fix(router): apply cache_kwargs when no redis connection params are given

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/router.py

@@ -155,6 +155,7 @@ from litellm.router_utils.router_callbacks.track_deployment_metrics import (
     increment_deployment_successes_for_current_minute,
 )
 from litellm.scheduler import FlowItem, Scheduler
+from litellm.types.caching import LiteLLMCacheType
 from litellm.types.llms.openai import (
     AllMessageValues,
     FileTypes,
@@ -345,6 +346,17 @@ def _replay_live_router_model_cost() -> None:
 set_live_deployment_replay(_replay_live_router_model_cost)
 
 
+def _resolve_cache_type(requested_type: object) -> LiteLLMCacheType:
+    """Resolve the `type` entry of `cache_kwargs` when no redis connection params are given."""
+    if requested_type is None:
+        return LiteLLMCacheType.LOCAL
+    if isinstance(requested_type, LiteLLMCacheType):
+        return requested_type
+    if isinstance(requested_type, str):
+        return LiteLLMCacheType(requested_type)
+    raise ValueError(f"cache_kwargs['type'] must be a string cache type, got {type(requested_type).__name__}")
+
+
 # Kwargs that log_retry must not copy into a retry breadcrumb. The breadcrumbs reach spend
 # logs and logging callbacks, and these carry either the request payload or router-internal
 # walk state rather than anything that identifies the failed attempt.
@@ -525,14 +537,15 @@ 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
+        uses_redis: Final[bool] = redis_url is not None or (redis_host is not None and redis_port is not None)
+        cache_type: Final[LiteLLMCacheType] = (
+            LiteLLMCacheType.REDIS if uses_redis else _resolve_cache_type(cache_kwargs.get("type"))
+        )
         redis_cache = None
         cache_config: Final[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"
-
+        if uses_redis:
             if redis_url is not None:
                 cache_config["url"] = redis_url
 
@@ -551,8 +564,9 @@ class Router:
                 )
                 cache_config["db"] = str(redis_db)
 
-            # Add additional key-value pairs from cache_kwargs
-            cache_config.update(cache_kwargs)
+        cache_config.update({key: value for key, value in cache_kwargs.items() if key != "type"})
+
+        if uses_redis:
             redis_cache = self._create_redis_cache(cache_config)
 
         if cache_responses:

tests/test_litellm/test_router.py

@@ -7615,3 +7615,66 @@ def test_ensure_deployment_affinity_callback_is_idempotent():
     finally:
         for cb in router.optional_callbacks or []:
             litellm.logging_callback_manager.remove_callback_from_all_lists(cb)
+
+@pytest.fixture
+def reset_litellm_cache():
+    previous = litellm.cache
+    litellm.cache = None
+    try:
+        yield
+    finally:
+        litellm.cache = previous
+
+
+def test_router_cache_kwargs_applied_without_redis(reset_litellm_cache):
+    litellm.Router(
+        model_list=[{"model_name": "gpt-4o", "litellm_params": {"model": "gpt-4o"}}],
+        cache_responses=True,
+        cache_kwargs={"ttl": 60, "namespace": "router-ns"},
+    )
+
+    assert litellm.cache is not None
+    assert litellm.cache.ttl == 60
+    assert litellm.cache.namespace == "router-ns"
+
+
+def test_router_cache_kwargs_type_selects_backend_without_redis(tmp_path, reset_litellm_cache):
+    pytest.importorskip("diskcache")
+    from litellm.caching.disk_cache import DiskCache
+
+    litellm.Router(
+        model_list=[{"model_name": "gpt-4o", "litellm_params": {"model": "gpt-4o"}}],
+        cache_responses=True,
+        cache_kwargs={"type": "disk", "disk_cache_dir": str(tmp_path), "ttl": 60},
+    )
+
+    assert litellm.cache is not None
+    assert isinstance(litellm.cache.cache, DiskCache)
+    assert litellm.cache.ttl == 60
+
+
+def test_router_cache_kwargs_type_ignored_when_redis_params_given(reset_litellm_cache):
+    created_caches = []
+
+    class _FakeRedisCache:
+        def __init__(self, **kwargs):
+            created_caches.append(kwargs)
+
+    with patch("litellm.router.RedisCache", _FakeRedisCache):
+        litellm.Router(
+            model_list=[{"model_name": "gpt-4o", "litellm_params": {"model": "gpt-4o"}}],
+            redis_host="localhost",
+            redis_port=6379,
+            cache_kwargs={"type": "disk", "socket_timeout": 5},
+        )
+
+    assert created_caches == [{"host": "localhost", "port": "6379", "socket_timeout": 5}]
+
+
+def test_router_invalid_cache_kwargs_type_raises(reset_litellm_cache):
+    with pytest.raises(ValueError):
+        litellm.Router(
+            model_list=[{"model_name": "gpt-4o", "litellm_params": {"model": "gpt-4o"}}],
+            cache_responses=True,
+            cache_kwargs={"type": "not-a-cache"},
+        )