From 2e854dddb3a0b45880d4e768dfd6a7fd219acd05 Mon Sep 17 00:00:00 2001 From: Tin Chi Lo Date: Thu, 27 Aug 2026 20:55:29 -0700 Subject: [PATCH] fix(router): drop a tier param the routed target cannot take A complexity tier's litellm_params are an operator override applied to every request that tier routes, and they were written into the request kwargs unconditionally. When the tier set a param the target does not declare, get_optional_params raised UnsupportedParamsError before the request left the proxy, so the whole tier answered 400. The bundled Lite preset sets reasoning_effort on its complex tier, and four of the thirteen kimi-k3 map entries reject that param, so a router built from a first-party template failed on every complex prompt Filter the tier params at both store sites against what the group's deployments declare. The candidate set is asked of the module that raises rather than derived from a second list, so credentials, endpoint and transport controls are never at risk: base_url, timeout, default_headers, organization and deployment_id are not chat completion params and never reach that comparison. A param survives if any deployment could take it, since routing has not picked one yet, and it survives an unresolvable provider or an empty group, since a best-effort filter must not narrow what the request already did The skip list _check_valid_arg applies before rejecting a param now has one owner both it and the router read, so the two cannot drift --- litellm/router.py | 72 +++++++++- litellm/utils.py | 26 ++-- .../router_strategy/test_complexity_router.py | 6 +- tests/test_litellm/test_router.py | 125 ++++++++++++++++++ 4 files changed, 216 insertions(+), 13 deletions(-) diff --git a/litellm/router.py b/litellm/router.py index facb4978ae5..e00b495725b 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -229,6 +229,7 @@ from litellm.types.utils import ( StandardLoggingPayload, StandardLoggingRoutingDecision, Usage, + all_litellm_params, shared_backend_model_info, ) from litellm.types.utils import ModelInfo as ModelMapInfo @@ -243,6 +244,7 @@ from litellm.utils import ( get_secret, get_utc_datetime, is_region_allowed, + provider_rejectable_params, set_live_deployment_replay, ) @@ -10832,6 +10834,68 @@ class Router: } return {**deployment, "model_info": model_info} # mutable-ok: DeploymentTypedDict rows are plain dicts + TIER_PARAMS_NEVER_DROPPED: Final = frozenset(all_litellm_params) | frozenset( + {"additional_drop_params", "drop_params", "messages", "model", "extra_headers"} + ) + + @staticmethod + def _deployment_accepts_param(deployment: DeploymentTypedDict, group: str, param: str) -> bool: + deployment_params: Final = deployment.get("litellm_params") + if not deployment_params: + return True + try: + model, custom_llm_provider, _, _ = litellm.get_llm_provider( + model=deployment_params.get("model") or group, + custom_llm_provider=deployment_params.get("custom_llm_provider"), + ) + supported: Final = litellm.get_supported_openai_params(model=model, custom_llm_provider=custom_llm_provider) + except Exception as e: # noqa: BLE001 # best-effort filter: an unresolvable provider must not narrow the request + verbose_router_logger.debug( + "litellm.router.py::_deployment_accepts_param: keeping %s for model=%s. Got - %s", param, group, e + ) + return True + return supported is None or param in supported + + def _tier_params_the_target_accepts(self, model: str, tier_params: Mapping[str, object]) -> Mapping[str, object]: + """Drop an OpenAI param that no deployment behind ``model`` declares. + + A tier's litellm_params are an operator override applied to every request the tier routes, + so one the target cannot take turns that whole tier into a 400 raised before the request + leaves the proxy. The candidates are exactly what get_optional_params can reject, asked of + the module that raises, so credentials and endpoint controls are never at risk. + + TIER_PARAMS_NEVER_DROPPED is excluded on top of that, for two reasons. No provider lists a + litellm control among its supported params, so "no deployment declares it" means litellm + consumes it rather than that the target refuses it, and dropping one changes litellm's own + behavior: dropping drop_params or additional_drop_params silently disables the sanitization + the operator configured. Providers do list extra_headers, but it carries auth, tenancy and + routing information, so sending fewer headers than configured is worse than today's error. + + The trade this filter makes is a param for a working request, which is right for one that + only shapes how the model answers and wrong for anything else. + + A param survives if ANY deployment could take it, because routing has not chosen one yet, + and it survives both an unresolvable provider and a group with no deployments, because a + best-effort filter must never narrow what the request already did. + """ + deployments: Final = self.get_model_list(model_name=model) or () + if not deployments: + return tier_params + candidates: Final = provider_rejectable_params(tier_params) - self.TIER_PARAMS_NEVER_DROPPED + unsupported: Final = frozenset( + param + for param in candidates + if not any(self._deployment_accepts_param(deployment, model, param) for deployment in deployments) + ) + if not unsupported: + return tier_params + verbose_router_logger.warning( + "litellm.router.py: dropping tier params %s for model=%s, no deployment behind it declares them", + ", ".join(sorted(unsupported)), + model, + ) + return MappingProxyType({key: value for key, value in tier_params.items() if key not in unsupported}) + def get_model_list( self, model_name: str | None = None, team_id: str | None = None ) -> list[DeploymentTypedDict] | None: @@ -11887,7 +11951,9 @@ class Router: model = pre_routing_hook_response.model messages = pre_routing_hook_response.messages if pre_routing_hook_response.litellm_params: - request_kwargs.update(pre_routing_hook_response.litellm_params) + request_kwargs.update( + self._tier_params_the_target_accepts(model, pre_routing_hook_response.litellm_params) + ) ######################################################### # Resolve the strategy and logger AFTER the pre-routing hook, since @@ -11998,7 +12064,9 @@ class Router: model = pre_routing_hook_response.model messages = pre_routing_hook_response.messages if pre_routing_hook_response.litellm_params: - request_kwargs.update(pre_routing_hook_response.litellm_params) + request_kwargs.update( + self._tier_params_the_target_accepts(model, pre_routing_hook_response.litellm_params) + ) # 2. Get healthy deployments healthy_deployments: Final = await self.async_get_healthy_deployments( diff --git a/litellm/utils.py b/litellm/utils.py index 520c40f67c0..54b5448b7e9 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -4150,17 +4150,11 @@ def get_optional_params( unsupported_params: Final = {} for k in non_default_params: if k not in supported_params: - if k == "user" or k == "stream_options" or k == "stream": + if k in PROVIDER_UNVALIDATED_PARAMS: continue if k == "n" and n == 1: # langchain sends n=1 as a default value continue # skip this param - if ( - k == "max_retries" - ): # TODO: This is a patch. We support max retries for OpenAI, Azure. For non OpenAI LLMs we need to add support for max retries - continue # skip this param - # Always keeps this in elif code blocks - else: - unsupported_params[k] = non_default_params[k] + unsupported_params[k] = non_default_params[k] if unsupported_params: if litellm.drop_params is True or (drop_params is not None and drop_params is True): @@ -4729,6 +4723,22 @@ def _apply_openai_param_overrides(optional_params: dict, non_default_params: dic return optional_params +PROVIDER_UNVALIDATED_PARAMS: Final = frozenset({"user", "stream_options", "stream", "max_retries"}) + + +def provider_rejectable_params(passed_params: Mapping[str, object]) -> frozenset[str]: + """The params a provider can actually be rejected for, i.e. the ones _check_valid_arg compares + against its supported list. + + Anything outside this set never reaches that comparison. Endpoint and transport controls such as + base_url, timeout, default_headers, organization and deployment_id are not chat completion + params at all, so a caller filtering on "is this an OpenAI param" would discard configuration the + request needs while never touching what the provider would have rejected. + """ + params: Final = dict(passed_params) # mutable-ok: get_non_default_params takes a dict + return frozenset(get_non_default_params(params)) - PROVIDER_UNVALIDATED_PARAMS + + def get_non_default_params(passed_params: dict) -> dict: # filter out those parameters that were passed with non-default values non_default_params: Final = { diff --git a/tests/test_litellm/router_strategy/test_complexity_router.py b/tests/test_litellm/router_strategy/test_complexity_router.py index ae9e3907aeb..6f76be18256 100644 --- a/tests/test_litellm/router_strategy/test_complexity_router.py +++ b/tests/test_litellm/router_strategy/test_complexity_router.py @@ -2213,14 +2213,14 @@ class TestRouterPreRoutingAliasOverrides: "complexity_router_config": { "tiers": { "SIMPLE": { - "model_name": "gpt-4o-mini", + "model_name": "gpt-5-mini", "litellm_params": {"reasoning_effort": "xhigh"}, } } }, }, }, - {"model_name": "gpt-4o-mini", "litellm_params": {"model": "openai/gpt-4o-mini"}}, + {"model_name": "gpt-5-mini", "litellm_params": {"model": "openai/gpt-5-mini"}}, ] ) request_kwargs: Dict = {"reasoning_effort": "low"} @@ -2231,7 +2231,7 @@ class TestRouterPreRoutingAliasOverrides: messages=[{"role": "user", "content": "hi"}], ) - assert deployment["model_name"] == "gpt-4o-mini" + assert deployment["model_name"] == "gpt-5-mini" assert request_kwargs["reasoning_effort"] == "xhigh" @pytest.mark.asyncio diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 2948568198f..24941e88185 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -11194,3 +11194,128 @@ def test_resolved_litellm_models_answers_through_every_channel_a_request_uses( result is not "the call fails", so what to do about it stays each caller's policy. """ assert set(_resolution_router().resolved_litellm_models(model_name)) == set(expected) + + +class TestTierParamsTheTargetAccepts: + """A tier's litellm_params are