From 836bf0807b62fe346697e3a1b987cc5b05afbbf9 Mon Sep 17 00:00:00 2001 From: Shivam Rawat Date: Fri, 17 Jul 2026 18:19:02 -0700 Subject: [PATCH] fix(router): keep team wildcard routers fresh and prioritize them over global patterns team_pattern_routers retained deleted/replaced deployments, so team users could keep resolving stale credentials; now set_model_list resets the registry and deployment removal prunes it. Also consult the team wildcard router before the global pattern_router in get_deployment_credentials_with_provider so a global pattern like "openai/*" no longer shadows the team's own entry Co-authored-by: Cursor --- litellm/router.py | 26 ++++-- .../router_utils/pattern_match_deployments.py | 11 +++ tests/test_litellm/test_router.py | 92 +++++++++++++++++++ 3 files changed, 121 insertions(+), 8 deletions(-) diff --git a/litellm/router.py b/litellm/router.py index dbc6da106e7..6a055f54b0e 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -7876,6 +7876,7 @@ class Router: self.model_id_to_deployment_index_map = {} # Reset the index self.model_name_to_deployment_indices = {} # Reset the model_name index self.team_model_to_deployment_indices = {} # Reset the team_model index + self.team_pattern_routers = {} self.team_public_model_names = frozenset() # Reset per-strategy router registries so hot-reload doesn't leave # stale routers pointing at the old model_list. @@ -8232,6 +8233,12 @@ class Router: public_model_name for _, public_model_name in self.team_model_to_deployment_indices ) + for team_id in list(self.team_pattern_routers.keys()): + team_pattern_router = self.team_pattern_routers[team_id] + team_pattern_router.remove_deployment(model_id) + if not team_pattern_router.patterns: + del self.team_pattern_routers[team_id] + def _update_team_model_index(self, model: dict, idx: int) -> None: """ Helper to update team_model_to_deployment_indices for a single deployment. @@ -8460,8 +8467,8 @@ class Router: return None def get_deployment_credentials_with_provider( - self, model_id: str, team_id: Optional[str] = None - ) -> Optional[Dict[str, Any]]: + self, model_id: str, team_id: str | None = None + ) -> dict[str, Any] | None: """ Get API credentials and provider info from a model name in model_list. Useful for passthrough endpoints (files, batches, etc.) that need credentials. @@ -8492,19 +8499,22 @@ class Router: if deployment is None: deployment = self.get_deployment_by_model_group_name(model_group_name=model_id) - # If not found, check team-scoped deployments (team public model names, - # e.g. team wildcard models like "openai/*", live in a separate index). + # If not found, check team-scoped deployments whose team public model + # name exactly matches model_id (wildcard team names are matched via + # team_pattern_routers below). if deployment is None and team_id is not None: team_indices = self.team_model_to_deployment_indices.get((team_id, model_id), []) if team_indices: team_model = self.model_list[team_indices[0]] deployment = Deployment(**team_model) if isinstance(team_model, dict) else team_model - # If still not found, check for wildcard pattern matches + # If still not found, check for wildcard pattern matches. Team wildcard + # matches take priority so a global pattern (e.g. "openai/*") doesn't + # shadow the team's own entry. if deployment is None: - potential_wildcard_models = self.pattern_router.route(model_id) or [] - if not potential_wildcard_models and team_id is not None and team_id in self.team_pattern_routers: - potential_wildcard_models = self.team_pattern_routers[team_id].route(model_id) or [] + team_pattern_router = self.team_pattern_routers.get(team_id) if team_id is not None else None + team_wildcard_models = (team_pattern_router.route(model_id) or []) if team_pattern_router else [] + potential_wildcard_models = team_wildcard_models or self.pattern_router.route(model_id) or [] if potential_wildcard_models: # Use the first matching wildcard deployment deployment_dict = potential_wildcard_models[0] diff --git a/litellm/router_utils/pattern_match_deployments.py b/litellm/router_utils/pattern_match_deployments.py index c08f8e95cf4..7e1ed739ef8 100644 --- a/litellm/router_utils/pattern_match_deployments.py +++ b/litellm/router_utils/pattern_match_deployments.py @@ -73,6 +73,17 @@ class PatternMatchRouter: self.patterns[regex] = [] self.patterns[regex].append(llm_deployment) + def remove_deployment(self, model_id: str) -> None: + """ + Remove every deployment with the given model id from the pattern registry, + dropping any pattern whose deployment list becomes empty. + """ + self.patterns = { + regex: remaining + for regex, deployments in self.patterns.items() + if (remaining := [d for d in deployments if (d.get("model_info") or {}).get("id") != model_id]) + } + def _pattern_to_regex(self, pattern: str) -> str: """ Convert a wildcard pattern to a regex pattern diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index c2c98c8869c..0fe855151b0 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -3535,6 +3535,98 @@ def test_get_deployment_credentials_with_provider_resolves_credential_name(): litellm.credential_list = [] +def _team_wildcard_model(api_key: str, model_id: str = "team-wildcard-id") -> dict: + return { + "model_name": f"model_name_team-1_{model_id}", + "litellm_params": {"model": "openai/*", "api_key": api_key}, + "model_info": { + "id": model_id, + "team_id": "team-1", + "team_public_model_name": "openai/*", + }, + } + + +def test_get_deployment_credentials_with_provider_team_wildcard_priority(): + """ + Regression: a global wildcard pattern (e.g. "openai/*") must not shadow a + team's own wildcard entry. When team_id is provided, the team wildcard + deployment's credentials win; without team_id the global one is used. + """ + router = litellm.Router( + model_list=[ + { + "model_name": "openai/*", + "litellm_params": {"model": "openai/*", "api_key": "global-key"}, + }, + _team_wildcard_model(api_key="team-key"), + ], + ) + + team_credentials = router.get_deployment_credentials_with_provider( + model_id="openai/gpt-5.2", team_id="team-1" + ) + assert team_credentials is not None + assert team_credentials["api_key"] == "team-key" + + global_credentials = router.get_deployment_credentials_with_provider( + model_id="openai/gpt-5.2" + ) + assert global_credentials is not None + assert global_credentials["api_key"] == "global-key" + + +def test_team_wildcard_credentials_not_usable_after_delete_deployment(): + """ + Regression: team_pattern_routers retained deleted deployments, so a team + user could keep resolving credentials of a deleted wildcard deployment. + """ + router = litellm.Router(model_list=[_team_wildcard_model(api_key="old-key")]) + + assert ( + router.get_deployment_credentials_with_provider( + model_id="openai/gpt-5.2", team_id="team-1" + ) + is not None + ) + + router.delete_deployment(id="team-wildcard-id") + + assert ( + router.get_deployment_credentials_with_provider( + model_id="openai/gpt-5.2", team_id="team-1" + ) + is None + ) + + +def test_team_wildcard_credentials_refreshed_on_upsert_and_set_model_list(): + """ + Regression: replacing a team wildcard deployment (upsert or model list + reload) must serve the new credentials, not the stale cached ones. + """ + from litellm.types.router import Deployment + + router = litellm.Router(model_list=[_team_wildcard_model(api_key="old-key")]) + + router.upsert_deployment( + deployment=Deployment(**_team_wildcard_model(api_key="new-key")) + ) + credentials = router.get_deployment_credentials_with_provider( + model_id="openai/gpt-5.2", team_id="team-1" + ) + assert credentials is not None + assert credentials["api_key"] == "new-key" + + router.set_model_list(model_list=[]) + assert ( + router.get_deployment_credentials_with_provider( + model_id="openai/gpt-5.2", team_id="team-1" + ) + is None + ) + + def test_get_available_guardrail_single_deployment(): """ Test get_available_guardrail returns the single guardrail when only one exists.