feat(caching.py): enable caching on provider-specific optional params

Closes https://github.com/BerriAI/litellm/issues/5049

litellm/__init__.py

@@ -146,6 +146,9 @@ return_response_headers: bool = (
 )
 ##################
 logging: bool = True
+enable_caching_on_optional_params: bool = (
+    False  # feature-flag for caching on optional params - e.g. 'top_k'
+)
 caching: bool = (
     False  # Not used anymore, will be removed in next MAJOR release - https://github.com/BerriAI/litellm/discussions/648
 )

litellm/caching.py

@@ -23,6 +23,7 @@ import litellm
 from litellm._logging import verbose_logger
 from litellm.litellm_core_utils.core_helpers import _get_parent_otel_span_from_kwargs
 from litellm.types.services import ServiceLoggerPayload, ServiceTypes
+from litellm.types.utils import all_litellm_params
 
 
 def print_verbose(print_statement):
@@ -1838,6 +1839,7 @@ class Cache:
             "seed",
             "tools",
             "tool_choice",
+            "stream",
         ]
         embedding_only_kwargs = [
             "input",
@@ -1851,9 +1853,9 @@ class Cache:
         combined_kwargs = (
             completion_kwargs + embedding_only_kwargs + transcription_only_kwargs
         )
-        for param in combined_kwargs:
+        litellm_param_kwargs = all_litellm_params
-            # ignore litellm params here
+        for param in kwargs:
-            if param in kwargs:
+            if param in combined_kwargs:
                 # check if param == model and model_group is passed in, then override model with model_group
                 if param == "model":
                     model_group = None
@@ -1897,6 +1899,17 @@ class Cache:
                         continue  # ignore None params
                     param_value = kwargs[param]
                 cache_key += f"{str(param)}: {str(param_value)}"
+            elif (
+                param not in litellm_param_kwargs
+            ):  # check if user passed in optional param - e.g. top_k
+                if (
+                    litellm.enable_caching_on_optional_params is True
+                ):  # feature flagged for now
+                    if kwargs[param] is None:
+                        continue  # ignore None params
+                    param_value = kwargs[param]
+                    cache_key += f"{str(param)}: {str(param_value)}"
+
         print_verbose(f"\nCreated cache key: {cache_key}")
         # Use hashlib to create a sha256 hash of the cache key
         hash_object = hashlib.sha256(cache_key.encode())
@@ -2101,9 +2114,7 @@ class Cache:
         try:
             cache_list = []
             for idx, i in enumerate(kwargs["input"]):
-                preset_cache_key = litellm.cache.get_cache_key(
+                preset_cache_key = self.get_cache_key(*args, **{**kwargs, "input": i})
-                    *args, **{**kwargs, "input": i}
-                )
                 kwargs["cache_key"] = preset_cache_key
                 embedding_response = result.data[idx]
                 cache_key, cached_data, kwargs = self._add_cache_logic(

litellm/main.py

@@ -125,7 +125,11 @@ from .llms.vertex_ai_partner import VertexAIPartnerModels
 from .llms.vertex_httpx import VertexLLM
 from .llms.watsonx import IBMWatsonXAI
 from .types.llms.openai import HttpxBinaryResponseContent
-from .types.utils import AdapterCompletionStreamWrapper, ChatCompletionMessageToolCall
+from .types.utils import (
+    AdapterCompletionStreamWrapper,
+    ChatCompletionMessageToolCall,
+    all_litellm_params,
+)
 
