Merge 8d928cf865 into cdb60af024

litellm/router.py

@@ -700,14 +700,18 @@ class Router:
         self.deployment_names: list = []  # names of models under litellm_params. ex. azure/chatgpt-v-2
         self.deployment_latency_map = {}
         ### CACHING ###
-        cache_type: Literal["local", "redis", "redis-semantic", "s3", "disk"] = "local"  # default to an in-memory cache
+        cache_type: Literal["local", "redis", "redis-semantic", "s3", "disk"] = "local"
         redis_cache = None
-        cache_config: Final[dict[str, Any]] = {}
+        cache_config: dict[str, Any] = {}
 
         self.client_ttl = client_ttl
         if redis_url is not None or (redis_host is not None and redis_port is not None):
             cache_type = "redis"
 
+            # Seed with cache_kwargs first, then let explicit constructor
+            # arguments override so they always win when both are provided.
+            cache_config.update(cache_kwargs)
+
             if redis_url is not None:
                 cache_config["url"] = redis_url
 
@@ -726,9 +730,12 @@ class Router:
                 )
                 cache_config["db"] = str(redis_db)
 
-            # Add additional key-value pairs from cache_kwargs
-            cache_config.update(cache_kwargs)
             redis_cache = self._create_redis_cache(cache_config)
+        else:
+            # No Redis configured: honor cache_kwargs for non-Redis backends
+            # (e.g. disk). Explicit Redis params above take precedence when set.
+            cache_config.update(cache_kwargs)
+            cache_type = cache_config.pop("type", cache_type)
 
         if cache_responses:
             if litellm.cache is None:

tests/test_litellm/test_router.py

@@ -3783,6 +3783,62 @@ def test_pre_call_checks_no_messages_or_input_does_not_crash(monkeypatch):
     assert counted == []  # token counting skipped entirely, so no misleading error is logged
 
 
+@pytest.mark.asyncio
+async def test_router_cache_kwargs_applied_without_redis():
+    """cache_kwargs should be applied even when Redis is not configured.
+    Regression test for #36309: cache_kwargs were silently ignored outside
+    the Redis block, so disk cache requests fell back to InMemoryCache.
+    """
+    import tempfile
+    import shutil
+    from litellm.caching import DiskCache
+
+    tmpdir = tempfile.mkdtemp(prefix="litellm_cache_test_")
+    try:
+        router = litellm.Router(
+            model_list=[
+                {
+                    "model_name": "test",
+                    "litellm_params": {"model": "gpt-3.5-turbo", "api_key": "sk-test"},
+                }
+            ],
+            cache_responses=True,
+            cache_kwargs={"type": "disk", "disk_cache_dir": tmpdir, "ttl": 60},
+        )
+
+        assert isinstance(litellm.cache.cache, DiskCache), (
+            f"Expected DiskCache, got {type(litellm.cache.cache).__name__}"
+        )
+    finally:
+        shutil.rmtree(tmpdir, ignore_errors=True)
+
+
+@pytest.mark.asyncio
+async def test_router_explicit_redis_params_override_cache_kwargs():
+    """Explicit Redis constructor args must win over cache_kwargs when both set.
+
+    Regression test for Greptile review on #36309: cache_kwargs applied before
+    the explicit Redis block previously let a cache_kwargs 'url' override the
+    explicit redis_url.
+    """
+    with patch("litellm.Router._create_redis_cache") as mock_create:
+        mock_create.return_value = None
+        litellm.Router(
+            model_list=[
+                {
+                    "model_name": "test",
+                    "litellm_params": {"model": "gpt-3.5-turbo", "api_key": "sk-test"},
+                }
+            ],
+            redis_url="redis://explicit-host:6379",
+            cache_kwargs={"url": "redis://kwargs-host:6379", "type": "redis"},
+        )
+        # Explicit redis_url must take precedence over cache_kwargs url
+        mock_create.assert_called_once()
+        passed_config = mock_create.call_args[0][0]
+        assert passed_config["url"] == "redis://explicit-host:6379"
+
+
 @pytest.mark.asyncio
 async def test_aresponses_enforces_context_window_pre_call_check():
     """