From 07d32b87adf1df3b25a2dda71e9ed728c026d853 Mon Sep 17 00:00:00 2001
From: DragonAssassin-one <2819637808@qq.com>
Date: Mon, 11 May 2026 11:45:10 +0800
Subject: [PATCH] refactor(router): Reorder async callback filter for
 deployments to improve request handling

---
 litellm/router.py                             | 20 +++----
 .../test_router_tag_routing.py                | 54 +++++++++++++++++++
 2 files changed, 64 insertions(+), 10 deletions(-)

diff --git a/litellm/router.py b/litellm/router.py
index e0ed9b85ec1..8694bdad65f 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -9262,16 +9262,6 @@ class Router:
             )
             healthy_deployments = _pre_cooldown_deployments
 
-        healthy_deployments = await self.async_callback_filter_deployments(
-            model=model,
-            healthy_deployments=healthy_deployments,
-            messages=(
-                cast(List[AllMessageValues], messages) if messages is not None else None
-            ),
-            request_kwargs=request_kwargs,
-            parent_otel_span=parent_otel_span,
-        )
-
         if self.enable_pre_call_checks and messages is not None:
             healthy_deployments = self._pre_call_checks(
                 model=model,
@@ -9290,6 +9280,16 @@ class Router:
             ),
         )
 
+        healthy_deployments = await self.async_callback_filter_deployments(
+            model=model,
+            healthy_deployments=healthy_deployments,
+            messages=(
+                cast(List[AllMessageValues], messages) if messages is not None else None
+            ),
+            request_kwargs=request_kwargs,
+            parent_otel_span=parent_otel_span,
+        )
+
         ## ORDER FILTERING ## -> if user set 'order' in deployments, return deployments with lowest order (e.g. order=1 > order=2)
         _target_order = (request_kwargs or {}).pop("_target_order", None)
         healthy_deployments = litellm.utils._get_order_filtered_deployments(
diff --git a/tests/test_litellm/router_strategy/test_router_tag_routing.py b/tests/test_litellm/router_strategy/test_router_tag_routing.py
index 1fdd3dad4da..10fb25c81d8 100644
--- a/tests/test_litellm/router_strategy/test_router_tag_routing.py
+++ b/tests/test_litellm/router_strategy/test_router_tag_routing.py
@@ -25,6 +25,9 @@ from dotenv import load_dotenv
 import litellm
 from litellm import Router
 from litellm._logging import verbose_logger
+from litellm.router_utils.pre_call_checks.prompt_prefix_affinity_check import (
+    PromptPrefixAffinityCheck,
+)
 
 
 @pytest.mark.asyncio()
@@ -219,6 +222,57 @@ async def test_default_tagged_deployments():
 
         assert response_extra_info["model_id"] == "default-model"
 
+
+@pytest.mark.asyncio()
+async def test_prompt_prefix_affinity_does_not_bypass_default_tag_routing_for_untagged_requests():
+    """
+    Ensure prompt-prefix affinity selection cannot bypass tag-based routing defaults.
+
+    For untagged requests, tag routing prefers deployments tagged "default" when present.
+    Prompt-prefix affinity must select only from that tag-filtered candidate set.
+    """
+
+    router = litellm.Router(
+        model_list=[
+            {
+                "model_name": "gpt-4",
+                "litellm_params": {
+                    "model": "gpt-4o",
+                    "api_base": "https://exampleopenaiendpoint-production.up.railway.app/",
+                    "tags": ["default"],
+                },
+                "model_info": {"id": "default-model"},
+            },
+            {
+                "model_name": "gpt-4",
+                "litellm_params": {
+                    "model": "gpt-4o-mini",
+                    "api_base": "https://exampleopenaiendpoint-production.up.railway.app/",
+                    "tags": ["teamA"],
+                },
+                "model_info": {"id": "teamA-model"},
+            },
+        ],
+        enable_tag_filtering=True,
+        optional_pre_call_checks=["prompt_prefix_affinity"],
+        prompt_prefix_affinity_tokens=64,
+        prompt_prefix_affinity_min_tokens=0,
+    )
+
+    # Force the affinity scorer to prefer the non-default deployment if it ever sees it.
+    def _prefer_team_a(self, prefix_hash: str, deployment_model_id: str) -> int:  # noqa: ANN001
+        return 2 if deployment_model_id == "teamA-model" else 1
+
+    with patch.object(PromptPrefixAffinityCheck, "_get_prefix_hash", return_value="hash"):
+        with patch.object(PromptPrefixAffinityCheck, "_score_deployment", _prefer_team_a):
+            response = await router.acompletion(
+                model="gpt-4",
+                messages=[{"role": "user", "content": "Tell me a joke."}],
+                mock_response="Tell me a joke.",
+            )
+
+    assert response._hidden_params["model_id"] == "default-model"
+
     for _ in range(5):
         # requests tagged with "default", this should pick model with id == "default-model"
         response = await router.acompletion(