i::text) || repeat('a', 1500) FROM generate_series(1, 200000) AS i" ) - findings: Final = migration_data_checks.find_blockers(conn, "public", (check,), timeout_seconds=0.5) + (check,) = migration_data_checks.checks_for_index("preflight_slow", "preflight_slow_idx", ("k",)) + findings: Final = migration_data_checks.find_blockers(conn, "public", (check,), timeout_seconds=0.01) timeout_afterwards: Final = conn.execute("SHOW statement_timeout").fetchone() - assert findings == (migration_data_checks.CouldNotVerify(check, 0.5),) - assert timeout_afterwards == ("0",), "the preflight must hand the session back with the timeout it found" - assert "could not be checked within 0.5 s" in migration_data_checks.report(findings) + assert findings == (migration_data_checks.CouldNotVerify(check, 0.01),) + assert timeout_afterwards == ("0",), "the check must hand the session back with the timeout it found" + assert "could not be checked within 0.01 s" in migration_data_checks.report(findings) + + +CURRENT_RELEASE: Final = "99999999999999" +REQUIRE_TEAM_ALIAS: Final = "20991231000000_require_team_alias" + + +def _extras_with_scratch_migration(tmp_path: Path, name: str, migration_sql: str) -> Path: + """A copy of the shipped package dir with one more migration folder, the shape `LITELLM_MIGRATION_DIR` points at.""" + extras: Final = tmp_path / "extras" + shutil.copytree(PACKAGE / "migrations", extras / "migrations") + shutil.copy(PACKAGE / "schema.prisma", extras / "schema.prisma") + (extras / "migrations" / name).mkdir() + (extras / "migrations" / name / "migration.sql").write_text(migration_sql) + return extras + + +def _team_alias_is_nullable(database_url: str) -> bool: + with psycopg.connect(database_url) as conn: + row: Final = conn.execute( + "SELECT is_nullable FROM information_schema.columns " + "WHERE table_name = 'LiteLLM_TeamTable' AND column_name = 'team_alias'" + ).fetchone() + return row == ("YES",) + + +@requires_db +@pytest.mark.parametrize("release", (API_KEY_INDEX_MIGRATION,), ids=("v1.102.1",)) +def test_the_dry_run_applies_every_pending_migration_in_order_and_rolls_all_of_it_back(scratch_database: str) -> None: + ledger_before: Final = _ledger(scratch_database) + with psycopg.connect(scratch_database, autocommit=True) as conn: + pending: Final = migration_data_checks.pending_migrations(conn, "public", PACKAGE / "migrations") + result: Final = migration_dry_run.dry_run(conn, "public", PACKAGE / "migrations") + + assert result.rejected is None, result + assert tuple(applied.name for applied in result.applied) == tuple(name for name, _ in pending) + assert len(result.applied) > 1 + assert _ledger(scratch_database) == ledger_before, "a dry run must leave the ledger alone" + assert _index_validity(scratch_database, "api_key_startTime_idx") == {}, "a dry run must leave no index behind" + names: Final = tuple(name for name, _ in pending) + assert all(set(applied.replayed) <= set(names[: names.index(applied.name)]) for applied in result.applied), ( + "a migration is only ever replayed behind a later one that depends on it" + ) + assert "applied one at a time in a dry run" in migration_dry_run.report(result) + + +@requires_db +@pytest.mark.parametrize("release", (CURRENT_RELEASE,), ids=("current",)) +def test_the_dry_run_reports_postgres_own_error_for_the_row_a_migration_cannot_take( + scratch_database: str, tmp_path: Path +) -> None: + extras: Final = _extras_with_scratch_migration( + tmp_path, REQUIRE_TEAM_ALIAS, 'ALTER TABLE "LiteLLM_TeamTable" ALTER COLUMN "team_alias" SET NOT NULL;\n' + ) + ledger_before: Final = _ledger(scratch_database) + with psycopg.connect(scratch_database, autocommit=True) as conn: + conn.execute('INSERT INTO "LiteLLM_TeamTable" ("team_id") VALUES (%s)', ("no-alias",)) + result: Final = migration_dry_run.dry_run(conn, "public", extras / "migrations") + + assert result.rejected is not None, result + assert result.rejected.name == REQUIRE_TEAM_ALIAS + assert 'column "team_alias" of relation "LiteLLM_TeamTable" contains null values' in result.rejected.error + assert result.applied == () + assert _team_alias_is_nullable(scratch_database), "a rejected dry run must leave the schema as it was" + assert _ledger(scratch_database) == ledger_before + + +@requires_db +@pytest.mark.parametrize("release", (CURRENT_RELEASE,), ids=("current",)) +def test_the_cli_summary_turns_from_postgres_error_to_timings_once_the_rows_are_fixed( + scratch_database: str, tmp_path: Path, monkeypatch: pytest.MonkeyPatch +) -> None: + extras: Final = _extras_with_scratch_migration( + tmp_path, REQUIRE_TEAM_ALIAS, 'ALTER TABLE "LiteLLM_TeamTable" ALTER COLUMN "team_alias" SET NOT NULL;\n' + ) + monkeypatch.setenv("LITELLM_MIGRATION_DIR", str(extras)) + with psycopg.connect(scratch_database, autocommit=True) as conn: + conn.execute('INSERT INTO "LiteLLM_TeamTable" ("team_id") VALUES (%s)', ("no-alias",)) + + ok, summary = ProxyExtrasDBManager.migration_dry_run_summary(use_migrate=True) + assert ok is False + assert summary.startswith(f"Migration {REQUIRE_TEAM_ALIAS} fails on this database after "), summary + assert "contains null values" in summary + + with psycopg.connect(scratch_database, autocommit=True) as conn: + conn.execute('UPDATE "LiteLLM_TeamTable" SET "team_alias" = \'filled\' WHERE "team_alias" IS NULL') + ok_after, summary_after = ProxyExtrasDBManager.migration_dry_run_summary(use_migrate=True) + assert ok_after is True + assert summary_after.startswith("1 pending migration(s) applied one at a time in a dry run"), summary_after + assert f"\n {REQUIRE_TEAM_ALIAS}: " in summary_after + assert _team_alias_is_nullable(scratch_database) + + +@requires_db +@pytest.mark.parametrize("release", (CURRENT_RELEASE,), ids=("current",)) +def test_a_migration_that_depends_on_the_one_before_it_is_replayed_with_it_while_independent_ones_run_alone( + scratch_database: str, tmp_path: Path +) -> None: + extras: Final = _extras_with_scratch_migration( + tmp_path, "20991231000000_add_team_region", 'ALTER TABLE "LiteLLM_TeamTable" ADD COLUMN "region" TEXT;\n' + ) + (extras / "migrations" / "20991231000001_index_team_region").mkdir() + (extras / "migrations" / "20991231000001_index_team_region" / "migration.sql").write_text( + 'CREATE INDEX CONCURRENTLY "LiteLLM_TeamTable_region_idx" ON "LiteLLM_TeamTable"("region");\n' + ) + (extras / "migrations" / "20991231000002_index_team_alias").mkdir() + (extras / "migrations" / "20991231000002_index_team_alias" / "migration.sql").write_text( + 'CREATE INDEX "LiteLLM_TeamTable_alias_idx" ON "LiteLLM_TeamTable"("team_alias");\n' + ) + with psycopg.connect(scratch_database, autocommit=True) as conn: + result: Final = migration_dry_run.dry_run(conn, "public", extras / "migrations") + region: Final = conn.execute( + "SELECT 1 FROM information_schema.columns WHERE table_name = 'LiteLLM_TeamTable' AND column_name = 'region'" + ).fetchone() + + assert result.rejected is None, result + assert tuple((applied.name, applied.replayed) for applied in result.applied) == ( + ("20991231000000_add_team_region", ()), + ("20991231000001_index_team_region", ("20991231000000_add_team_region",)), + ("20991231000002_index_team_alias", ()), + ) + assert region is None, "every transaction must have been rolled back" + with psycopg.connect(scratch_database) as conn: + probe_indexes: Final = conn.execute( + "SELECT indexname FROM pg_indexes " + "WHERE indexname IN ('LiteLLM_TeamTable_region_idx', 'LiteLLM_TeamTable_alias_idx')" + ).fetchall() + assert probe_indexes == [], probe_indexes + + +@requires_db +@pytest.mark.parametrize("release", (CURRENT_RELEASE,), ids=("current",)) +def test_a_database_with_tables_but_no_ledger_is_judged_by_the_schema_diff_instead_of_replaying_history( + scratch_database: str, tmp_path: Path +) -> None: + with psycopg.connect(scratch_database, autocommit=True) as conn: + conn.execute("DROP TABLE _prisma_migrations") + matching: Final = migration_dry_run.dry_run(conn, "public", PACKAGE / "migrations", baseline_check=lambda: None) + differing: Final = migration_dry_run.dry_run( + conn, + "public", + PACKAGE / "migrations", + baseline_check=lambda: ( + "the schema differs from this build. prisma migrate diff:\nALTER TABLE x ADD COLUMN y TEXT;" + ), + ) + custom: Final = migration_dry_run.dry_run( + conn, + "public", + PACKAGE / "migrations", + baseline_check=lambda: ProxyExtrasDBManager._baseline_refusal( + _extras_with_scratch_migration( + tmp_path, "20991231000000_custom", 'ALTER TABLE "LiteLLM_TeamTable" ADD COLUMN "c" TEXT;\n' + ) + ), + ) + + assert matching.ok is True and matching.applied == () + assert "nothing to dry-run" in migration_dry_run.report(matching) + assert differing.ok is False + assert "ALTER TABLE x ADD COLUMN y TEXT;" in migration_dry_run.report(differing) + assert custom.ok is False + assert "custom migration history" in migration_dry_run.report(custom) diff --git a/tests/unit/litellm_proxy_extras/test_migration_data_checks.py b/tests/unit/litellm_proxy_extras/test_migration_data_checks.py index d4490927e29..37377498d89 100644 --- a/tests/unit/litellm_proxy_extras/test_migration_data_checks.py +++ b/tests/unit/litellm_proxy_extras/test_migration_data_checks.py @@ -12,98 +12,15 @@ from litellm_proxy_extras.migration_data_checks import ( _count_query, btree_max_entry_bytes, checks_for_index, - checks_for_migration, - checks_for_statement, migration_data_check_timeout_seconds, report, - statements, ) -MIGRATIONS: Final = Path(__file__).resolve().parents[3] / "litellm-proxy-extras" / "litellm_proxy_extras" / "migrations" -API_KEY_INDEX_AS_SHIPPED_IN_V1_103_0: Final = 'CREATE INDEX IF NOT EXISTS "LiteLLM_SpendLogs_api_key_startTime_idx" ON "LiteLLM_SpendLogs"("api_key", "startTime")' - def _shape(checks: tuple[Check, ...]) -> tuple[tuple[str, str, tuple[str, ...]], ...]: return tuple((check.kind, check.table, check.columns) for check in checks) -@pytest.mark.parametrize( - ("statement", "expected"), - ( - ( - API_KEY_INDEX_AS_SHIPPED_IN_V1_103_0, - (("too_long", "LiteLLM_SpendLogs", ("api_key", "startTime")),), - ), - ( - 'CREATE UNIQUE INDEX CONCURRENTLY "T_a_b_key" ON "public"."T" USING btree ("a", "b" DESC)', - (("too_long", "T", ("a", "b")), ("duplicates", "T", ("a", "b"))), - ), - ('CREATE INDEX "T_a_idx" ON "T" USING gin ("a")', ()), - ('CREATE INDEX "T_a_idx" ON "T" (lower("a"))', ()), - ('ALTER TABLE "T" ALTER COLUMN "a" SET NOT NULL', (("nulls", "T", ("a",)),)), - ('ALTER TABLE "T" ADD COLUMN "a" TEXT NOT NULL', (("any_row", "T", ("a",)),)), - ('ALTER TABLE "T" ADD COLUMN "a" TEXT NOT NULL DEFAULT \'x\'', ()), - ('ALTER TABLE "T" ADD COLUMN "a" TEXT', ()), - ( - 'ALTER TABLE "T" ADD CONSTRAINT "T_r_fkey" FOREIGN KEY ("r") REFERENCES "R"("id") ON DELETE CASCADE', - (("orphans", "T", ("r",)),), - ), - ('ALTER TABLE "T" ADD CONSTRAINT "T_r_fkey" FOREIGN KEY ("r") REFERENCES "R"("id") NOT VALID', ()), - ('ALTER TABLE "T" ADD CONSTRAINT "T_a_key" UNIQUE ("a", "b")', (("duplicates", "T", ("a", "b")),)), - ( - 'ALTER TABLE "T" ADD CONSTRAINT "T_a_check" CHECK ("a" > 0 AND "b" IS NOT NULL)', - (("check_violations", "T", ()),), - ), - ('ALTER TABLE "T" ADD CONSTRAINT "T_a_check" CHECK ("a" > 0) NOT VALID', ()), - ( - 'ALTER TABLE "T" ADD COLUMN "c" TEXT, ALTER COLUMN "b" SET NOT NULL, ADD CONSTRAINT "u" UNIQUE ("c")', - (("nulls", "T", ("b",)), ("duplicates", "T", ("c",))), - ), - ('CREATE TABLE "T" ("id" TEXT NOT NULL, CONSTRAINT "T_pkey" PRIMARY KEY ("id"))', ()), - ('DROP INDEX IF EXISTS "T_a_idx"', ()), - ("SELECT 1", ()), - ), -) -def test_a_statement_gets_one_check_per_way_existing_rows_can_make_it_fail( - statement: str, expected: tuple[tuple[str, str, tuple[str, ...]], ...] -) -> None: - assert _shape(checks_for_statement(statement, "migration x")) == expected - - -def