 encoding = tiktoken.get_encoding("cl100k_base")
 from litellm.utils import (
@@ -744,64 +748,9 @@ def completion(
         "top_logprobs",
         "extra_headers",
     ]
-    litellm_params = [
+    litellm_params = (
-        "metadata",
+        all_litellm_params  # use the external var., used in creating cache key as well.
-        "tags",
+    )
-        "acompletion",
-        "atext_completion",
-        "text_completion",
-        "caching",
-        "mock_response",
-        "api_key",
-        "api_version",
-        "api_base",
-        "force_timeout",
-        "logger_fn",
-        "verbose",
-        "custom_llm_provider",
-        "litellm_logging_obj",
-        "litellm_call_id",
-        "use_client",
-        "id",
-        "fallbacks",
-        "azure",
-        "headers",
-        "model_list",
-        "num_retries",
-        "context_window_fallback_dict",
-        "retry_policy",
-        "roles",
-        "final_prompt_value",
-        "bos_token",
-        "eos_token",
-        "request_timeout",
-        "complete_response",
-        "self",
-        "client",
-        "rpm",
-        "tpm",
-        "max_parallel_requests",
-        "input_cost_per_token",
-        "output_cost_per_token",
-        "input_cost_per_second",
-        "output_cost_per_second",
-        "hf_model_name",
-        "model_info",
-        "proxy_server_request",
-        "preset_cache_key",
-        "caching_groups",
-        "ttl",
-        "cache",
-        "no-log",
-        "base_model",
-        "stream_timeout",
-        "supports_system_message",
-        "region_name",
-        "allowed_model_region",
-        "model_config",
-        "fastest_response",
-        "cooldown_time",
-    ]
 
     default_params = openai_params + litellm_params
     non_default_params = {
@@ -5205,7 +5154,7 @@ def stream_chunk_builder(
         response["choices"][0]["message"]["function_call"][
             "arguments"
         ] = combined_arguments
-    
+
     content_chunks = [
         chunk
         for chunk in chunks

litellm/tests/test_caching.py

@@ -207,11 +207,17 @@ async def test_caching_with_cache_controls(sync_flag):
         else:
             ## TTL = 0
             response1 = await litellm.acompletion(
-                model="gpt-3.5-turbo", messages=messages, cache={"ttl": 0}
+                model="gpt-3.5-turbo",
+                messages=messages,
+                cache={"ttl": 0},
+                mock_response="Hello world",
             )
             await asyncio.sleep(10)
             response2 = await litellm.acompletion(
-                model="gpt-3.5-turbo", messages=messages, cache={"s-maxage": 10}
+                model="gpt-3.5-turbo",
+                messages=messages,
+                cache={"s-maxage": 10},
+                mock_response="Hello world",
             )
 
             assert response2["id"] != response1["id"]
@@ -220,21 +226,33 @@ async def test_caching_with_cache_controls(sync_flag):
         ## TTL = 5
         if sync_flag:
             response1 = completion(
-                model="gpt-3.5-turbo", messages=messages, cache={"ttl": 5}
+                model="gpt-3.5-turbo",
+                messages=messages,
+                cache={"ttl": 5},
+                mock_response="Hello world",
             )
             response2 = completion(
-                model="gpt-3.5-turbo", messages=messages, cache={"s-maxage": 5}
+                model="gpt-3.5-turbo",
+                messages=messages,
+                cache={"s-maxage": 5},
+                mock_response="Hello world",
             )
             print(f"response1: {response1}")
             print(f"response2: {response2}")
             assert response2["id"] == response1["id"]
         else:
             response1 = await litellm.acompletion(
-                model="gpt-3.5-turbo", messages=messages, cache={"ttl": 25}
+                model="gpt-3.5-turbo",
+                messages=messages,
+                cache={"ttl": 25},
+                mock_response="Hello world",
             )
             await asyncio.sleep(10)
             response2 = await litellm.acompletion(
-                model="gpt-3.5-turbo", messages=messages, cache={"s-maxage": 25}
+                model="gpt-3.5-turbo",
+                messages=messages,
+                cache={"s-maxage": 25},
+                mock_response="Hello world",
             )
             print(f"response1: {response1}")
             print(f"response2: {response2}")
@@ -282,6 +300,61 @@ def test_caching_with_models_v2():
 
