Merge 3cf31e0c4d into b781d157d7

litellm/caching/caching.py

@@ -41,6 +41,7 @@ from .redis_cluster_cache import RedisClusterCache
 from .redis_semantic_cache import RedisSemanticCache
 from .s3_cache import S3Cache
 
+_warned_dropped_cache_params: set = set()
 if TYPE_CHECKING:
     from litellm.rust_bridge.response_cache import NativeCacheRequest, ResponseCacheRuntime
 
@@ -400,6 +401,19 @@ class Cache:
                         continue  # ignore None params
                     param_value = kwargs[param]
                     cache_key += f"{param}: {param_value}"
+                elif (
+                    kwargs[param] is not None
+                    and param not in _warned_dropped_cache_params
+                    and len(_warned_dropped_cache_params) < 100
+                ):
+                    _warned_dropped_cache_params.add(param)
+                    verbose_logger.warning(
+                        "litellm.cache: provider-specific param '%s' is excluded from the cache key by default, "
+                        "so requests differing only in '%s' will return the same cached response. "
+                        "Set litellm.enable_caching_on_provider_specific_optional_params=True to include it in the key.",
+                        param,
+                        param,
+                    )
 
         if is_semantic_cache:
             cache_key += self._get_semantic_cache_tenant_scope(kwargs)

tests/local_testing/test_caching.py

@@ -24,6 +24,7 @@ import hashlib
 import random
 
 import pytest
+import logging
 
 import litellm
 from litellm import aembedding, completion, embedding
@@ -74,6 +75,28 @@ async def test_dual_cache_async_batch_get_cache():
         await dual_cache.async_batch_get_cache(keys=["test_value", "test_value_2"])
 
         assert mock_redis_cache.call_count == 1
+        
+        
+def test_cache_key_warns_on_dropped_provider_specific_param(caplog):
+    import litellm
+    from litellm.caching.caching import Cache, _warned_dropped_cache_params
+
+    litellm.enable_caching_on_provider_specific_optional_params = False
+    _warned_dropped_cache_params.discard("num_ctx")
+    cache = Cache()
+    try:
+        with caplog.at_level(logging.WARNING):
+            cache.get_cache_key(model="ollama/llama3.2",
+                messages=[{"role": "user", "content": "hello"}], num_ctx=2048)
+        assert any("num_ctx" in r.message for r in caplog.records)
+
+        caplog.clear()
+        with caplog.at_level(logging.WARNING):
+            cache.get_cache_key(model="ollama/llama3.2",
+                messages=[{"role": "user", "content": "hello"}], num_ctx=4096)
+        assert not any("num_ctx" in r.message for r in caplog.records)
+    finally:
+        _warned_dropped_cache_params.discard("num_ctx")
 
 
 def test_dual_cache_batch_get_cache():