fix(router): honor ServiceUnavailableErrorRetries and InternalServerErrorRetries in retry policy

litellm/router_utils/get_retry_from_policy.py

@@ -8,7 +8,9 @@ from litellm.exceptions import (
     AuthenticationError,
     BadRequestError,
     ContentPolicyViolationError,
+    InternalServerError,
     RateLimitError,
+    ServiceUnavailableError,
     Timeout,
 )
 from litellm.types.router import RetryPolicy
@@ -26,6 +28,8 @@ def get_num_retries_from_retry_policy(
     TimeoutErrorRetries: Optional[int] = None
     RateLimitErrorRetries: Optional[int] = None
     ContentPolicyViolationErrorRetries: Optional[int] = None
+    InternalServerErrorRetries: Optional[int] = None
+    ServiceUnavailableErrorRetries: Optional[int] = None
     """
     # if we can find the exception then in the retry policy -> return the number of retries
 
@@ -48,6 +52,10 @@ def get_num_retries_from_retry_policy(
         and retry_policy.ContentPolicyViolationErrorRetries is not None
     ):
         return retry_policy.ContentPolicyViolationErrorRetries
+    if isinstance(exception, ServiceUnavailableError) and retry_policy.ServiceUnavailableErrorRetries is not None:
+        return retry_policy.ServiceUnavailableErrorRetries
+    if isinstance(exception, InternalServerError) and retry_policy.InternalServerErrorRetries is not None:
+        return retry_policy.InternalServerErrorRetries
     if isinstance(exception, BadRequestError) and retry_policy.BadRequestErrorRetries is not None:
         return retry_policy.BadRequestErrorRetries
 

litellm/types/router.py

@@ -95,6 +95,7 @@ class RetryPolicy(BaseModel):
     RateLimitErrorRetries: Optional[int] = None
     ContentPolicyViolationErrorRetries: Optional[int] = None
     InternalServerErrorRetries: Optional[int] = None
+    ServiceUnavailableErrorRetries: Optional[int] = None
 
 
 class UpdateRouterConfig(BaseModel):

tests/test_litellm/router_utils/test_get_retry_from_policy.py

@@ -0,0 +1,102 @@
+import litellm
+from litellm.router_utils.get_retry_from_policy import (
+    get_num_retries_from_retry_policy,
+)
+from litellm.types.router import RetryPolicy
+
+
+def _service_unavailable_error() -> litellm.ServiceUnavailableError:
+    return litellm.ServiceUnavailableError(
+        message="model is down",
+        llm_provider="openai",
+        model="gpt-5.6",
+    )
+
+
+def _internal_server_error() -> litellm.InternalServerError:
+    return litellm.InternalServerError(
+        message="upstream 500",
+        llm_provider="openai",
+        model="gpt-5.6",
+    )
+
+
+def test_service_unavailable_error_retries_honored():
+    policy = RetryPolicy(ServiceUnavailableErrorRetries=0)
+
+    assert (
+        get_num_retries_from_retry_policy(
+            exception=_service_unavailable_error(),
+            retry_policy=policy,
+        )
+        == 0
+    )
+
+
+def test_service_unavailable_error_retries_nonzero():
+    policy = RetryPolicy(ServiceUnavailableErrorRetries=4)
+
+    assert (
+        get_num_retries_from_retry_policy(
+            exception=_service_unavailable_error(),
+            retry_policy=policy,
+        )
+        == 4
+    )
+
+
+def test_internal_server_error_retries_honored():
+    policy = RetryPolicy(InternalServerErrorRetries=0)
+
+    assert (
+        get_num_retries_from_retry_policy(
+            exception=_internal_server_error(),
+            retry_policy=policy,
+        )
+        == 0
+    )
+
+
+def test_service_unavailable_not_covered_by_internal_server_error_retries():
+    policy = RetryPolicy(InternalServerErrorRetries=0)
+
+    assert (
+        get_num_retries_from_retry_policy(
+            exception=_service_unavailable_error(),
+            retry_policy=policy,
+        )
+        is None
+    )
+
+
+def test_internal_server_error_not_covered_by_service_unavailable_retries():
+    policy = RetryPolicy(ServiceUnavailableErrorRetries=0)
+
+    assert (
+        get_num_retries_from_retry_policy(
+            exception=_internal_server_error(),
+            retry_policy=policy,
+        )
+        is None
+    )
+
+
+def test_service_unavailable_error_retries_from_dict_policy():
+    assert (
+        get_num_retries_from_retry_policy(
+            exception=_service_unavailable_error(),
+            retry_policy={"ServiceUnavailableErrorRetries": 0},
+        )
+        == 0
+    )
+
+
+def test_service_unavailable_error_retries_from_model_group_policy():
+    assert (
+        get_num_retries_from_retry_policy(
+            exception=_service_unavailable_error(),
+            model_group="gpt-5.6",
+            model_group_retry_policy={"gpt-5.6": RetryPolicy(ServiceUnavailableErrorRetries=1)},
+        )
+        == 1
+    )

tests/test_litellm/test_router.py

@@ -6574,3 +6574,34 @@ def test_model_info_is_active_for_environment_matrix(monkeypatch):
     monkeypatch.delenv("LITELLM_ENVIRONMENT")
     with pytest.raises(ValueError, match="LITELLM_ENVIRONMENT"):
         model_info_is_active_for_environment(model_info={"supported_environments": ["production"]})
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("policy_retries,expected_calls", [(0, 1), (1, 2)])
+async def test_router_retry_policy_service_unavailable_retries(policy_retries, expected_calls):
+    from litellm.types.router import RetryPolicy
+
+    router = litellm.Router(
+        model_list=[
+            {
+                "model_name": "gpt-5.6",
+                "litellm_params": {"model": "openai/gpt-5.6", "api_key": "fake-key"},
+            }
+        ],
+        retry_policy=RetryPolicy(ServiceUnavailableErrorRetries=policy_retries),
+        disable_cooldowns=True,
+    )
+
+    error = litellm.ServiceUnavailableError(
+        message="model is down",
+        llm_provider="openai",
+        model="gpt-5.6",
+    )
+    with patch.object(litellm, "acompletion", AsyncMock(side_effect=error)) as mock_acompletion:
+        with pytest.raises(litellm.ServiceUnavailableError):
+            await router.acompletion(
+                model="gpt-5.6",
+                messages=[{"role": "user", "content": "hi"}],
+            )
+
+    assert mock_acompletion.call_count == expected_calls

ui/litellm-dashboard/src/app/(dashboard)/models-and-endpoints/components/ModelRetrySettingsTab.tsx

@@ -30,6 +30,7 @@ const retryPolicyMap: Record<string, string> = {
   "RateLimitError (429)": "RateLimitErrorRetries",
   "ContentPolicyViolationError (400)": "ContentPolicyViolationErrorRetries",
   "InternalServerError (500)": "InternalServerErrorRetries",
+  "ServiceUnavailableError (503)": "ServiceUnavailableErrorRetries",
 };
 
 const ModelRetrySettingsTab = ({

ui/litellm-dashboard/src/lib/http/schema.d.ts

@@ -30947,6 +30947,8 @@ export interface components {
             InternalServerErrorRetries?: number | null;
             /** Ratelimiterrorretries */
             RateLimitErrorRetries?: number | null;
+            /** Serviceunavailableerrorretries */
+            ServiceUnavailableErrorRetries?: number | null;
             /** Timeouterrorretries */
             TimeoutErrorRetries?: number | null;
         };