 # test_caching_with_models_v2()
 
+
+def test_caching_with_optional_params():
+    litellm.enable_caching_on_optional_params = True
+    messages = [
+        {"role": "user", "content": "who is ishaan CTO of litellm from litellm 2023"}
+    ]
+    litellm.cache = Cache()
+    print("test2 for caching")
+    litellm.set_verbose = True
+
+    response1 = completion(
+        model="gpt-3.5-turbo",
+        messages=messages,
+        top_k=10,
+        caching=True,
+        mock_response="Hello: {}".format(uuid.uuid4()),
+    )
+    response2 = completion(
+        model="gpt-3.5-turbo",
+        messages=messages,
+        top_k=10,
+        caching=True,
+        mock_response="Hello: {}".format(uuid.uuid4()),
+    )
+    response3 = completion(
+        model="gpt-3.5-turbo",
+        messages=messages,
+        top_k=9,
+        caching=True,
+        mock_response="Hello: {}".format(uuid.uuid4()),
+    )
+    print(f"response1: {response1}")
+    print(f"response2: {response2}")
+    print(f"response3: {response3}")
+    litellm.cache = None
+    litellm.success_callback = []
+    litellm._async_success_callback = []
+    if (
+        response3["choices"][0]["message"]["content"]
+        == response2["choices"][0]["message"]["content"]
+    ):
+        # if models are different, it should not return cached response
+        print(f"response2: {response2}")
+        print(f"response3: {response3}")
+        pytest.fail(f"Error occurred:")
+    if (
+        response1["choices"][0]["message"]["content"]
+        != response2["choices"][0]["message"]["content"]
+    ):
+        print(f"response1: {response1}")
+        print(f"response2: {response2}")
+        pytest.fail(f"Error occurred:")
+    litellm.enable_caching_on_optional_params = False
+
+
 embedding_large_text = (
     """
 small text
@@ -1347,7 +1420,7 @@ def test_get_cache_key():
                 "litellm_logging_obj": {},
             }
         )
-        cache_key_str = "model: gpt-3.5-turbomessages: [{'role': 'user', 'content': 'write a one sentence poem about: 7510'}]temperature: 0.2max_tokens: 40"
+        cache_key_str = "model: gpt-3.5-turbomessages: [{'role': 'user', 'content': 'write a one sentence poem about: 7510'}]max_tokens: 40temperature: 0.2stream: True"
         hash_object = hashlib.sha256(cache_key_str.encode())
         # Hexadecimal representation of the hash
         hash_hex = hash_object.hexdigest()

litellm/types/utils.py

@@ -1052,6 +1052,68 @@ class ResponseFormatChunk(TypedDict, total=False):
     response_schema: dict
 
 
+all_litellm_params = [
+    "metadata",
+    "tags",
+    "acompletion",
+    "atext_completion",
+    "text_completion",
+    "caching",
+    "mock_response",
+    "api_key",
+    "api_version",
+    "api_base",
+    "force_timeout",
+    "logger_fn",
+    "verbose",
+    "custom_llm_provider",
+    "litellm_logging_obj",
+    "litellm_call_id",
+    "use_client",
+    "id",
+    "fallbacks",
+    "azure",
+    "headers",
+    "model_list",
+    "num_retries",
+    "context_window_fallback_dict",
+    "retry_policy",
+    "roles",
+    "final_prompt_value",
+    "bos_token",
+    "eos_token",
+    "request_timeout",
+    "complete_response",
+    "self",
+    "client",
+    "rpm",
+    "tpm",
+    "max_parallel_requests",
+    "input_cost_per_token",
+    "output_cost_per_token",
+    "input_cost_per_second",
+    "output_cost_per_second",
+    "hf_model_name",
+    "model_info",
+    "proxy_server_request",
+    "preset_cache_key",
+    "caching_groups",
+    "ttl",
+    "cache",
+    "no-log",
+    "base_model",
+    "stream_timeout",
+    "supports_system_message",
+    "region_name",
+    "allowed_model_region",
+    "model_config",
+    "fastest_response",
+    "cooldown_time",
+    "cache_key",
+    "max_retries",
+]
+
+
 class LoggedLiteLLMParams(TypedDict, total=False):
     force_timeout: Optional[float]
     custom_llm_provider: Optional[str]

litellm/utils.py

@@ -1084,7 +1084,7 @@ def client(original_function):
                     and str(original_function.__name__)
                     in litellm.cache.supported_call_types
                 ):
-                    print_verbose(f"Checking Cache")
+                    print_verbose("Checking Cache")
                     if call_type == CallTypes.aembedding.value and isinstance(
                         kwargs["input"], list
                     ):