applied to every request that tier routes, so one the target + cannot take raised UnsupportedParamsError before the request left the proxy, turning the whole + tier into a 400.""" + + @pytest.fixture(autouse=True) + def force_local_model_cost(self, monkeypatch): + from litellm.litellm_core_utils.get_model_cost_map import GetModelCostMap + + monkeypatch.setattr(litellm, "model_cost", GetModelCostMap.load_local_model_cost_map()) + + @staticmethod + def _router(model: str) -> litellm.Router: + return litellm.Router( + model_list=[{"model_name": "tiered", "litellm_params": {"model": model, "api_key": "sk-x"}}] + ) + + def test_drops_a_param_no_deployment_declares(self): + router = self._router("novita/moonshotai/kimi-k3") + + accepted = router._tier_params_the_target_accepts("tiered", {"reasoning_effort": "max"}) + + assert accepted == {} + + def test_keeps_a_param_the_deployment_declares(self): + router = self._router("fireworks_ai/kimi-k3") + + accepted = router._tier_params_the_target_accepts("tiered", {"reasoning_effort": "max"}) + + assert accepted == {"reasoning_effort": "max"} + + @pytest.mark.parametrize( + "control, value", + [ + ("api_base", "https://example.invalid"), + ("api_key", "sk-tier"), + ("base_url", "https://example.invalid"), + ("timeout", 30), + ("default_headers", {"x-tier": "1"}), + ("organization", "org-tier"), + ("deployment_id", "dep-tier"), + ], + ) + def test_keeps_credentials_and_transport_controls(self, control, value): + """These are not chat completion params, so get_optional_params never compares them against + a provider's supported list. Filtering on "is this an OpenAI param" would discard the + configuration the request needs while never touching what the provider would reject.""" + router = self._router("novita/moonshotai/kimi-k3") + + accepted = router._tier_params_the_target_accepts("tiered", {control: value, "reasoning_effort": "max"}) + + assert accepted == {control: value} + + @pytest.mark.parametrize( + "control, value", + [ + ("additional_drop_params", ["seed"]), + ("drop_params", True), + ("allowed_openai_params", ["reasoning_effort"]), + ("api_version", "2024-02-01"), + ("metadata", {"tier": "complex"}), + ], + ) + def test_keeps_litellm_controls_the_provider_never_lists(self, control, value): + """No provider lists a litellm control among its supported params, so "no deployment + declares it" means litellm consumes it, not that the target refuses it. Dropping + drop_params or additional_drop_params would silently disable the operator's sanitization.""" + router = self._router("novita/moonshotai/kimi-k3") + + accepted = router._tier_params_the_target_accepts("tiered", {control: value, "reasoning_effort": "max"}) + + assert accepted == {control: value} + + def test_keeps_extra_headers_even_when_the_provider_omits_it(self): + """Several providers leave extra_headers out of their supported params, so the filter would + drop it. Headers carry auth and tenancy, so sending fewer than the operator configured is + worse than the error they already get.""" + router = self._router("ai21/jamba-1.5-mini") + + accepted = router._tier_params_the_target_accepts( + "tiered", {"extra_headers": {"x-tenant": "acme"}, "reasoning_effort": "max"} + ) + + assert accepted == {"extra_headers": {"x-tenant": "acme"}} + + def test_keeps_a_param_any_deployment_in_the_group_declares(self): + """Routing has not picked a deployment yet, so one capable member keeps the param alive.""" + router = litellm.Router( + model_list=[ + {"model_name": "tiered", "litellm_params": {"model": "novita/moonshotai/kimi-k3", "api_key": "k"}}, + {"model_name": "tiered", "litellm_params": {"model": "fireworks_ai/kimi-k3", "api_key": "k"}}, + ] + ) + + accepted = router._tier_params_the_target_accepts("tiered", {"reasoning_effort": "max"}) + + assert accepted == {"reasoning_effort": "max"} + + def test_deployment_accepts_param_reads_the_provider(self): + deployment = {"model_name": "x", "litellm_params": {"model": "fireworks_ai/kimi-k3"}} + + assert litellm.Router._deployment_accepts_param(deployment, "x", "reasoning_effort") is True + + def test_deployment_accepts_param_is_false_when_the_provider_omits_it(self): + deployment = {"model_name": "x", "litellm_params": {"model": "novita/moonshotai/kimi-k3"}} + + assert litellm.Router._deployment_accepts_param(deployment, "x", "reasoning_effort") is False + + @pytest.mark.parametrize( + "deployment", + [{"model_name": "x"}, {"model_name": "x", "litellm_params": {}}, {"model_name": "x", "litellm_params": {"model": "not-a-real-provider/nope"}}], + ) + def test_deployment_accepts_param_fails_open(self, deployment): + """An unresolvable deployment must not be the reason a param is dropped.""" + assert litellm.Router._deployment_accepts_param(deployment, "x", "reasoning_effort") is True + + def test_keeps_everything_for_an_unknown_group(self): + """An unresolvable target must never narrow what the request already did.""" + router = self._router("fireworks_ai/kimi-k3") + + accepted = router._tier_params_the_target_accepts("no-such-group", {"reasoning_effort": "max"}) + + assert accepted == {"reasoning_effort": "max"}