fix(o_series_transformation.py): fix optional param check for o-serie… (#8787)

* fix(o_series_transformation.py): fix optional param check for o-series models

o3-mini and o-1 do not support parallel tool calling

* fix(utils.py): support 'drop_params' for 'thinking' param across models

allows switching to older claude versions (or non-anthropic models) and param to be safely dropped

* fix: fix passing thinking param in optional params

allows dropping thinking_param where not applicable

* test: update old model

* fix(utils.py): fix linting errors

* fix(main.py): add param to acompletion

litellm/constants.py

@@ -120,6 +120,7 @@ OPENAI_CHAT_COMPLETION_PARAMS = [
     "top_logprobs",
     "reasoning_effort",
     "extra_headers",
+    "thinking",
 ]
 
 openai_compatible_endpoints: List = [

litellm/llms/anthropic/chat/transformation.py

@@ -80,7 +80,7 @@ class AnthropicConfig(BaseConfig):
         return super().get_config()
 
     def get_supported_openai_params(self, model: str):
-        return [
+        params = [
             "stream",
             "stop",
             "temperature",
@@ -95,6 +95,11 @@ class AnthropicConfig(BaseConfig):
             "user",
         ]
 
+        if "claude-3-7-sonnet" in model:
+            params.append("thinking")
+
+        return params
+
     def get_json_schema_from_pydantic_object(
         self, response_format: Union[Any, Dict, None]
     ) -> Optional[dict]:
@@ -302,6 +307,7 @@ class AnthropicConfig(BaseConfig):
         model: str,
         drop_params: bool,
     ) -> dict:
+
         for param, value in non_default_params.items():
             if param == "max_tokens":
                 optional_params["max_tokens"] = value
@@ -358,7 +364,8 @@ class AnthropicConfig(BaseConfig):
                 optional_params["json_mode"] = True
             if param == "user":
                 optional_params["metadata"] = {"user_id": value}
-
+            if param == "thinking":
+                optional_params["thinking"] = value
         return optional_params
 
     def _create_json_tool_call_for_response_format(

litellm/llms/openai/chat/o_series_transformation.py

@@ -19,6 +19,7 @@ from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider
 from litellm.types.llms.openai import AllMessageValues, ChatCompletionUserMessage
 from litellm.utils import (
     supports_function_calling,
+    supports_parallel_function_calling,
     supports_response_schema,
     supports_system_messages,
 )
@@ -76,14 +77,19 @@ class OpenAIOSeriesConfig(OpenAIGPTConfig):
             model, custom_llm_provider
         )
         _supports_response_schema = supports_response_schema(model, custom_llm_provider)
+        _supports_parallel_tool_calls = supports_parallel_function_calling(
+            model, custom_llm_provider
+        )
 
         if not _supports_function_calling:
             non_supported_params.append("tools")
             non_supported_params.append("tool_choice")
-            non_supported_params.append("parallel_tool_calls")
             non_supported_params.append("function_call")
             non_supported_params.append("functions")
 
+        if not _supports_parallel_tool_calls:
+            non_supported_params.append("parallel_tool_calls")
+
         if not _supports_response_schema:
             non_supported_params.append("response_format")
 

litellm/main.py

@@ -166,6 +166,7 @@ from .llms.vertex_ai.vertex_model_garden.main import VertexAIModelGardenModels
 from .llms.vllm.completion import handler as vllm_handler
 from .llms.watsonx.chat.handler import WatsonXChatHandler
 from .llms.watsonx.common_utils import IBMWatsonXMixin
+from .types.llms.anthropic import AnthropicThinkingParam
 from .types.llms.openai import (
     ChatCompletionAssistantMessage,
     ChatCompletionAudioParam,
@@ -341,6 +342,7 @@ async def acompletion(
     model_list: Optional[list] = None,  # pass in a list of api_base,keys, etc.
     extra_headers: Optional[dict] = None,
     # Optional liteLLM function params
+    thinking: Optional[AnthropicThinkingParam] = None,
     **kwargs,
 ) -> Union[ModelResponse, CustomStreamWrapper]:
     """
@@ -800,6 +802,7 @@ def completion(  # type: ignore # noqa: PLR0915
     api_key: Optional[str] = None,
     model_list: Optional[list] = None,  # pass in a list of api_base,keys, etc.
     # Optional liteLLM function params
+    thinking: Optional[AnthropicThinkingParam] = None,
     **kwargs,
 ) -> Union[ModelResponse, CustomStreamWrapper]:
     """
@@ -1106,6 +1109,7 @@ def completion(  # type: ignore # noqa: PLR0915
             parallel_tool_calls=parallel_tool_calls,
             messages=messages,
             reasoning_effort=reasoning_effort,
+            thinking=thinking,
             **non_default_params,
         )
 