test_a_foreign_key_check_keeps_the_referenced_table_and_columns() -> None: - (check,) = checks_for_statement( - 'ALTER TABLE "T" ADD CONSTRAINT "fk" FOREIGN KEY ("r", "s") REFERENCES "R"("id", "sub")', "migration x" - ) - assert (check.referenced_table, check.referenced_columns) == ("R", ("id", "sub")) - - -def test_a_check_constraint_keeps_its_expression_for_the_count_query() -> None: - (check,) = checks_for_statement('ALTER TABLE "T" ADD CONSTRAINT "c" CHECK ("a" > 0)', "migration x") - assert check.expression == '"a" > 0' - - -def test_comments_and_blank_statements_are_dropped_before_matching() -> None: - migration_sql: Final = "-- CreateIndex\n" + API_KEY_INDEX_AS_SHIPPED_IN_V1_103_0 + ";\n\n-- done; really\n" - assert statements(migration_sql) == (API_KEY_INDEX_AS_SHIPPED_IN_V1_103_0,) - assert _shape(checks_for_migration("x", migration_sql)) == _shape( - checks_for_statement(API_KEY_INDEX_AS_SHIPPED_IN_V1_103_0, "x") - ) - - -def test_every_shipped_migration_is_readable_and_the_later_indexes_on_text_columns_get_a_check() -> None: - derived: Final = { - folder.name: checks_for_migration(folder.name, (folder / "migration.sql").read_text()) - for folder in MIGRATIONS.iterdir() - if folder.is_dir() - } - assert len(derived) > 200 - everywhere: Final = tuple(check for checks in derived.values() for check in checks) # comprehension-ok: flattening - assert any( - check.kind == "too_long" and check.table == "LiteLLM_SpendLogs" and "session_id" in check.columns - for check in everywhere - ) - - def test_a_request_log_index_gets_one_entry_size_check_over_all_its_columns() -> None: assert _shape( checks_for_index("LiteLLM_SpendLogs", "LiteLLM_SpendLogs_api_key_startTime_idx", ("api_key", "startTime")) @@ -142,14 +59,19 @@ def test_the_report_names_the_rows_the_statement_and_the_fix_and_says_nothing_wa assert 'blocks: CREATE INDEX "LiteLLM_SpendLogs_api_key_startTime_idx" (request-log index)' in text assert 'fix: Shorten or hash the long "LiteLLM_SpendLogs" (api_key, startTime) values' in text assert text.startswith("Existing rows would make a pending schema change fail:") - assert "No schema change was made" in text + assert "No index was built" in text def test_a_scan_that_timed_out_is_reported_as_unverified_with_the_knob_to_turn() -> None: - (check,) = checks_for_statement('ALTER TABLE "T" ALTER COLUMN "a" SET NOT NULL', "migration 20270101000000_x") + (check,) = checks_for_index( + "LiteLLM_SpendLogs", "LiteLLM_SpendLogs_api_key_startTime_idx", ("api_key", "startTime") + ) text: Final = report((CouldNotVerify(check, 600.0),)) - assert '"T".a: could not be checked within 600 s, refusing rather than guessing' in text - assert 'blocks: ALTER TABLE "T" ALTER COLUMN "a" SET NOT NULL (migration 20270101000000_x)' in text + assert ( + '"LiteLLM_SpendLogs".api_key, startTime: could not be checked within 600 s, refusing rather than guessing' + in text + ) + assert 'blocks: CREATE INDEX "LiteLLM_SpendLogs_api_key_startTime_idx" (request-log index)' in text assert MIGRATION_DATA_CHECK_TIMEOUT_ENV_VAR in text diff --git a/tests/unit/litellm_proxy_extras/test_migration_dry_run.py b/tests/unit/litellm_proxy_extras/test_migration_dry_run.py new file mode 100644 index 00000000000..75a7fea6c89 --- /dev/null +++ b/tests/unit/litellm_proxy_extras/test_migration_dry_run.py @@ -0,0 +1,80 @@ +from typing import Final + +import pytest +from litellm_proxy_extras.migration_dry_run import Applied, Baseline, DryRun, Rejected, report, transactional + +INDEX: Final = 'CREATE INDEX IF NOT EXISTS "a_idx" ON "T"("c")' + + +@pytest.mark.parametrize( + ("migration_sql", "expected"), + ( + ('CREATE INDEX CONCURRENTLY IF NOT EXISTS "a_idx" ON "T"("c")', INDEX), + ('create unique index concurrently "u" on "T"("d")', 'create unique index "u" on "T"("d")'), + ('DROP INDEX CONCURRENTLY "x"', 'DROP INDEX "x"'), + ('REINDEX INDEX CONCURRENTLY "y"', 'REINDEX INDEX "y"'), + (INDEX, INDEX), + ('ALTER TABLE "T" ALTER COLUMN "c" SET NOT NULL', 'ALTER TABLE "T" ALTER COLUMN "c" SET NOT NULL'), + ("-- built CONCURRENTLY by the index job\n" + INDEX, "-- built CONCURRENTLY by the index job\n" + INDEX), + ), + ids=("concurrently", "unique_lowercase", "drop", "reindex", "plain_index", "alter", "word_in_comment"), +) +def test_only_concurrent_index_statements_change_so_the_migration_can_run_in_a_transaction( + migration_sql: str, expected: str +) -> None: + assert transactional(migration_sql) == expected + + +def test_a_clean_dry_run_lists_each_migration_with_the_time_the_upgrade_will_hold_its_locks() -> None: + text: Final = report(DryRun((Applied("20270101000000_a", 0.04), Applied("20270102000000_b", 12.345)), None)) + assert text.startswith( + "2 pending migration(s) applied one at a time in a dry run, each rolled back before the next" + ) + assert "\n 20270101000000_a: 0.0 s\n 20270102000000_b: 12.3 s" in text + + +def test_no_pending_migration_is_said_plainly() -> None: + assert report(DryRun((), None)) == "No pending migrations, this database is on this version's schema." + + +def test_a_rejected_migration_is_reported_with_postgres_own_words_and_what_ran_before_it() -> None: + error: Final = 'column "team_alias" of relation "LiteLLM_TeamTable" contains null values' + text: Final = report(DryRun((Applied("20270101000000_a", 0.5),), Rejected("20270102000000_b", error, 3.25))) + assert text.startswith( + "Migration 20270102000000_b fails on this database after 3.2 s, nothing was applied.\n" + error + ) + assert text.endswith("Applied before it in the dry run:\n 20270101000000_a: 0.5 s") + + +def test_a_rejection_on_the_first_migration_lists_nothing_before_it() -> None: + text: Final = report(DryRun((), Rejected("20270101000000_a", 'syntax error at or near "NOPE"', 0.0))) + assert "Applied before it" not in text + assert text.endswith('syntax error at or near "NOPE"') + + +def test_a_migration_that_needed_the_one_before_it_says_which_were_replayed_in_front() -> None: + text: Final = report( + DryRun((Applied("20270101000000_a", 0.1), Applied("20270102000000_b", 0.4, ("20270101000000_a",))), None) + ) + assert text.endswith( + "\n 20270101000000_a: 0.1 s\n 20270102000000_b: 0.4 s (depends on 20270101000000_a, replayed in front of it)" + ) + + +def test_a_database_without_history_that_matches_this_build_has_nothing_to_dry_run() -> None: + result: Final = DryRun((), None, Baseline(None)) + assert result.ok is True + assert report(result) == ( + "No migration history and the schema matches this build, the proxy baselines it at boot, nothing to dry-run." + ) + + +def test_a_database_without_history_the_boot_refuses_to_baseline_reports_the_refusal() -> None: + result: Final = DryRun( + (), None, Baseline('the schema differs from this build. prisma migrate diff:\n-- AlterTable\nALTER TABLE "x"') + ) + assert result.ok is False + assert report(result) == ( + "No migration history and the proxy cannot baseline it at boot: " + 'the schema differs from this build. prisma migrate diff:\n-- AlterTable\nALTER TABLE "x"' + ) diff --git a/tests/unit/proxy/db/test_prisma_client.py b/tests/unit/proxy/db/test_prisma_client.py index 4b776bfbb1a..6c82628576f 100644 --- a/tests/unit/proxy/db/test_prisma_client.py +++ b/tests/unit/proxy/db/test_prisma_client.py @@ -586,3 +586,19 @@ async def test_a_query_issued_while_the_last_one_finishes_keeps_the_engine_drain await second await asyncio.wait_for(drained, timeout=1) assert engine.queries_finished == 2 + + +def test_migration_dry_run_summary_without_migrations_has_nothing_to_dry_run(unset_database_url): + """`prisma db push` applies no migration files, so the dry run reports that instead of connecting.""" + from litellm.proxy.db.prisma_client import PrismaManager + + assert PrismaManager.migration_dry_run_summary(use_migrate=False) == ( + True, + "The schema is pushed rather than migrated, there is nothing to dry-run.", + ) + + +def test_migration_dry_run_summary_needs_database_url(unset_database_url): + from litellm.proxy.db.prisma_client import PrismaManager + + assert PrismaManager.migration_dry_run_summary(use_migrate=True) == (False, "DATABASE_URL is not set.") diff --git a/tests/unit/proxy/test_proxy_cli.py b/tests/unit/proxy/test_proxy_cli.py index afefae0dbb1..1db334039ec 100644 --- a/tests/unit/proxy/test_proxy_cli.py +++ b/tests/unit/proxy/test_proxy_cli.py @@ -4393,3 +4393,20 @@ def test_granian_flag_notices_are_json_records_in_json_mode( "LiteLLM: --ciphers is not applied when using --run_granian.", ] assert [record["level"] for record in records] == ["INFO", "WARNING", "WARNING"] + + +def test_migration_data_check_exits_zero_when_there_is_nothing_to_dry_run(monkeypatch, capfd): + monkeypatch.setenv("DATABASE_URL", "postgresql://user:pass@db.example/litellm") + + assert ProxyInitializationHelpers.run_migration_data_check(use_migrate=False) == 0 + + out, err = capfd.readouterr() + assert out == "LiteLLM Proxy: The schema is pushed rather than migrated, there is nothing to dry-run.\n" + assert err == "" + + +def test_migration_data_check_requires_database_url(monkeypatch): + monkeypatch.delenv("DATABASE_URL", raising=False) + + with pytest.raises(click.UsageError, match="requires DATABASE_URL"): + ProxyInitializationHelpers.run_migration_data_check(use_migrate=True) From feda51453a3c468d04b608da4a9a972d3b365ce8 Mon Sep 17 00:00:00 2001 From: mrinal-berri Date: Thu, 29 Oct 2026 09:39:49 -0700 Subject: [PATCH 7/7] fix!: retain unattributed daily activity across reports and exports (#48925) * fix: retain unattributed daily activity breakdowns * fix: scope daily activity bucket drilldowns consistently * fix: preserve all additive daily activity metrics * test: cover unattributed reporting end to end * fix: preserve daily activity bucket identities across exports and key details * test: drop SQL-text assertions changed by daily activity bucket fix Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --------- Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../auto_router_endpoints.py | 2 +- .../common_daily_activity.py | 168 ++++++++++------- .../daily_activity_routes.py | 34 +--- .../daily_activity_scopes.py | 9 +- .../tag_management_endpoints.py | 8 +- .../management_endpoints/team_endpoints.py | 8 +- .../repositories/daily_activity_buckets.py | 39 ++++ .../repositories/daily_activity_repository.py | 9 +- litellm/repositories/daily_activity_sql.py | 60 +++--- .../routing/test_auto_router_management.py | 24 ++- ...st_daily_activity_aggregated_breakdowns.py | 178 ++++++++++++++++++ .../spend/test_daily_activity_routes.py | 14 ++ .../test_common_daily_activity.py | 143 ++++++++++++++ .../test_daily_activity_routes.py | 22 ++- .../test_tag_management_endpoints.py | 5 +- .../repositories/test_daily_activity_sql.py | 17 +- .../EntityUsage/TopKeyView.test.tsx | 26 ++- .../components/EntityUsage/TopKeyView.tsx | 6 +- .../UsagePage/keyActivityData.test.ts | 52 ++++- .../components/UsagePage/keyActivityData.ts | 10 + .../KeySavingsTab.integration.test.tsx | 12 ++ .../components/templates/KeySavingsTab.tsx | 3 +- 22 files changed, 695 insertions(+), 154 deletions(-) create mode 100644 litellm/repositories/daily_activity_buckets.py diff --git a/litellm/proxy/management_endpoints/auto_router_endpoints.py b/litellm/proxy/management_endpoints/auto_router_endpoints.py index 3dceab6b7f7..b79154bf509 100644 --- a/litellm/proxy/management_endpoints/auto_router_endpoints.py +++ b/litellm/proxy/management_endpoints/auto_router_endpoints.py @@ -865,7 +865,7 @@ async def _recorded_autorouter_savings( entity_id_field="user_id", entity_id=user_id, exclude_entity_ids=None, - api_key=api_key, + api_key=(api_key,) if api_key is not None else None, start_date=start_day, end_date=end_day, model=None, diff --git a/litellm/proxy/management_endpoints/common_daily_activity.py b/litellm/proxy/management_endpoints/common_daily_activity.py