From 34c95d53e164b8398c3146cc68e055e141497abf Mon Sep 17 00:00:00 2001
From: Tin Chi Lo <tin@berri.ai>
Date: Fri, 7 Aug 2026 17:15:37 -0700
Subject: [PATCH] perf(router): cache per-model-group strategy resolution,
 invalidated with the model-group info cache

---
 litellm/router.py                             | 10 ++++++--
 .../test_model_group_routing_strategy.py      | 23 +++++++++++++++++--
 2 files changed, 29 insertions(+), 4 deletions(-)

diff --git a/litellm/router.py b/litellm/router.py
index faabac87fbd..f2cbecd8390 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -609,6 +609,9 @@ class Router:
         # ``id()``-reuse risk after GC). See
         # ``litellm.proxy.auth.auth_checks._is_model_cost_zero``.
         self._zero_cost_cache: dict[str, bool] = {}
+        self._get_model_group_strategy_config = lru_cache(maxsize=DEFAULT_MAX_LRU_CACHE_SIZE)(
+            self._compute_model_group_strategy_config
+        )
 
         if model_list is not None:
             # set_model_list will build indices automatically
@@ -1153,7 +1156,7 @@ class Router:
         normalized: Final = self._normalize_strategy(raw) if isinstance(raw, (str, RoutingStrategy)) else None
         return normalized, info.get("routing_strategy_args") or _EMPTY_MAPPING
 
-    def _get_model_group_strategy_config(self, model: str) -> tuple[str, Mapping[str, object]] | None:
+    def _compute_model_group_strategy_config(self, model: str) -> tuple[str, Mapping[str, object]] | None:
         """
         Reads `model_info.routing_strategy` (+ `routing_strategy_args`) off the
         deployments of `model`. When deployments of the same model_name disagree,
@@ -1161,7 +1164,9 @@ class Router:
         stable across edits and upserts (those re-append the deployment, so
         model_list order is not); invalid or conflicting values are reported
         once per offending config so a bad stored value can never take down
-        that model's traffic.
+        that model's traffic. Runs on every request, so the constructor wraps
+        it as `_get_model_group_strategy_config` with a per-instance lru_cache
+        that `_invalidate_model_group_info_cache` clears on model-list changes.
         """
         indices: Final = self.model_name_to_deployment_indices.get(model)
         if not indices:
@@ -10112,6 +10117,7 @@ class Router:
         result and bypass budget enforcement.
         """
         self._cached_get_model_group_info.cache_clear()
+        self._get_model_group_strategy_config.cache_clear()
         self._zero_cost_cache.clear()
 
     def _invalidate_access_groups_cache(self) -> None:
diff --git a/tests/test_litellm/router_strategy/test_model_group_routing_strategy.py b/tests/test_litellm/router_strategy/test_model_group_routing_strategy.py
index c6392b2eccc..274f68cda5f 100644
--- a/tests/test_litellm/router_strategy/test_model_group_routing_strategy.py
+++ b/tests/test_litellm/router_strategy/test_model_group_routing_strategy.py
@@ -244,6 +244,20 @@ def test_deleting_deployment_evicts_its_selector():
     assert all(c is not selector for c in litellm.callbacks)
 
 
+def test_strategy_config_is_cached_until_model_list_changes():
+    router = _build_router(
+        [_deployment("quality", "openai/gpt-4o", "d1", {"routing_strategy": "cost-based-routing"})]
+    )
+    assert router._get_routing_context("quality")[0] == "cost-based-routing"
+
+    for idx in router.model_name_to_deployment_indices["quality"]:
+        router.model_list[idx]["model_info"]["routing_strategy"] = "least-busy"
+    assert router._get_routing_context("quality")[0] == "cost-based-routing"
+
+    _upsert(router, "quality", "openai/gpt-4o", "d1", {"routing_strategy": "latency-based-routing"})
+    assert router._get_routing_context("quality")[0] == "latency-based-routing"
+
+
 def test_config_and_args_warnings_fire_independently(caplog):
     router = _build_router(
         [
@@ -281,8 +295,13 @@ def test_changed_bad_args_warn_again(caplog):
         router._get_routing_context("quality")
         router._get_routing_context("quality")
 
-        for idx in router.model_name_to_deployment_indices["quality"]:
-            router.model_list[idx]["model_info"]["routing_strategy_args"] = {"ttl": "still-bogus"}
+        _upsert(
+            router,
+            "quality",
+            "openai/gpt-4o",
+            "d1",
+            {"routing_strategy": "latency-based-routing", "routing_strategy_args": {"ttl": "still-bogus"}},
+        )
         router._get_routing_context("quality")
     args_warnings = [r for r in caplog.records if "cannot initialize strategy" in r.getMessage()]
     assert len(args_warnings) == 2