fix(router): honor explicit retry opt-out

litellm/router.py

@@ -7513,7 +7513,7 @@ class Router:
             # Check retry policy FIRST, before should_retry_this_error
             # This allows retry policies to override the healthy deployments check
             _retry_policy_applies = False
-            if self.retry_policy is not None or model_group_retry_policy is not None:
+            if request_num_retries != 0 and (self.retry_policy is not None or model_group_retry_policy is not None):
                 # get num_retries from retry policy
                 # Use the model_group captured at the start of the function, or get it from metadata
                 # kwargs.get("model") at this point is the deployment model, not the model_group

tests/test_litellm/test_router_per_deployment_num_retries.py

@@ -490,7 +490,7 @@ class TestRequestNumRetriesBeatsGlobal:
         litellm.callbacks = prev_callbacks
 
     @staticmethod
-    def _router(global_num_retries):
+    def _router(global_num_retries, retry_policy=None):
         return Router(
             model_list=[
                 {
@@ -503,6 +503,7 @@ class TestRequestNumRetriesBeatsGlobal:
                 }
             ],
             num_retries=global_num_retries,
+            retry_policy=retry_policy,
         )
 
     async def _count_attempts(self, *, global_num_retries, request_num_retries):
@@ -530,6 +531,26 @@ class TestRequestNumRetriesBeatsGlobal:
         attempts = await self._count_attempts(global_num_retries=3, request_num_retries=0)
         assert attempts == 1
 
+    @pytest.mark.asyncio
+    async def test_request_num_retries_zero_disables_retry_policy(self):
+        """An explicit zero remains a single attempt when a retry policy matches the error."""
+        counter = _AttemptCounter()
+        litellm.callbacks = [counter]
+        router = self._router(
+            global_num_retries=3,
+            retry_policy=RetryPolicy(InternalServerErrorRetries=2),
+        )
+
+        with patch("asyncio.sleep", return_value=None):
+            with pytest.raises(litellm.InternalServerError):
+                await router.acompletion(
+                    model="mock",
+                    messages=[{"role": "user", "content": "hi"}],
+                    num_retries=0,
+                )
+
+        assert counter.attempts == 1
+
     @pytest.mark.asyncio
     async def test_global_num_retries_applies_when_request_omits_it(self):
         """No request num_retries -> the global still applies: 1 initial + 3 retries = 4."""