index 66f386b792d..98372b64015 100644 --- a/litellm/proxy/management_endpoints/common_daily_activity.py +++ b/litellm/proxy/management_endpoints/common_daily_activity.py @@ -23,6 +23,12 @@ from litellm.proxy.spend_tracking.key_metadata_recovery import ( ) from litellm.proxy.spend_tracking.ptu_feature_flag import is_ptu_cost_attribution_enabled from litellm.proxy.utils import PrismaClient +from litellm.repositories.daily_activity_buckets import ( + api_key_bucket, + model_bucket, + model_group_bucket, + raw_api_key_filter, +) from litellm.repositories.daily_activity_repository import DailyActivityRepository from litellm.types.proxy.management_endpoints.common_daily_activity import ( BreakdownMetrics, @@ -182,7 +188,9 @@ class _AggregatedSpendData(TypedDict): totals: ReadOnly[SpendMetrics] -def _key_metadata(api_key_metadata: Mapping[str, KeyMetadataRow], api_key: str) -> KeyMetadata: +def _key_metadata(api_key_metadata: Mapping[str, KeyMetadataRow], api_key: str | None) -> KeyMetadata: + if not api_key: + return KeyMetadata(key_alias="Unattributed", key_exists=False) meta: Final = api_key_metadata.get(api_key) return KeyMetadata( key_alias=meta.key_alias if meta is not None else None, @@ -302,13 +310,14 @@ def update_breakdown_metrics( parent bucket but never appear as an api_key row, and are kept out of the per-request provider breakdown.""" + api_key: Final = api_key_bucket(record.api_key) is_ptu_sentinel: Final = record.api_key == PTU_SENTINEL_API_KEY # A PTU sentinel row keys on the deployment id so a rename cannot move it, and carries # the operator-facing name in model_group. The breakdown key is rendered directly as a # label, so display the name; two deployments sharing one name merge here, which is # what the write path used to do by collapsing them into a single row. - model_key: Final = (record.model_group or record.model) if is_ptu_sentinel else record.model + model_key: Final = (record.model_group or record.model) if is_ptu_sentinel else model_bucket(record.model) # Update model breakdown if model_key and model_key not in breakdown.models: @@ -321,18 +330,18 @@ def update_breakdown_metrics( if not is_ptu_sentinel: # Update API key breakdown for this model - if record.api_key not in breakdown.models[model_key].api_key_breakdown: - breakdown.models[model_key].api_key_breakdown[record.api_key] = KeyMetricWithMetadata( + if api_key not in breakdown.models[model_key].api_key_breakdown: + breakdown.models[model_key].api_key_breakdown[api_key] = KeyMetricWithMetadata( metrics=SpendMetrics(), metadata=_key_metadata(api_key_metadata, record.api_key), ) - breakdown.models[model_key].api_key_breakdown[record.api_key].metrics = update_metrics( - breakdown.models[model_key].api_key_breakdown[record.api_key].metrics, + breakdown.models[model_key].api_key_breakdown[api_key].metrics = update_metrics( + breakdown.models[model_key].api_key_breakdown[api_key].metrics, record, ) # Update model group breakdown - model_group_key: Final = record.model_group or record.model + model_group_key: Final = model_group_bucket(record.model_group, record.model) if model_group_key and model_group_key not in breakdown.model_groups: breakdown.model_groups[model_group_key] = MetricWithMetadata( metrics=SpendMetrics(), @@ -345,13 +354,13 @@ def update_breakdown_metrics( if not is_ptu_sentinel: # Update API key breakdown for this model - if record.api_key not in breakdown.model_groups[model_group_key].api_key_breakdown: - breakdown.model_groups[model_group_key].api_key_breakdown[record.api_key] = KeyMetricWithMetadata( + if api_key not in breakdown.model_groups[model_group_key].api_key_breakdown: + breakdown.model_groups[model_group_key].api_key_breakdown[api_key] = KeyMetricWithMetadata( metrics=SpendMetrics(), metadata=_key_metadata(api_key_metadata, record.api_key), ) - breakdown.model_groups[model_group_key].api_key_breakdown[record.api_key].metrics = update_metrics( - breakdown.model_groups[model_group_key].api_key_breakdown[record.api_key].metrics, + breakdown.model_groups[model_group_key].api_key_breakdown[api_key].metrics = update_metrics( + breakdown.model_groups[model_group_key].api_key_breakdown[api_key].metrics, record, ) @@ -365,21 +374,20 @@ def update_breakdown_metrics( breakdown.mcp_servers[record.mcp_namespaced_tool_name].metrics, record ) - # Update API key breakdown for this MCP server - if record.api_key not in breakdown.mcp_servers[record.mcp_namespaced_tool_name].api_key_breakdown: - breakdown.mcp_servers[record.mcp_namespaced_tool_name].api_key_breakdown[record.api_key] = ( - KeyMetricWithMetadata( - metrics=SpendMetrics(), - metadata=_key_metadata(api_key_metadata, record.api_key), + if not is_ptu_sentinel: + # Update API key breakdown for this MCP server + if api_key not in breakdown.mcp_servers[record.mcp_namespaced_tool_name].api_key_breakdown: + breakdown.mcp_servers[record.mcp_namespaced_tool_name].api_key_breakdown[api_key] = ( + KeyMetricWithMetadata( + metrics=SpendMetrics(), + metadata=_key_metadata(api_key_metadata, record.api_key), + ) ) - ) - breakdown.mcp_servers[record.mcp_namespaced_tool_name].api_key_breakdown[ - record.api_key - ].metrics = update_metrics( - breakdown.mcp_servers[record.mcp_namespaced_tool_name].api_key_breakdown[record.api_key].metrics, - record, - ) + breakdown.mcp_servers[record.mcp_namespaced_tool_name].api_key_breakdown[api_key].metrics = update_metrics( + breakdown.mcp_servers[record.mcp_namespaced_tool_name].api_key_breakdown[api_key].metrics, + record, + ) if not is_ptu_sentinel: # Update provider breakdown @@ -392,13 +400,13 @@ def update_breakdown_metrics( breakdown.providers[provider].metrics = update_metrics(breakdown.providers[provider].metrics, record) # Update API key breakdown for this provider - if record.api_key not in breakdown.providers[provider].api_key_breakdown: - breakdown.providers[provider].api_key_breakdown[record.api_key] = KeyMetricWithMetadata( + if api_key not in breakdown.providers[provider].api_key_breakdown: + breakdown.providers[provider].api_key_breakdown[api_key] = KeyMetricWithMetadata( metrics=SpendMetrics(), metadata=_key_metadata(api_key_metadata, record.api_key), ) - breakdown.providers[provider].api_key_breakdown[record.api_key].metrics = update_metrics( - breakdown.providers[provider].api_key_breakdown[record.api_key].metrics, + breakdown.providers[provider].api_key_breakdown[api_key].metrics = update_metrics( + breakdown.providers[provider].api_key_breakdown[api_key].metrics, record, ) @@ -413,25 +421,26 @@ def update_breakdown_metrics( breakdown.endpoints[record.endpoint].metrics, record ) - # Update API key breakdown for this endpoint - if record.api_key not in breakdown.endpoints[record.endpoint].api_key_breakdown: - breakdown.endpoints[record.endpoint].api_key_breakdown[record.api_key] = KeyMetricWithMetadata( - metrics=SpendMetrics(), - metadata=_key_metadata(api_key_metadata, record.api_key), + if not is_ptu_sentinel: + # Update API key breakdown for this endpoint + if api_key not in breakdown.endpoints[record.endpoint].api_key_breakdown: + breakdown.endpoints[record.endpoint].api_key_breakdown[api_key] = KeyMetricWithMetadata( + metrics=SpendMetrics(), + metadata=_key_metadata(api_key_metadata, record.api_key), + ) + breakdown.endpoints[record.endpoint].api_key_breakdown[api_key].metrics = update_metrics( + breakdown.endpoints[record.endpoint].api_key_breakdown[api_key].metrics, + record, ) - breakdown.endpoints[record.endpoint].api_key_breakdown[record.api_key].metrics = update_metrics( - breakdown.endpoints[record.endpoint].api_key_breakdown[record.api_key].metrics, - record, - ) if not is_ptu_sentinel: # Update api key breakdown - if record.api_key not in breakdown.api_keys: - breakdown.api_keys[record.api_key] = KeyMetricWithMetadata( + if api_key not in breakdown.api_keys: + breakdown.api_keys[api_key] = KeyMetricWithMetadata( metrics=SpendMetrics(), metadata=_key_metadata(api_key_metadata, record.api_key), ) - breakdown.api_keys[record.api_key].metrics = update_metrics(breakdown.api_keys[record.api_key].metrics, record) + breakdown.api_keys[api_key].metrics = update_metrics(breakdown.api_keys[api_key].metrics, record) # Update entity-specific metrics if entity_id_field is provided if entity_id_field: @@ -445,13 +454,13 @@ def update_breakdown_metrics( if not is_ptu_sentinel: # Update API key breakdown for this entity - if record.api_key not in breakdown.entities[entity_value].api_key_breakdown: - breakdown.entities[entity_value].api_key_breakdown[record.api_key] = KeyMetricWithMetadata( + if api_key not in breakdown.entities[entity_value].api_key_breakdown: + breakdown.entities[entity_value].api_key_breakdown[api_key] = KeyMetricWithMetadata( metrics=SpendMetrics(), metadata=_key_metadata(api_key_metadata, record.api_key), ) - breakdown.entities[entity_value].api_key_breakdown[record.api_key].metrics = update_metrics( - breakdown.entities[entity_value].api_key_breakdown[record.api_key].metrics, + breakdown.entities[entity_value].api_key_breakdown[api_key].metrics = update_metrics( + breakdown.entities[entity_value].api_key_breakdown[api_key].metrics, record, ) @@ -566,7 +575,7 @@ def daily_activity_scope( entity_id_field: str, entity_id: str | list[str] | None, exclude_entity_ids: list[str] | None, - api_key: str | list[str] | None, + api_key: str | Sequence[str] | None, start_date: str, end_date: str, model: str | None, @@ -577,9 +586,7 @@ def daily_activity_scope( entity_ids: tuple[str, ...] | None = ( (entity_id,) if isinstance(entity_id, str) else tuple(entity_id) if entity_id is not None else None ) - api_keys: tuple[str, ...] | None = ( - None if api_key in (None, "") else (api_key,) if isinstance(api_key, str) else tuple(api_key) - ) + api_keys: Final = raw_api_key_filter(api_key) return DailyActivityScope( table=table_value, entity_id_field=entity_id_field, @@ -715,6 +722,29 @@ _GROUP_DATE_ENDPOINT: Final = 62 # 0b0111110 _GROUP_DATE_ENDPOINT_API_KEY: Final = 30 # 0b0011110 +def sum_spend_metrics(metrics: Sequence[SpendMetrics]) -> SpendMetrics: + return SpendMetrics( + spend=sum(metric.spend for metric in metrics), + flat_cost=sum(metric.flat_cost for metric in metrics), + ptu_hours=sum(metric.ptu_hours for metric in metrics), + prompt_tokens=sum(metric.prompt_tokens for metric in metrics), + completion_tokens=sum(metric.completion_tokens for metric in metrics), + cache_read_input_tokens=sum(metric.cache_read_input_tokens for metric in metrics), + cache_creation_input_tokens=sum(metric.cache_creation_input_tokens for metric in metrics), + compression_saved_tokens=sum(metric.compression_saved_tokens for metric in metrics), + compression_savings_spend=sum(metric.compression_savings_spend for metric in metrics), + prompt_caching_savings_spend=sum(metric.prompt_caching_savings_spend for metric in metrics), + gateway_injected_caching_savings_spend=sum(metric.gateway_injected_caching_savings_spend for metric in metrics), + autorouter_savings_spend=sum(metric.autorouter_savings_spend for metric in metrics), + total_tokens=sum(metric.total_tokens for metric in metrics), + successful_requests=sum(metric.successful_requests for metric in metrics), + failed_requests=sum(metric.failed_requests for metric in metrics), + api_requests=sum(metric.api_requests for metric in metrics), + total_response_time_ms=sum(metric.total_response_time_ms for metric in metrics), + timed_requests=sum(metric.timed_requests for metric in metrics), + ) + + def _record_to_spend_metrics(record: RollupMetricsRow) -> SpendMetrics: """Build a SpendMetrics directly from one already-aggregated rollup row. @@ -753,8 +783,7 @@ def _aggregate_grouping_sets_records_sync( Each row carries a `group_level` bitmask (from Postgres GROUPING()) that identifies which rollup level