diff --git a/litellm/router.py b/litellm/router.py index a5523d7af79..e3e9454ddce 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -3877,11 +3877,17 @@ class Router: ) _router_timeout: Final = ( - float(self._explicit_timeout) if isinstance(self._explicit_timeout, (int, float)) else None + self.request_timeout + if self.request_timeout is not None + else float(self._explicit_timeout) + if isinstance(self._explicit_timeout, (int, float)) + else None ) _router_stream_timeout: Final = ( self.stream_timeout if self.stream_timeout is not None + else self.request_timeout + if self.request_timeout is not None else self.default_litellm_params.get("stream_timeout") ) kwargs["timeout"] = resolve_llm_passthrough_timeout( diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index e92e5de23dc..148ca0b0ffa 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -8230,6 +8230,16 @@ class TestRouterRequestTimeoutPropagation: == 60 ) + def test_passthrough_prefers_request_timeout_over_router_timeout(self, explicit_request_timeout): + router = self._make_router(timeout=330) + deployment: Final = router.model_list[0] + assert _passthrough_timeout(router, deployment, stream=False) == 300.0 + assert _passthrough_timeout(router, deployment, stream=True) == 300.0 + + def test_passthrough_stream_timeout_still_wins_over_request_timeout(self, explicit_request_timeout): + router = self._make_router(timeout=330, stream_timeout=45) + assert _passthrough_timeout(router, router.model_list[0], stream=True) == 45.0 + # --------------------------------------------------------------------------- # Deferred-stream eager-fetch tests