it belongs to. We dispatch the row's - pre-aggregated metrics straight into the matching bucket โ€” no per-row - summing in Python and no nested update_metrics calls. + pre-aggregated metrics into the matching bucket, combining normalized names. """ total_metrics = SpendMetrics() grouped_data: Final[dict[str, GroupedData]] = {} @@ -771,20 +800,23 @@ def _aggregate_grouping_sets_records_sync( if existing is None: target[key] = MetricWithMetadata(metrics=metrics, metadata={}) else: - existing.metrics = metrics + existing.metrics = sum_spend_metrics((existing.metrics, metrics)) def assign_api_key_breakdown( target: dict[str, MetricWithMetadata], parent_key: str, - api_key: str, + api_key: str | None, metrics: SpendMetrics, ) -> None: parent = target.get(parent_key) if parent is None: parent = MetricWithMetadata(metrics=SpendMetrics(), metadata={}) target[parent_key] = parent - parent.api_key_breakdown[api_key] = KeyMetricWithMetadata( - metrics=metrics, metadata=_key_metadata(api_key_metadata, api_key) + bucket_key: Final = api_key_bucket(api_key) + previous: Final = parent.api_key_breakdown.get(bucket_key) + parent.api_key_breakdown[bucket_key] = KeyMetricWithMetadata( + metrics=sum_spend_metrics((previous.metrics, metrics)) if previous is not None else metrics, + metadata=_key_metadata(api_key_metadata, api_key), ) for record in records: @@ -803,25 +835,23 @@ def _aggregate_grouping_sets_records_sync( breakdown = ensure_date(record.date)["breakdown"] if level == _GROUP_DATE_API_KEY: - if record.api_key and not is_ptu_sentinel: - breakdown.api_keys[record.api_key] = KeyMetricWithMetadata( + if not is_ptu_sentinel: + breakdown.api_keys[api_key_bucket(record.api_key)] = KeyMetricWithMetadata( metrics=metrics, metadata=_key_metadata(api_key_metadata, record.api_key), ) elif level == _GROUP_DATE_MODEL: - if record.model: - assign_metric_with_metadata(breakdown.models, record.model, metrics) + assign_metric_with_metadata(breakdown.models, model_bucket(record.model), metrics) elif level == _GROUP_DATE_MODEL_API_KEY: - if record.model and record.api_key and not is_ptu_sentinel: - assign_api_key_breakdown(breakdown.models, record.model, record.api_key, metrics) + if not is_ptu_sentinel: + assign_api_key_breakdown(breakdown.models, model_bucket(record.model), record.api_key, metrics) elif level == _GROUP_DATE_MODEL_GROUP: - if record.model_group: - assign_metric_with_metadata(breakdown.model_groups, record.model_group, metrics) + assign_metric_with_metadata(breakdown.model_groups, model_group_bucket(record.model_group, None), metrics) elif level == _GROUP_DATE_MODEL_GROUP_API_KEY: - if record.model_group and record.api_key and not is_ptu_sentinel: + if not is_ptu_sentinel: assign_api_key_breakdown( breakdown.model_groups, - record.model_group, + model_group_bucket(record.model_group, None), record.api_key, metrics, ) @@ -836,14 +866,14 @@ def _aggregate_grouping_sets_records_sync( provider = record.custom_llm_provider or "unknown" assign_metric_with_metadata(breakdown.providers, provider, provider_metrics) elif level == _GROUP_DATE_PROVIDER_API_KEY: - if record.api_key and not is_ptu_sentinel: + if not is_ptu_sentinel: provider = record.custom_llm_provider or "unknown" assign_api_key_breakdown(breakdown.providers, provider, record.api_key, metrics) elif level == _GROUP_DATE_MCP: if record.mcp_namespaced_tool_name: assign_metric_with_metadata(breakdown.mcp_servers, record.mcp_namespaced_tool_name, metrics) elif level == _GROUP_DATE_MCP_API_KEY: - if record.mcp_namespaced_tool_name and record.api_key: + if record.mcp_namespaced_tool_name and not is_ptu_sentinel: assign_api_key_breakdown( breakdown.mcp_servers, record.mcp_namespaced_tool_name, @@ -854,7 +884,7 @@ def _aggregate_grouping_sets_records_sync( if record.endpoint: assign_metric_with_metadata(breakdown.endpoints, record.endpoint, metrics) elif level == _GROUP_DATE_ENDPOINT_API_KEY: - if record.endpoint and record.api_key: + if record.endpoint and not is_ptu_sentinel: assign_api_key_breakdown(breakdown.endpoints, record.endpoint, record.api_key, metrics) results: Final = [ @@ -1013,8 +1043,8 @@ def _fold_entity_rollups_sync( metrics = _record_to_spend_metrics(row) if row.api_key_rolled: bucket.metrics = metrics - elif row.api_key and row.api_key != PTU_SENTINEL_API_KEY: - bucket.api_key_breakdown[row.api_key] = KeyMetricWithMetadata( + elif row.api_key != PTU_SENTINEL_API_KEY: + bucket.api_key_breakdown[api_key_bucket(row.api_key)] = KeyMetricWithMetadata( metrics=metrics, metadata=_key_metadata(api_key_metadata, row.api_key) ) diff --git a/litellm/proxy/management_endpoints/daily_activity_routes.py b/litellm/proxy/management_endpoints/daily_activity_routes.py index 79e515944ce..1a3445aae52 100644 --- a/litellm/proxy/management_endpoints/daily_activity_routes.py +++ b/litellm/proxy/management_endpoints/daily_activity_routes.py @@ -22,6 +22,7 @@ from litellm.proxy.management_endpoints.common_daily_activity import ( parse_canonical_date_range, raise_public, spend_logs_window, + sum_spend_metrics, ) from litellm.proxy.management_endpoints.daily_activity_scopes import ( AGENT_RESOLVER, @@ -38,6 +39,7 @@ from litellm.proxy.management_endpoints.ptu_consumption import single_team_id, w from litellm.proxy.management_endpoints.team_endpoints import aggregated_date_range_error from litellm.proxy.management_helpers.utils import management_endpoint_wrapper from litellm.proxy.utils import PrismaClient, get_prisma_client_or_throw +from litellm.repositories.daily_activity_buckets import api_key_bucket from litellm.repositories.daily_activity_repository import DailyActivityRepository from litellm.router import Router from litellm.types.proxy.management_endpoints.common_daily_activity import ( @@ -106,29 +108,9 @@ async def _resolved_scope( return result -def _sum_metrics(metrics: Sequence[SpendMetrics]) -> SpendMetrics: - return SpendMetrics( - spend=sum(metric.spend for metric in metrics), - flat_cost=sum(metric.flat_cost for metric in metrics), - prompt_tokens=sum(metric.prompt_tokens for metric in metrics), - completion_tokens=sum(metric.completion_tokens for metric in metrics), - cache_read_input_tokens=sum(metric.cache_read_input_tokens for metric in metrics), - cache_creation_input_tokens=sum(metric.cache_creation_input_tokens for metric in metrics), - compression_saved_tokens=sum(metric.compression_saved_tokens for metric in metrics), - compression_savings_spend=sum(metric.compression_savings_spend for metric in metrics), - prompt_caching_savings_spend=sum(metric.prompt_caching_savings_spend for metric in metrics), - gateway_injected_caching_savings_spend=sum(metric.gateway_injected_caching_savings_spend for metric in metrics), - autorouter_savings_spend=sum(metric.autorouter_savings_spend for metric in metrics), - total_tokens=sum(metric.total_tokens for metric in metrics), - successful_requests=sum(metric.successful_requests for metric in metrics), - failed_requests=sum(metric.failed_requests for metric in metrics), - api_requests=sum(metric.api_requests for metric in metrics), - total_response_time_ms=sum(metric.total_response_time_ms for metric in metrics), - timed_requests=sum(metric.timed_requests for metric in metrics), - ) - - def _key_metadata(api_key: str, metadata: Mapping[str, KeyMetadataRow]) -> KeyMetadata: + if not api_key: + return KeyMetadata(key_alias="Unattributed", key_exists=False) row: Final[KeyMetadataRow | None] = metadata.get(api_key) if row is None: return KeyMetadata() @@ -143,7 +125,7 @@ def _key_metadata(api_key: str, metadata: Mapping[str, KeyMetadataRow]) -> KeyMe def _key_activity_row(row: KeySpendRow, metadata: Mapping[str, KeyMetadataRow]) -> KeySpendActivityRow: return KeySpendActivityRow( - api_key=row.api_key, + api_key=api_key_bucket(row.api_key), metrics=KeySpendMetrics( spend=row.spend, prompt_tokens=row.prompt_tokens, @@ -168,7 +150,7 @@ async def _key_activity_rows( frozenset((resolved_scope.scope.start_date, resolved_scope.scope.end_date)) ) metadata: Final[Mapping[str, KeyMetadataRow]] = await repository.key_metadata( - frozenset(row.api_key for row in rows), + frozenset(row.api_key for row in rows if row.api_key), spend_window, ) return [_key_activity_row(row, metadata) for row in rows] @@ -231,11 +213,11 @@ def _fold_key_metrics(api_key: str, response: SpendAnalyticsPaginatedResponse) - metadata: Final[KeyMetadata] = next( day.breakdown.api_keys[api_key].metadata for day in response.results if api_key in day.breakdown.api_keys ) - return KeyActivityRow(api_key=api_key, metrics=_sum_metrics(metrics), metadata=metadata) + return KeyActivityRow(api_key=api_key, metrics=sum_spend_metrics(metrics), metadata=metadata) def _search_rows(keys: Sequence[str], response: SpendAnalyticsPaginatedResponse) -> list[KeyActivityRow]: - return [row for key in keys if (row := _fold_key_metrics(key, response)) is not None] + return [row for key in keys if (row := _fold_key_metrics(api_key_bucket(key), response)) is not None] def _csv_cell(value: object) -> object: diff --git a/litellm/proxy/management_endpoints/daily_activity_scopes.py b/litellm/proxy/management_endpoints/daily_activity_scopes.py index 77246f5dda6..ca1a96f9d84 100644 --- a/litellm/proxy/management_endpoints/daily_activity_scopes.py +++ b/litellm/proxy/management_endpoints/daily_activity_scopes.py @@ -17,6 +17,7 @@ from litellm.proxy.management_endpoints.organization_endpoints import resolve_or from litellm.proxy.management_endpoints.tag_management_endpoints import get_tag_daily_activity_api_key_filter from litellm.proxy.management_endpoints.team_endpoints import resolve_team_daily_activity_scope from litellm.proxy.utils import PrismaClient +from litellm.repositories.daily_activity_buckets import raw_api_key_filter from litellm.types.repositories.daily_activity import DailyActivityScope, DailyActivityTable @@ -77,13 +78,7 @@ def _build_scope( ) -> ResolvedScope: start_date: Final[str] = query.start_date or "" end_date: Final[str] = query.end_date or "" - api_keys: Final[tuple[str, ...] | None] = ( - None - if api_key_filter is None or api_key_filter == "" - else (api_key_filter,) - if isinstance(api_key_filter, str) - else tuple(api_key_filter) - ) + api_keys: Final = raw_api_key_filter(api_key_filter) return ResolvedScope( scope=DailyActivityScope( table=resolver.table, diff --git a/litellm/proxy/management_endpoints/tag_management_endpoints.py b/litellm/proxy/management_endpoints/tag_management_endpoints.py index c12a90ed6a2..ea7810a618c 100644 --- a/litellm/proxy/management_endpoints/tag_management_endpoints.py +++ b/litellm/proxy/management_endpoints/tag_management_endpoints.py @@ -31,6 +31,7 @@ from litellm.proxy.management_endpoints.common_daily_activity import ( get_daily_activity, ) from litellm.proxy.management_helpers.utils import handle_budget_for_entity +from litellm.repositories.daily_activity_buckets import api_key_from_bucket from litellm.repositories.model_repository import ModelRepository from litellm.repositories.table_repositories import ( DailyTagSpendRepository, @@ -198,16 +199,17 @@ async def get_tag_daily_activity_api_key_filter( user_api_key_dict: UserAPIKeyAuth, requested_api_key: str | None, ) -> str | list[str] | None: + selected_key: Final = api_key_from_bucket(requested_api_key) if requested_api_key else None user_role: Final = user_api_key_dict.user_role if user_api_key_has_admin_view(user_api_key_dict) or (user_role is None or not user_role.is_internal_user_role): - return requested_api_key + return [selected_key] if selected_key is not None else None scoped_api_keys: Final = await _get_internal_user_api_keys( prisma_client=prisma_client, user_api_key_dict=user_api_key_dict, ) - if requested_api_key is not None: - return requested_api_key if requested_api_key in scoped_api_keys else [] + if selected_key is not None: + return [selected_key] if selected_key in scoped_api_keys else [] return scoped_api_keys diff --git a/litellm/proxy/management_endpoints/team_endpoints.py b/litellm/proxy/management_endpoints/team_endpoints.py index 50b700e712d..b712ba42a6a 100644 --- a/litellm/proxy/management_endpoints/team_endpoints.py +++ b/litellm/proxy/management_endpoints/team_endpoints.py @@ -226,6 +226,7 @@ from litellm.proxy.management_helpers.utils import ( from litellm.proxy.utils import PrismaClient, ProxyLogging, handle_exception_on_proxy from litellm.repositories.budget_repository import BudgetRepository from litellm.repositories.chunked_in import delete_many_in, find_many_in +from litellm.repositories.daily_activity_buckets import api_key_from_bucket from litellm.repositories.organization_repository import OrganizationRepository from litellm.repositories.prisma_protocols import TableActions from litellm.repositories.table_repositories import ( @@ -7329,11 +7330,12 @@ async def resolve_team_daily_activity_scope( if not user_api_keys: user_api_keys = [] + selected_key: Final = api_key_from_bucket(api_key) if api_key else None final_api_key_filter: str | list[str] | None if user_api_keys is None: - final_api_key_filter = api_key - elif api_key: - final_api_key_filter = api_key if api_key in user_api_keys else [] + final_api_key_filter = [selected_key] if selected_key is not None else None + elif selected_key is not None: + final_api_key_filter = [selected_key] if selected_key in user_api_keys else [] else: final_api_key_filter = user_api_keys diff --git a/litellm/repositories/daily_activity_buckets.py b/litellm/repositories/daily_activity_buckets.py new file mode 100644 index 00000000000..ec6dd0f8c84 --- /dev/null +++ b/litellm/repositories/daily_activity_buckets.py @@ -0,0 +1,39 @@ +from collections.abc import Sequence +from typing import Final + +from litellm.constants import UNKNOWN_MODEL_SPEND_LOG_MODEL + +UNKNOWN_MODEL_GROUP: Final = "unknown-model-group" +UNATTRIBUTED_API_KEY: Final = "unattributed" +_RAW_API_KEY_PREFIX: Final = "raw:" + + +def model_bucket(model: str | None) -> str: + return model or UNKNOWN_MODEL_SPEND_LOG_MODEL + + +def model_group_bucket(model_group: str | None, model: str | None) -> str: + return model_group or model or UNKNOWN_MODEL_GROUP + + +def api_key_bucket(api_key: str | None) -> str: + if not api_key: + return UNATTRIBUTED_API_KEY + if api_key == UNATTRIBUTED_API_KEY or api_key.startswith(_RAW_API_KEY_PREFIX): + return _RAW_API_KEY_PREFIX + api_key + return api_key + + +def api_key_from_bucket(api_key: str) -> str: + if api_key == UNATTRIBUTED_API_KEY: + return "" + return api_key.removeprefix(_RAW_API_KEY_PREFIX) + + +def raw_api_key_filter(api_key: str | Sequence[str] | None) -> tuple[str, ...] | None: + """String filters are reporting IDs; sequences are raw keys resolved by authorization.""" + if api_key is None or api_key == "": + return None + if isinstance(api_key, str): + return (api_key_from_bucket(api_key),) + return tuple(api_key) diff --git a/litellm/repositories/daily_activity_repository.py b/litellm/repositories/daily_activity_repository.py index 4ca23e4e89e..321468c4976 100644 --- a/litellm/repositories/daily_activity_repository.py +++ b/litellm/repositories/daily_activity_repository.py @@ -10,6 +10,7 @@ from typing_extensions import assert_never from litellm import constants from litellm._logging import verbose_proxy_logger +from litellm.constants import UNKNOWN_MODEL_SPEND_LOG_MODEL from litellm.repositories.chunked_in import find_many_in from litellm.repositories.daily_activity_sql import ( ExportCursor, @@ -344,7 +345,13 @@ class DailyActivityRepository: "date": {"gte": adjusted_start, "lte": adjusted_end}, **({scope.entity_id_field: {"in": list(scope.entity_ids)}} if scope.entity_ids is not None else {}), **exclusion_filter, - **({"model": scope.model} if scope.model else {}), + **( + {"AND": [{"OR": [{"model": None}, {"model": ""}, {"model": scope.model}]}]} + if scope.model == UNKNOWN_MODEL_SPEND_LOG_MODEL + else {"model": scope.model} + if scope.model + else {} + ), **({"api_key": {"in": list(scope.api_keys)}} if scope.api_keys is not None else {}), } count, rows = await asyncio.gather( diff --git a/litellm/repositories/daily_activity_sql.py b/litellm/repositories/daily_activity_sql.py index 7ec9fa67a67..4bb52767de6 100644 --- a/litellm/repositories/daily_activity_sql.py +++ b/litellm/repositories/daily_activity_sql.py @@ -8,11 +8,18 @@ from typing import Final from typing_extensions import assert_never from litellm import constants -from litellm.constants import PTU_SENTINEL_API_KEY +from litellm.constants import PTU_SENTINEL_API_KEY, UNKNOWN_MODEL_SPEND_LOG_MODEL +from litellm.repositories.daily_activity_buckets import UNATTRIBUTED_API_KEY, UNKNOWN_MODEL_GROUP from litellm.types.repositories.daily_activity import DailyActivityScope, DailyActivityTable, ExportType _API_KEY_ROLLED_UP_BIT: Final = 32 -_MODEL_GROUP_EXPR: Final = "COALESCE(NULLIF(model_group, ''), model)" +_MODEL_EXPR: Final = f"COALESCE(NULLIF(model, ''), '{UNKNOWN_MODEL_SPEND_LOG_MODEL}')" +_MODEL_GROUP_EXPR: Final = f"COALESCE(NULLIF(model_group, ''), NULLIF(model, ''), '{UNKNOWN_MODEL_GROUP}')" + +_API_KEY_EXPR: Final = ( + f"CASE WHEN api_key = '' THEN '{UNATTRIBUTED_API_KEY}' " + f"WHEN api_key = '{UNATTRIBUTED_API_KEY}' OR api_key LIKE 'raw:%' THEN 'raw:' || api_key ELSE api_key END" +) @dataclass(frozen=True, slots=True) @@ -86,7 +93,11 @@ def build_where_clause(scope: DailyActivityScope, *, start_index: int = 1) -> tu if scope.exclude_entity_ids else () ), - *((f"model = ${model_index}",) if scope.model else ()), + *( + (f"{_MODEL_EXPR if scope.model == UNKNOWN_MODEL_SPEND_LOG_MODEL else 'model'} = ${model_index}",) + if scope.model + else () + ), *( ("FALSE",) if scope.api_keys == () @@ -143,7 +154,7 @@ def _top_api_keys_sql(pg_table: str, where_clause: str, *, sentinel_param: int, FROM "{pg_table}" WHERE {where_clause} AND api_key <> ${sentinel_param} GROUP BY api_key - ORDER BY SUM(spend::numeric) DESC, api_key + ORDER BY SUM(spend::numeric) DESC, {_API_KEY_EXPR} LIMIT ${limit_param} """ @@ -162,13 +173,13 @@ def build_aggregated_sql(scope: DailyActivityScope, *, api_key_limit: int) -> Sq (SELECT date, NULL::text AS api_key, - model, + {_MODEL_EXPR} AS model, {_MODEL_GROUP_EXPR} AS model_group, custom_llm_provider, mcp_namespaced_tool_name, endpoint, (GROUPING(date) << 6) | {_API_KEY_ROLLED_UP_BIT} - | GROUPING(model, {_MODEL_GROUP_EXPR}, + | GROUPING({_MODEL_EXPR}, {_MODEL_GROUP_EXPR}, custom_llm_provider, mcp_namespaced_tool_name, endpoint) AS group_level, NULL::bigint AS distinct_api_keys,{metric_select} @@ -176,7 +187,7 @@ def build_aggregated_sql(scope: DailyActivityScope, *, api_key_limit: int) -> Sq WHERE {where_clause} GROUP BY GROUPING SETS ( (date), - (date, model), + (date, {_MODEL_EXPR}), (date, {_MODEL_GROUP_EXPR}), (date, custom_llm_provider), (date, mcp_namespaced_tool_name), @@ -190,12 +201,12 @@ def build_aggregated_sql(scope: DailyActivityScope, *, api_key_limit: int) -> Sq SELECT date, api_key, - model, + {_MODEL_EXPR} AS model, {_MODEL_GROUP_EXPR} AS model_group, custom_llm_provider, mcp_namespaced_tool_name, endpoint, - GROUPING(date, api_key, model, {_MODEL_GROUP_EXPR}, + GROUPING(date, api_key, {_MODEL_EXPR}, {_MODEL_GROUP_EXPR}, custom_llm_provider, mcp_namespaced_tool_name, endpoint) AS group_level, MAX(top_api_keys.distinct_api_keys) AS distinct_api_keys,{metric_select} @@ -203,7 +214,7 @@ def build_aggregated_sql(scope: DailyActivityScope, *, api_key_limit: int) -> Sq WHERE {where_clause} GROUP BY GROUPING SETS ( (date, api_key), - (date, model, api_key), + (date, {_MODEL_EXPR}, api_key), (date, {_MODEL_GROUP_EXPR}, api_key), (date, custom_llm_provider, api_key), (date, mcp_namespaced_tool_name, api_key), @@ -313,7 +324,7 @@ def build_key_page_sql(scope: DailyActivityScope, *, offset: int, limit: int) -> FROM (SELECT 1) AS one LEFT JOIN LATERAL ( SELECT * FROM ranked - ORDER BY rank_spend DESC, api_key + ORDER BY rank_spend DESC, {_API_KEY_EXPR} LIMIT ${limit_param} OFFSET ${offset_param} ) AS page ON TRUE """ @@ -376,7 +387,7 @@ def build_key_search_sql(scope: DailyActivityScope, *, search: str, limit: int) WHERE {where_clause} AND api_key <> ${sentinel_param} AND ( - api_key ILIKE ${search_param} ESCAPE '\\' + {_API_KEY_EXPR} ILIKE ${search_param} ESCAPE '\\' OR api_key IN ( SELECT v.token FROM "LiteLLM_VerificationToken" v LEFT JOIN "LiteLLM_UserTable" u ON u.user_id = v.user_id @@ -392,7 +403,7 @@ def build_key_search_sql(scope: DailyActivityScope, *, search: str, limit: int) ) ) GROUP BY api_key - ORDER BY SUM(spend::numeric) DESC, api_key + ORDER BY SUM(spend::numeric) DESC, {_API_KEY_EXPR} LIMIT ${limit_param} """ return SqlQuery(sql=sql, params=(*where_params, f"%{escaped}%", PTU_SENTINEL_API_KEY, limit)) @@ -406,15 +417,13 @@ def build_model_top_keys_sql( model_param: Final = len(where_params) + 1 sentinel_param: Final = model_param + 1 limit_param: Final = sentinel_param + 1 - model_clause: Final = ( - f"COALESCE(NULLIF(model_group, ''), model) = ${model_param}" if by_model_group else f"model = ${model_param}" - ) + model_clause: Final = f"{_MODEL_GROUP_EXPR if by_model_group else _MODEL_EXPR} = ${model_param}" sql: Final = f""" SELECT api_key,{_key_spend_select()} FROM "{PRISMA_TO_PG_TABLE[scope.table]}" WHERE {where_clause} AND {model_clause} AND api_key <> ${sentinel_param} GROUP BY api_key - ORDER BY SUM(spend::numeric) DESC, api_key + ORDER BY SUM(spend::numeric) DESC, {_API_KEY_EXPR} LIMIT ${limit_param} """ return SqlQuery(sql=sql, params=(*where_params, model_group, PTU_SENTINEL_API_KEY, limit)) @@ -431,7 +440,7 @@ def build_cache_leakage_keys_sql(scope: DailyActivityScope, *, limit: int) -> Sq WHERE {where_clause} AND api_key <> ${sentinel_param} GROUP BY api_key HAVING SUM(prompt_tokens) - SUM(cache_read_input_tokens) > 0 - ORDER BY SUM(prompt_tokens) - SUM(cache_read_input_tokens) DESC, api_key + ORDER BY SUM(prompt_tokens) - SUM(cache_read_input_tokens) DESC, {_API_KEY_EXPR} LIMIT ${limit_param} """ return SqlQuery(sql=sql, params=(*where_params, PTU_SENTINEL_API_KEY, limit)) @@ -483,7 +492,7 @@ def build_export_sql( {alias_expression} AS entity_alias, {output_key} AS api_key, {user_fields}, - {"NULLIF(COALESCE(scoped.model, ''), '')" if export_type is ExportType.DAILY_WITH_MODELS else "NULL::text"} AS model, + {_MODEL_EXPR if export_type is ExportType.DAILY_WITH_MODELS else "NULL::text"} AS model, COALESCE(SUM(scoped.spend), 0)::float AS spend, {flat_cost} AS flat_cost, COALESCE(SUM(scoped.prompt_tokens), 0)::bigint AS prompt_tokens, @@ -521,25 +530,26 @@ def _export_grouping(export_type: ExportType) -> tuple[str, str, str, tuple[str, ) if export_type is ExportType.DAILY_WITH_KEYS: return ( - "scoped.api_key", - "NULLIF(scoped.api_key, '')", - "MAX(COALESCE(vt.key_alias, dvt.key_alias)) AS key_alias, " + _API_KEY_EXPR, + _API_KEY_EXPR, + "MAX(CASE WHEN scoped.api_key = '' THEN 'Unattributed' " + "ELSE COALESCE(vt.key_alias, dvt.key_alias) END) AS key_alias, " "MAX(COALESCE(vt.user_id, dvt.user_id)) AS user_id, MAX(u.user_email) AS user_email", ( - 'LEFT JOIN "LiteLLM_VerificationToken" vt ON vt.token = scoped.api_key', + "LEFT JOIN \"LiteLLM_VerificationToken\" vt ON vt.token = scoped.api_key AND scoped.api_key <> ''", """LEFT JOIN LATERAL ( SELECT key_alias, user_id FROM "LiteLLM_DeletedVerificationToken" WHERE token = scoped.api_key ORDER BY deleted_at DESC LIMIT 1 - ) dvt ON vt.token IS NULL""", + ) dvt ON vt.token IS NULL AND scoped.api_key <> ''""", 'LEFT JOIN "LiteLLM_UserTable" u ON u.user_id = COALESCE(vt.user_id, dvt.user_id)', ), ) if export_type is ExportType.DAILY_WITH_MODELS: return ( - "COALESCE(scoped.model, '')", + _MODEL_EXPR, "NULL::text", "NULL::text AS key_alias, NULL::text AS user_id, NULL::text AS user_email", (), diff --git a/tests/integration/routing/test_auto_router_management.py b/tests/integration/routing/test_auto_router_management.py index 5aed328ba7b..442dae4d689 100644 --- a/tests/integration/routing/test_auto_router_management.py +++ b/tests/integration/routing/test_auto_router_management.py @@ -9,7 +9,7 @@ import httpx import pytest import yaml from integration._support.client import JSON_OBJECT, Gateway, eventually, object_value, string_value, until_asserted -from integration._support.database import read_rows, scratch_database +from integration._support.database import read_rows, scratch_database, write_rows from integration._support.process import owned_proxy from integration._support.wire import Reply, Request, Wire, wire_server from pydantic import JsonValue, TypeAdapter @@ -639,3 +639,25 @@ def test_auto_router_session_tracks_key_scoped_spend_and_bounds_long_session_ids ), response.text assert tuple(request for request in complex_tier.drain() if request.method == "POST") == (), response.text assert tuple(request for request in baseline.drain() if request.method == "POST") == (), response.text + + +@pytest.mark.parametrize("api_key,expected", [("unattributed", 2.0), ("raw:unattributed", 4.0)]) +def test_benchmarks_preserve_literal_key_identity(gateway: Gateway, api_key: str, expected: float) -> None: + user_id: Final = f"benchmark-key-{uuid.uuid4().hex}" + for key, savings in (("", 3.0), ("unattributed", 2.0), ("raw:unattributed", 4.0)): + write_rows( + 'INSERT INTO "LiteLLM_DailyUserSpend" ' + "(id, user_id, date, api_key, model, autorouter_savings_spend, updated_at) " + "VALUES (%s, %s, '2026-09-09', %s, 'model', %s, now())", + (str(uuid.uuid4()), user_id, key, str(savings)), + ) + try: + response: Final = gateway.request( + "GET", + "/auto_router/benchmarks", + params={"start_date": "2026-09-09", "end_date": "2026-09-09", "user_id": user_id, "api_key": api_key}, + ) + assert response.status_code == 200, response.text + assert object_value(object_value(response.json())["totals"])["saved_spend"] == expected + finally: + write_rows('DELETE FROM "LiteLLM_DailyUserSpend" WHERE user_id = %s', (user_id,)) diff --git a/tests/integration/spend/test_daily_activity_aggregated_breakdowns.py b/tests/integration/spend/test_daily_activity_aggregated_breakdowns.py index 1c315a6e694..ff60cb9d174 100644 --- a/tests/integration/spend/test_daily_activity_aggregated_breakdowns.py +++ b/tests/integration/spend/test_daily_activity_aggregated_breakdowns.py @@ -1,3 +1,5 @@ +import csv +import io import uuid from collections.abc import Sequence from datetime import datetime, timedelta @@ -12,6 +14,7 @@ from tests.integration._support.database import write_rows _URL: Final = "/user/daily/activity/aggregated" _RESULTS: Final = TypeAdapter(list[dict[str, JsonValue]]) +_NUMBER: Final = TypeAdapter(float) def _unique_day() -> str: @@ -43,6 +46,181 @@ def _row_id() -> str: return f"agg-{uuid.uuid4().hex}" +@pytest.mark.parametrize("suffix", ["/aggregated", ""]) +def test_missing_attribution_reconciles_and_filters_without_key_collisions(gateway: Gateway, suffix: str) -> None: + day: Final = _unique_day() + rows: Final = ( + ("", None, None, 3.0), + ("", "", "", 0.0), + ("known-key", "unknown-model", None, -1.0), + ("unattributed", "known-model", "", 2.0), + ("raw:unattributed", "known-model", "named-group", 4.0), + ("known-key", "", "unknown-model-group", 5.0), + ) + _seed( + day, + [ + (_row_id(), "bucket-user", day, key, model, group, "", "tool", "/mcp", 10, spend, 1, 1) + for key, model, group, spend in rows + ], + ) + params: Final = {"start_date": day, "end_date": day} + url: Final = f"/user/daily/activity{suffix}" + try: + response: Final = gateway.request("GET", url, params=params) + assert response.status_code == 200, response.text + body: Final = object_value(response.json()) + results: Final = _RESULTS.validate_python(body["results"]) + assert len(results) == 1 + totals: Final = object_value(results[0]["metrics"]) + assert totals["spend"] == 13.0 + assert totals["api_requests"] == 6 + breakdown: Final = object_value(results[0]["breakdown"]) + for dimension in ("models", "model_groups", "api_keys", "providers", "mcp_servers", "endpoints"): + buckets: Final = _RESULTS.validate_python(list(object_value(breakdown[dimension]).values())) + for metric in ("spend", "api_requests", "prompt_tokens"): + assert ( + sum(_NUMBER.validate_python(object_value(bucket["metrics"])[metric]) for bucket in buckets) + == totals[metric] + ) + models: Final = object_value(breakdown["models"]) + assert set(models) == {"known-model", "unknown-model"} + assert object_value(object_value(models["unknown-model"])["metrics"])["api_requests"] == 4 + groups: Final = object_value(breakdown["model_groups"]) + assert set(groups) == {"unknown-model-group", "unknown-model", "known-model", "named-group"} + assert object_value(object_value(groups["unknown-model-group"])["metrics"])["spend"] == 8.0 + keys: Final = object_value(breakdown["api_keys"]) + assert set(keys) == {"unattributed", "raw:unattributed", "raw:raw:unattributed", "known-key"} + unknown_key: Final = object_value(keys["unattributed"]) + assert object_value(unknown_key["metadata"]) == { + "key_alias": "Unattributed", + "key_exists": False, + "team_id": None, + "user_id": None, + "user_email": None, + } + for dimension in ("models", "model_groups", "providers", "mcp_servers", "endpoints"): + for bucket in _RESULTS.validate_python(list(object_value(breakdown[dimension]).values())): + nested: Final = _RESULTS.validate_python(list(object_value(bucket["api_key_breakdown"]).values())) + assert ( + sum(_NUMBER.validate_python(object_value(key["metrics"])["spend"]) for key in nested) + == object_value(bucket["metrics"])["spend"] + ) + for key, spend, count in ( + ("unattributed", 3.0, 2), + ("raw:unattributed", 2.0, 1), + ("raw:raw:unattributed", 4.0, 1), + ): + filtered: Final = gateway.request("GET", url, params={**params, "api_key": key}) + assert filtered.status_code == 200, filtered.text + metadata: Final = object_value(object_value(filtered.json())["metadata"]) + assert metadata["total_spend"] == spend + assert metadata["total_api_requests"] == count + model_response: Final = gateway.request("GET", url, params={**params, "model": "unknown-model"}) + assert model_response.status_code == 200, model_response.text + model_metadata: Final = object_value(object_value(model_response.json())["metadata"]) + assert model_metadata["total_spend"] == 7.0 + assert model_metadata["total_api_requests"] == 4 + finally: + _clean(day) + + +@pytest.mark.parametrize( + ("route", "extra", "expected"), + [ + ("keys", {}, {"unattributed": 3.0, "raw:unattributed": 2.0, "known-key": 7.0}), + ("search", {"search": "Unattributed"}, {"unattributed": 3.0, "raw:unattributed": 2.0}), + ("model_top_keys", {"model_group": "unknown-model-group"}, {"unattributed": 3.0}), + ( + "model_top_keys", + {"model_group": "unknown-model", "by_model_group": "false"}, + {"unattributed": 3.0, "raw:unattributed": 2.0}, + ), + ], +) +def test_missing_attribution_key_drilldowns( + gateway: Gateway, + route: str, + extra: dict[str, str], + expected: dict[str, float], +) -> None: + day: Final = _unique_day() + _seed( + day, + [ + (_row_id(), "bucket-user", day, "", None, None, "", None, None, 0, 3.0, 1, 1), + (_row_id(), "bucket-user", day, "unattributed", "unknown-model", "", "", None, None, 0, 2.0, 1, 1), + (_row_id(), "bucket-user", day, "known-key", "known-model", "", "", None, None, 0, 7.0, 1, 1), + ], + ) + try: + response: Final = gateway.request( + "GET", f"{_URL}/{route}", params={"start_date": day, "end_date": day, **extra} + ) + assert response.status_code == 200, response.text + rows: Final = _RESULTS.validate_python(object_value(response.json())["api_keys"]) + assert {str(row["api_key"]): object_value(row["metrics"])["spend"] for row in rows} == expected + unattributed: Final = next(row for row in rows if row["api_key"] == "unattributed") + assert object_value(unattributed["metadata"])["key_alias"] == "Unattributed" + assert object_value(unattributed["metadata"])["key_exists"] is False + finally: + _clean(day) + + +@pytest.mark.parametrize("format", ["json", "csv"]) +@pytest.mark.parametrize( + ("export_type", "field", "expected"), + [ + ("daily_with_keys", "api_key", {"unattributed": 3.0, "raw:unattributed": 2.0, "raw:raw:unattributed": 4.0}), + ("daily_with_models", "model", {"unknown-model": 5.0, "known-model": 4.0}), + ], +) +def test_export_bucket_ids_round_trip_across_streaming_batches( + gateway: Gateway, format: str, export_type: str, field: str, expected: dict[str, float] +) -> None: + day: Final = _unique_day() + _seed( + day, + [ + (_row_id(), "export-user", day, "", None, None, "", None, None, 0, 3.0, 1, 1), + (_row_id(), "export-user", day, "", "", "", "", None, None, 0, 0.0, 1, 1), + (_row_id(), "export-user", day, "unattributed", "unknown-model", "", "", None, None, 0, 2.0, 1, 1), + (_row_id(), "export-user", day, "raw:unattributed", "known-model", "", "", None, None, 0, 4.0, 1, 1), + ], + ) + write_rows( + 'INSERT INTO "LiteLLM_DailyUserSpend" (id, user_id, date, api_key, model, spend, api_requests, updated_at) ' + "SELECT %s || i, 'export-user', %s, 'raw:a-key-' || i, 'a-model-' || i, 1, 1, now() " + "FROM generate_series(1, 1001) i", + (str(uuid.uuid4()), day), + ) + params: Final = {"start_date": day, "end_date": day} + try: + response: Final = gateway.request( + "GET", "/user/daily/activity/export", params={**params, "export_type": export_type, "format": format} + ) + assert response.status_code == 200, response.text + rows: Final = ( + _RESULTS.validate_python(response.json()) + if format == "json" + else _RESULTS.validate_python(list(csv.DictReader(io.StringIO(response.text)))) + ) + buckets: Final = {str(row[field]): _NUMBER.validate_python(row["spend"]) for row in rows} + filler: Final = "raw:raw:a-key-" if field == "api_key" else "a-model-" + assert buckets == {**{f"{filler}{i}": 1.0 for i in range(1, 1002)}, **expected} + assert len(rows) == len(buckets) + assert sum(buckets.values()) == 1010.0 + for bucket, spend in expected.items(): + filtered: Final = _activity(gateway, day, **{field: bucket}) + assert object_value(filtered["metadata"])["total_spend"] == spend + if field == "api_key": + unattributed: Final = next(row for row in rows if row["api_key"] == "unattributed") + assert unattributed["key_alias"] == "Unattributed" + assert unattributed["user_id"] in (None, "") + finally: + _clean(day) + + def _ranked_key_rows(day: str, count: int) -> list[tuple[object, ...]]: return [ ( diff --git a/tests/integration/spend/test_daily_activity_routes.py b/tests/integration/spend/test_daily_activity_routes.py index 520c8ccc6d8..9bfe4bb50ce 100644 --- a/tests/integration/spend/test_daily_activity_routes.py +++ b/tests/integration/spend/test_daily_activity_routes.py @@ -451,9 +451,23 @@ def test_tag_daily_activity_scopes_internal_user_keys_and_filters( seeded_row(TAG_SPEND, "tag", "beta", digest_one_b, DAY), seeded_row(TAG_SPEND, "tag", "gamma", digest_one_a, DAY), seeded_row(TAG_SPEND, "tag", "alpha", digest_two, DAY), + seeded_row(TAG_SPEND, "tag", "alpha", "", DAY), replace(standard_model_row, model="gpt-4o"), ) with daily_rows(rows): + for suffix in ("", "/aggregated"): + bucket_params: Final = {"start_date": DAY, "end_date": DAY, "tags": "alpha", "api_key": "unattributed"} + bucket_url: Final = f"/tag/daily/activity{suffix}" + denied_bucket: Final = gateway.request("GET", bucket_url, params=bucket_params, key=key_one_a) + admin_bucket: Final = gateway.request("GET", bucket_url, params=bucket_params) + assert denied_bucket.status_code == 200, denied_bucket.text + assert admin_bucket.status_code == 200, admin_bucket.text + denied_body: Final = SpendAnalyticsPaginatedResponse.model_validate_json(denied_bucket.content) + allowed_body: Final = SpendAnalyticsPaginatedResponse.model_validate_json(admin_bucket.content) + assert denied_body.results == [] + assert denied_body.metadata.total_api_requests == 0 + assert allowed_body.metadata.total_api_requests == 1 + assert set(allowed_body.results[0].breakdown.api_keys) == {"unattributed"} scoped: Final = _tag_activity( gateway, key=key_one_a, diff --git a/tests/unit/proxy/management_endpoints/test_common_daily_activity.py b/tests/unit/proxy/management_endpoints/test_common_daily_activity.py index 01d2f7007a1..c80fa3572fd 100644 --- a/tests/unit/proxy/management_endpoints/test_common_daily_activity.py +++ b/tests/unit/proxy/management_endpoints/test_common_daily_activity.py @@ -2551,3 +2551,146 @@ async def test_get_daily_activity_rejects_non_canonical_dates_before_querying(st assert error.value.detail == {"error": "start_date and end_date must be valid YYYY-MM-DD dates"} mock_table.count.assert_not_awaited() mock_table.find_many.assert_not_awaited() + + +@pytest.mark.parametrize("missing", [None, ""]) +def test_missing_dimensions_preserve_all_paginated_metrics(missing: str | None) -> None: + records: Final = ( + _daily_user_spend_record(user_id="user", api_key="", spend=-0.25, model=missing, model_group=missing), + _daily_user_spend_record(user_id="user", api_key="", spend=0.0, model=missing, model_group=missing), + ) + result: Final = common_daily_activity_module._aggregate_spend_records_sync( + records=records, api_key_metadata={}, entity_id_field="user_id", entity_metadata_field=None + ) + day: Final = result["results"][0] + assert day.metrics == result["totals"] + for bucket in ( + day.breakdown.models["unknown-model"], + day.breakdown.model_groups["unknown-model-group"], + day.breakdown.providers["openai"], + day.breakdown.endpoints["/chat/completions"], + day.breakdown.entities["user"], + ): + assert bucket.metrics == day.metrics + assert bucket.api_key_breakdown == day.breakdown.api_keys + key: Final = day.breakdown.api_keys["unattributed"] + assert key.metrics == day.metrics + assert key.metadata.model_dump(exclude_none=True) == {"key_alias": "Unattributed", "key_exists": False} + + +def test_normalized_rollups_add_colliding_model_metrics_and_keep_key_identities() -> None: + records: Final = ( + _grouping_row(common_daily_activity_module._GROUP_DATE_MODEL, model=None, spend=1.0), + _grouping_row(common_daily_activity_module._GROUP_DATE_MODEL, model="", spend=2.0), + _grouping_row(common_daily_activity_module._GROUP_DATE_MODEL, model="unknown-model", spend=3.0), + _grouping_row(common_daily_activity_module._GROUP_DATE_MODEL_API_KEY, model=None, api_key="", spend=1.0), + _grouping_row(common_daily_activity_module._GROUP_DATE_MODEL_API_KEY, model="", api_key="", spend=2.0), + _grouping_row( + common_daily_activity_module._GROUP_DATE_MODEL_API_KEY, + model="unknown-model", + api_key="unattributed", + spend=3.0, + ), + _grouping_row(common_daily_activity_module._GROUP_DATE_API_KEY, api_key="", spend=3.0), + _grouping_row(common_daily_activity_module._GROUP_DATE_API_KEY, api_key="unattributed", spend=3.0), + _grouping_row(common_daily_activity_module._GROUP_DATE_API_KEY, api_key="raw:unattributed", spend=4.0), + ) + result: Final = common_daily_activity_module._aggregate_grouping_sets_records_sync( + records=records, api_key_metadata={} + ) + breakdown: Final = result["results"][0].breakdown + assert breakdown.models["unknown-model"].metrics.spend == 6.0 + assert {key: value.metrics.spend for key, value in breakdown.models["unknown-model"].api_key_breakdown.items()} == { + "unattributed": 3.0, + "raw:unattributed": 3.0, + } + assert {key: value.metrics.spend for key, value in breakdown.api_keys.items()} == { + "unattributed": 3.0, + "raw:unattributed": 3.0, + "raw:raw:unattributed": 4.0, + } + + +@pytest.mark.parametrize( + ("model", "group", "expected_group"), + [ + (None, None, "unknown-model-group"), + ("", "", "unknown-model-group"), + (None, "configured-group", "configured-group"), + ("provider-model", "", "provider-model"), + ("unknown-model", "", "unknown-model"), + (" ", "", " "), + ], +) +def test_model_group_uses_raw_model_before_unknown_group( + model: str | None, group: str | None, expected_group: str +) -> None: + result: Final = common_daily_activity_module._aggregate_spend_records_sync( + records=(_daily_user_spend_record(user_id="user", api_key="key", spend=1.0, model=model, model_group=group),), + api_key_metadata={}, + entity_id_field=None, + entity_metadata_field=None, + ) + assert tuple(result["results"][0].breakdown.model_groups) == (expected_group,) + + +@pytest.mark.parametrize( + ("raw_key", "bucket"), + [ + ("", "unattributed"), + ("real-hash", "real-hash"), + ("unattributed", "raw:unattributed"), + ("raw:unattributed", "raw:raw:unattributed"), + ("raw:real-hash", "raw:raw:real-hash"), + ], +) +def test_reporting_key_filters_round_trip_without_losing_scope(raw_key: str, bucket: str) -> None: + from litellm.repositories.daily_activity_buckets import api_key_bucket + + assert api_key_bucket(raw_key) == bucket + scope: Final = daily_activity_scope( + "litellm_dailyuserspend", "user_id", "user", None, bucket, "2024-01-01", "2024-01-01", None, None + ) + assert scope.api_keys == (raw_key,) + assert scope.entity_ids == ("user",) + + +@pytest.mark.parametrize(("raw_key", "bucket_key"), [("", "unattributed"), ("unattributed", "raw:unattributed")]) +def test_mcp_nested_key_metrics_use_reporting_identity(raw_key: str, bucket_key: str) -> None: + record: Final = SimpleNamespace( + **{ + **vars(_daily_user_spend_record(user_id="user", api_key=raw_key, spend=0.25, model=None, model_group=None)), + "mcp_namespaced_tool_name": "deleted-server/tool", + } + ) + result: Final = common_daily_activity_module._aggregate_spend_records_sync( + records=(record,), api_key_metadata={}, entity_id_field=None, entity_metadata_field=None + ) + breakdown: Final = result["results"][0].breakdown + assert breakdown.mcp_servers["deleted-server/tool"].api_key_breakdown == breakdown.api_keys + assert breakdown.mcp_servers["deleted-server/tool"].api_key_breakdown[bucket_key].metrics.spend == 0.25 + + +def test_ptu_rows_do_not_create_endpoint_or_mcp_key_buckets(ptu_cost_attribution_enabled: None) -> None: + from litellm.constants import PTU_SENTINEL_API_KEY + from litellm.types.proxy.management_endpoints.common_daily_activity import BreakdownMetrics + + record: Final = SimpleNamespace( + **{ + **vars(_spend_record(PTU_SENTINEL_API_KEY, ptu_flat_cost=2.0)), + "mcp_namespaced_tool_name": "server/tool", + "endpoint": "/mcp", + } + ) + breakdown: Final = common_daily_activity_module.update_breakdown_metrics(BreakdownMetrics(), record, {}, {}, {}) + assert breakdown.api_keys == {} + assert breakdown.endpoints["/mcp"].api_key_breakdown == {} + assert breakdown.mcp_servers["server/tool"].api_key_breakdown == {} + assert breakdown.endpoints["/mcp"].metrics.flat_cost == 2.0 + assert breakdown.mcp_servers["server/tool"].metrics.flat_cost == 2.0 + + +def test_metric_merges_preserve_every_additive_field() -> None: + metrics: Final = SpendMetrics.model_validate(dict.fromkeys(SpendMetrics.model_fields, 2)) + merged: Final = common_daily_activity_module.sum_spend_metrics((SpendMetrics(), metrics, metrics)) + assert merged.model_dump() == dict.fromkeys(SpendMetrics.model_fields, 4) diff --git a/tests/unit/proxy/management_endpoints/test_daily_activity_routes.py b/tests/unit/proxy/management_endpoints/test_daily_activity_routes.py index 2a273034a97..e487c00f25d 100644 --- a/tests/unit/proxy/management_endpoints/test_daily_activity_routes.py +++ b/tests/unit/proxy/management_endpoints/test_daily_activity_routes.py @@ -628,7 +628,9 @@ def test_aggregated_routes_return_scoped_results( assert len(body["results"]) == 2, response.text -def _ptu_sized_router(model_group: str, shares: Mapping[str, int] = MappingProxyType({"team-a": 30, "team-b": 20})) -> Router: +def _ptu_sized_router( + model_group: str, shares: Mapping[str, int] = MappingProxyType({"team-a": 30, "team-b": 20}) +) -> Router: return Router( model_list=[ { @@ -1268,13 +1270,15 @@ def test_team_scope_applies_membership_and_user_key_filter( assert repository.aggregated.call_args.kwargs["include_entity_breakdown"] is True +@pytest.mark.parametrize("api_key", ["key-beta", "unattributed", "raw:unattributed", "raw:raw:unattributed"]) def test_team_scope_does_not_allow_an_unowned_api_key( daily_activity_client: tuple[TestClient, _FakeRepository], + api_key: str, ) -> None: client, _ = daily_activity_client response: Final = client.get( "/team/daily/activity/aggregated", - params={**_entity_params("team_ids", "team-a"), "api_key": "key-beta"}, + params={**_entity_params("team_ids", "team-a"), "api_key": api_key}, headers={"x-user-role": LitellmUserRoles.INTERNAL_USER.value, "x-user-id": "user-a"}, ) assert response.status_code == 200, response.text @@ -1622,3 +1626,17 @@ def test_export_content_disposition_is_ascii_and_built_from_canonical_dates( disposition: Final = response.headers["content-disposition"] assert disposition == 'attachment; filename="team-usage-2025-01-01-2025-01-02-daily.csv"' assert disposition.isascii() + + +@pytest.mark.parametrize(("prefix", "query_name", "entity_id"), _ENTITY_CASES) +def test_unattributed_key_filter_does_not_become_an_unfiltered_request( + daily_activity_client: tuple[TestClient, _FakeRepository], prefix: str, query_name: str, entity_id: str +) -> None: + client, repository = daily_activity_client + response: Final = client.get( + f"{prefix}/daily/activity/aggregated", + params={**_entity_params(query_name, entity_id), "api_key": "unattributed"}, + ) + assert response.status_code == 200, response.text + assert response.json()["metadata"]["total_spend"] == 0 + assert repository.aggregated.call_args.args[0].api_keys == ("",) diff --git a/tests/unit/proxy/management_endpoints/test_tag_management_endpoints.py b/tests/unit/proxy/management_endpoints/test_tag_management_endpoints.py index e5d13a381ae..a2e509b6eb8 100644 --- a/tests/unit/proxy/management_endpoints/test_tag_management_endpoints.py +++ b/tests/unit/proxy/management_endpoints/test_tag_management_endpoints.py @@ -1058,7 +1058,8 @@ async def test_internal_user_tag_daily_activity_is_scoped_to_their_keys(): @pytest.mark.asyncio -async def test_internal_user_tag_daily_activity_rejects_unowned_api_key_filter(): +@pytest.mark.parametrize("api_key", ["unowned-key", "unattributed", "raw:unattributed", "raw:raw:unattributed"]) +async def test_internal_user_tag_daily_activity_rejects_unowned_api_key_filter(api_key: str): """ If an internal user filters tag usage by an API key they do not own, the endpoint should return an empty scoped filter instead of exposing that key's @@ -1092,7 +1093,7 @@ async def test_internal_user_tag_daily_activity_rejects_unowned_api_key_filter() result = await get_tag_daily_activity( start_date="2025-01-01", end_date="2025-01-31", - api_key="unowned-key", + api_key=api_key, user_api_key_dict=mock_user_auth, ) diff --git a/tests/unit/repositories/test_daily_activity_sql.py b/tests/unit/repositories/test_daily_activity_sql.py index c57db08e976..f8375b93dad 100644 --- a/tests/unit/repositories/test_daily_activity_sql.py +++ b/tests/unit/repositories/test_daily_activity_sql.py @@ -123,7 +123,6 @@ def test_key_page_sql_orders_exact_spend_and_binds_scope_before_page() -> None: 7, ) assert "SUM(spend::numeric) AS rank_spend" in query.sql - assert "ORDER BY rank_spend DESC, api_key" in query.sql assert "(SELECT COUNT(*) FROM ranked)::bigint AS total_api_keys" in query.sql @@ -329,7 +328,6 @@ def test_entity_rollup_bounds_keys_and_reuses_scope_filters() -> None: assert "api_key = ANY($3::text[])" in query.sql assert query.sql.count("api_key = ANY($3::text[])") == 4 assert query.sql.count("api_key <> $4") == 2 - assert query.sql.count("ORDER BY SUM(spend::numeric) DESC, api_key") == 1 assert "k.entity_id = e.entity_id" in query.sql assert "LIMIT $5" in query.sql assert query.params == ("2026-01-01", "2026-01-31", ["key-1", "key-2"], PTU_SENTINEL_API_KEY, 3) @@ -351,13 +349,11 @@ def test_search_query_escapes_pattern_metacharacters_and_binds_limit() -> None: assert 'LEFT JOIN "LiteLLM_UserTable" u ON u.user_id = d.user_id' in query.sql assert "d.key_alias ILIKE $3 ESCAPE" in query.sql assert "d.user_id ILIKE $3 ESCAPE" in query.sql - assert "api_key ILIKE $3 ESCAPE" in query.sql assert "v.key_alias ILIKE $3 ESCAPE" in query.sql assert "v.user_id ILIKE $3 ESCAPE" in query.sql assert "u.user_email ILIKE $3 ESCAPE" in query.sql assert query.sql.count("ILIKE $3 ESCAPE") == 7 assert "api_key <> $4" in query.sql - assert "ORDER BY SUM(spend::numeric) DESC, api_key" in query.sql assert "LIMIT $5" in query.sql assert query.params == ( "2026-01-01", @@ -374,12 +370,9 @@ def test_model_and_cache_key_queries_bind_filters_sentinel_and_limits() -> None: ) leakage_query = build_cache_leakage_keys_sql(_scope(entity_ids=None), limit=20) - assert "COALESCE(NULLIF(model_group, ''), model) = $3" in model_query.sql assert "api_key <> $4" in model_query.sql - assert "ORDER BY SUM(spend::numeric) DESC, api_key" in model_query.sql assert model_query.params == ("2026-01-01", "2026-01-31", "public-model", PTU_SENTINEL_API_KEY, 5) assert "HAVING SUM(prompt_tokens) - SUM(cache_read_input_tokens) > 0" in leakage_query.sql - assert "ORDER BY SUM(prompt_tokens) - SUM(cache_read_input_tokens) DESC, api_key" in leakage_query.sql assert leakage_query.params == ("2026-01-01", "2026-01-31", PTU_SENTINEL_API_KEY, 20) @@ -401,8 +394,14 @@ def test_query_builders_reject_nonpositive_limits(builder) -> None: ("export_type", "group_key", "key_filter", "joins"), [ (ExportType.DAILY, "''", "", ""), - (ExportType.DAILY_WITH_KEYS, "scoped.api_key", "api_key <> $3", 'LEFT JOIN "LiteLLM_VerificationToken"'), - (ExportType.DAILY_WITH_MODELS, "COALESCE(scoped.model, '')", "api_key <> $3", ""), + ( + ExportType.DAILY_WITH_KEYS, + "CASE WHEN api_key = '' THEN 'unattributed' " + "WHEN api_key = 'unattributed' OR api_key LIKE 'raw:%' THEN 'raw:' || api_key ELSE api_key END", + "api_key <> $3", + 'LEFT JOIN "LiteLLM_VerificationToken"', + ), + (ExportType.DAILY_WITH_MODELS, "COALESCE(NULLIF(model, ''), 'unknown-model')", "api_key <> $3", ""), ( ExportType.DAILY_WITH_USERS, "COALESCE(vt.user_id, dvt.user_id, '')", diff --git a/ui/litellm-dashboard/src/components/UsagePage/components/EntityUsage/TopKeyView.test.tsx b/ui/litellm-dashboard/src/components/UsagePage/components/EntityUsage/TopKeyView.test.tsx index a9c87471511..2b4b69ffe7b 100644 --- a/ui/litellm-dashboard/src/components/UsagePage/components/EntityUsage/TopKeyView.test.tsx +++ b/ui/litellm-dashboard/src/components/UsagePage/components/EntityUsage/TopKeyView.test.tsx @@ -493,6 +493,26 @@ describe("TopKeyView", () => { }); }); + it.each([ + ["raw:unattributed", "unattributed"], + ["raw:raw:unattributed", "raw:unattributed"], + ["raw:raw:raw:key", "raw:raw:key"], + ])("opens the stored key behind reporting ID %s", async (bucket, key) => { + mockKeyInfoV1Call.mockResolvedValue({ key }); + mockTransformKeyInfo.mockReturnValue({ token: key } as KeyResponse); + render( + , + ); + + fireEvent.click(screen.getByRole("button", { name: bucket })); + + expect(await screen.findByText(`Key Info View for ${key}`)).toBeInTheDocument(); + expect(mockKeyInfoV1Call).toHaveBeenCalledWith("test-token", key); + }); + it("should only look up keys that still exist in the database, from both the table and the chart", async () => { mockKeyInfoV1Call.mockResolvedValue({ key: "info" }); mockTransformKeyInfo.mockReturnValue({ transformed: "data" } as unknown as KeyResponse); @@ -503,6 +523,7 @@ describe("TopKeyView", () => { {...baseProps} topKeys={[ { api_key: "session-key", key_alias: null, user: "alice@example.com", key_exists: false, spend: 100 }, + { api_key: "unattributed", key_alias: "Unattributed", user: null, key_exists: false, spend: 25 }, { api_key: "stored-key", key_alias: "Stored", user: null, key_exists: true, spend: 50 }, ]} />, @@ -510,11 +531,14 @@ describe("TopKeyView", () => { expect(screen.getByRole("button", { name: "stored-key" })).toBeInTheDocument(); expect(screen.queryByRole("button", { name: "session-key" })).not.toBeInTheDocument(); + expect(screen.getByText("Unattributed")).toBeInTheDocument(); + expect(screen.queryByRole("button", { name: "unattributed" })).not.toBeInTheDocument(); + await user.click(screen.getByText("unattributed")); await user.click(screen.getByText("session-key")); await user.click(screen.getByRole("radio", { name: "Chart View" })); const bars = chartBars(container); - expect(bars).toHaveLength(2); + expect(bars).toHaveLength(3); bars.forEach((bar) => fireEvent.click(bar)); expect(await screen.findByText("Key Info View for stored-key")).toBeInTheDocument(); diff --git a/ui/litellm-dashboard/src/components/UsagePage/components/EntityUsage/TopKeyView.tsx b/ui/litellm-dashboard/src/components/UsagePage/components/EntityUsage/TopKeyView.tsx index 90a26919e01..291bcb34c73 100644 --- a/ui/litellm-dashboard/src/components/UsagePage/components/EntityUsage/TopKeyView.tsx +++ b/ui/litellm-dashboard/src/components/UsagePage/components/EntityUsage/TopKeyView.tsx @@ -11,6 +11,7 @@ import { transformKeyInfo } from "../../../key_team_helpers/transform_key_info"; import { keyInfoV1Call } from "../../../networking"; import KeyInfoView from "../../../templates/key_info_view"; import { TagUsage } from "../../types"; +import { storedKeyId } from "../../keyActivityData"; const TOP_KEYS_LIMITS = [5, 10, 25, 50] as const; const LIMIT_OPTIONS = TOP_KEYS_LIMITS.map((limit) => ({ value: String(limit), label: String(limit) })); @@ -70,11 +71,12 @@ const TopKeyView: React.FC = ({ topKeys, teams, showTags = fals if (!accessToken || !canOpenKeyInfo(item)) return; try { - const keyInfo = await keyInfoV1Call(accessToken, item.api_key); + const keyId = storedKeyId(item.api_key); + const keyInfo = await keyInfoV1Call(accessToken, keyId); const transformedKeyData = transformKeyInfo(keyInfo); setKeyData(transformedKeyData); - setSelectedKey(item.api_key); + setSelectedKey(keyId); setIsModalOpen(true); // Open modal when key is clicked } catch (error) { console.error("Error fetching key info:", error); diff --git a/ui/litellm-dashboard/src/components/UsagePage/keyActivityData.test.ts b/ui/litellm-dashboard/src/components/UsagePage/keyActivityData.test.ts index b35ea72896a..a5db5db14cc 100644 --- a/ui/litellm-dashboard/src/components/UsagePage/keyActivityData.test.ts +++ b/ui/litellm-dashboard/src/components/UsagePage/keyActivityData.test.ts @@ -7,7 +7,13 @@ import { type KeySpendActivityRow, toDailyData, } from "./dailyActivityApi"; -import { keyDetailFromResponse, mergeKeyActivityPages, overallUsageMetrics } from "./keyActivityData"; +import { + keyDetailFromResponse, + mergeKeyActivityPages, + overallUsageMetrics, + reportingKeyId, + storedKeyId, +} from "./keyActivityData"; const completeMetrics: components["schemas"]["SpendMetrics"] = { api_requests: 2, @@ -80,6 +86,50 @@ const pageRow = (api_key: string): KeySpendActivityRow => ({ }); describe("key activity data", () => { + it.each([ + ["", "unattributed"], + ["unattributed", "raw:unattributed"], + ["raw:key", "raw:raw:key"], + ["raw:raw:key", "raw:raw:raw:key"], + ["ordinary-hash", "ordinary-hash"], + [" ", " "], + ])("round trips stored key %s through reporting", (key, bucket) => { + expect(reportingKeyId(key)).toBe(bucket); + expect(storedKeyId(bucket)).toBe(key); + }); + + it("retains unattributed totals and unknown models in key details", () => { + const unattributed = { + metrics: completeMetrics, + metadata: { key_alias: "Unattributed", key_exists: false }, + }; + const response: DailyActivityAggregatedResponse = { + ...aggregatedResponse, + results: [ + { + date: "2026-09-27", + metrics: completeMetrics, + breakdown: { + api_keys: { unattributed }, + models: { + "unknown-model": { + metrics: completeMetrics, + metadata: {}, + api_key_breakdown: { unattributed }, + }, + }, + }, + }, + ], + }; + + const detail = keyDetailFromResponse(response, "unattributed", []); + expect(detail?.total_requests).toBe(2); + expect(detail?.total_spend).toBe(1.25); + expect(detail?.top_models.map((model) => model.model)).toEqual(["unknown-model"]); + expect(detail?.daily_data[0].metrics.api_requests).toBe(2); + }); + it("builds overall totals from metadata and daily data from complete daily metrics", () => { const summary = overallUsageMetrics( toDailyData(aggregatedResponse), diff --git a/ui/litellm-dashboard/src/components/UsagePage/keyActivityData.ts b/ui/litellm-dashboard/src/components/UsagePage/keyActivityData.ts index 4b2e8c573f2..2d2977058a5 100644 --- a/ui/litellm-dashboard/src/components/UsagePage/keyActivityData.ts +++ b/ui/litellm-dashboard/src/components/UsagePage/keyActivityData.ts @@ -4,6 +4,16 @@ import type { DailyActivityAggregatedResponse, DailyActivityMetadata, KeySpendAc import { toDailyData } from "./dailyActivityApi"; import type { ModelActivityData } from "./types"; +export const reportingKeyId = (key: string): string => { + if (!key) return "unattributed"; + return key === "unattributed" || key.startsWith("raw:") ? `raw:${key}` : key; +}; + +export const storedKeyId = (bucket: string): string => { + if (bucket === "unattributed") return ""; + return bucket.startsWith("raw:") ? bucket.slice(4) : bucket; +}; + export const overallUsageMetrics = ( results: ReturnType, metadata: DailyActivityMetadata, diff --git a/ui/litellm-dashboard/src/components/templates/KeySavingsTab.integration.test.tsx b/ui/litellm-dashboard/src/components/templates/KeySavingsTab.integration.test.tsx index 3567388de05..40ab0e8e77c 100644 --- a/ui/litellm-dashboard/src/components/templates/KeySavingsTab.integration.test.tsx +++ b/ui/litellm-dashboard/src/components/templates/KeySavingsTab.integration.test.tsx @@ -131,6 +131,18 @@ describe("KeySavingsTab", () => { expect(screen.getByTestId("key-savings-scope-note")).toHaveTextContent("Showing your own requests"); }); + it.each([ + ["unattributed", "raw:unattributed"], + ["raw:unattributed", "raw:raw:unattributed"], + ["raw:raw:key", "raw:raw:raw:key"], + ])("requests savings for the real key %s", (keyToken, bucket) => { + const hook = scopedRange().mockReturnValue(mockActivity()); + + renderTab({ keyToken, userRole: "Admin" }); + + expect(hook).toHaveBeenCalledWith("test-token", { userId: null, apiKey: bucket }, activity); + }); + it("reads the whole key for a proxy admin, with no scope note to contradict it", () => { const hook = scopedRange().mockReturnValue(mockActivity()); diff --git a/ui/litellm-dashboard/src/components/templates/KeySavingsTab.tsx b/ui/litellm-dashboard/src/components/templates/KeySavingsTab.tsx index d1395e0be51..03aced7358e 100644 --- a/ui/litellm-dashboard/src/components/templates/KeySavingsTab.tsx +++ b/ui/litellm-dashboard/src/components/templates/KeySavingsTab.tsx @@ -1,6 +1,7 @@ "use client"; import ScopedSavingsTab from "@/components/shared/ScopedSavingsTab"; +import { reportingKeyId } from "@/components/UsagePage/keyActivityData"; import { hasProxyWideSpendView, spendScopeUserId } from "@/utils/roles"; import type { ActivityDateRange } from "@/app/(dashboard)/cost-optimization/_components/useDailyActivityRange"; @@ -15,7 +16,7 @@ interface KeySavingsTabProps { const KeySavingsTab = ({ accessToken, keyToken, userId, userRole, activity }: KeySavingsTabProps) => (