litellm/types/llms/anthropic.py

@@ -359,3 +359,8 @@ ANTHROPIC_API_HEADERS = {
 ANTHROPIC_API_ONLY_HEADERS = {  # fails if calling anthropic on vertex ai / bedrock
     "anthropic-beta",
 }
+
+
+class AnthropicThinkingParam(TypedDict, total=False):
+    type: Literal["enabled"]
+    budget_tokens: int

litellm/types/llms/openai.py

@@ -383,6 +383,7 @@ class ChatCompletionImageObject(TypedDict):
     type: Literal["image_url"]
     image_url: Union[str, ChatCompletionImageUrlObject]
 
+
 class ChatCompletionVideoUrlObject(TypedDict, total=False):
     url: Required[str]
     detail: str

litellm/utils.py

@@ -119,7 +119,10 @@ from litellm.router_utils.get_retry_from_policy import (
     reset_retry_policy,
 )
 from litellm.secret_managers.main import get_secret
-from litellm.types.llms.anthropic import ANTHROPIC_API_ONLY_HEADERS
+from litellm.types.llms.anthropic import (
+    ANTHROPIC_API_ONLY_HEADERS,
+    AnthropicThinkingParam,
+)
 from litellm.types.llms.openai import (
     AllMessageValues,
     AllPromptValues,
@@ -1969,6 +1972,19 @@ def supports_response_schema(
     )
 
 
+def supports_parallel_function_calling(
+    model: str, custom_llm_provider: Optional[str] = None
+) -> bool:
+    """
+    Check if the given model supports parallel tool calls and return a boolean value.
+    """
+    return _supports_factory(
+        model=model,
+        custom_llm_provider=custom_llm_provider,
+        key="supports_parallel_function_calling",
+    )
+
+
 def supports_function_calling(
     model: str, custom_llm_provider: Optional[str] = None
 ) -> bool:
@@ -2118,30 +2134,6 @@ def supports_embedding_image_input(
     )
 
 
-def supports_parallel_function_calling(model: str):
-    """
-    Check if the given model supports parallel function calling and return True if it does, False otherwise.
-
-    Parameters:
-        model (str): The model to check for support of parallel function calling.
-
-    Returns:
-        bool: True if the model supports parallel function calling, False otherwise.
-
-    Raises:
-        Exception: If the model is not found in the model_cost dictionary.
-    """
-    if model in litellm.model_cost:
-        model_info = litellm.model_cost[model]
-        if model_info.get("supports_parallel_function_calling", False) is True:
-            return True
-        return False
-    else:
-        raise Exception(
-            f"Model not supports parallel function calling. You passed model={model}."
-        )
-
-
 ####### HELPER FUNCTIONS ################
 def _update_dictionary(existing_dict: Dict, new_dict: dict) -> dict:
     for k, v in new_dict.items():
@@ -2752,6 +2744,7 @@ def get_optional_params(  # noqa: PLR0915
     reasoning_effort=None,
     additional_drop_params=None,
     messages: Optional[List[AllMessageValues]] = None,
+    thinking: Optional[AnthropicThinkingParam] = None,
     **kwargs,
 ):
     # retrieve all parameters passed to the function
@@ -2836,9 +2829,11 @@ def get_optional_params(  # noqa: PLR0915
         "additional_drop_params": None,
         "messages": None,
         "reasoning_effort": None,
+        "thinking": None,
     }
 
     # filter out those parameters that were passed with non-default values
+
     non_default_params = {
         k: v
         for k, v in passed_params.items()

tests/llm_translation/test_anthropic_completion.py

@@ -1166,6 +1166,8 @@ def test_anthropic_citations_api_streaming():
 def test_anthropic_thinking_output():
     from litellm import completion
 
+    litellm._turn_on_debug()
+
     resp = completion(
         model="anthropic/claude-3-7-sonnet-20250219",
         messages=[{"role": "user", "content": "What is the capital of France?"}],

tests/llm_translation/test_openai.py

@@ -338,3 +338,18 @@ def test_openai_max_retries_0(mock_get_openai_client):
 
     mock_get_openai_client.assert_called_once()
     assert mock_get_openai_client.call_args.kwargs["max_retries"] == 0
+
+
+@pytest.mark.parametrize("model", ["o1", "o1-preview", "o1-mini", "o3-mini"])
+def test_o1_parallel_tool_calls(model):
+    litellm.completion(
+        model=model,
+        messages=[
+            {
+                "role": "user",
+                "content": "foo",
+            }
+        ],
+        parallel_tool_calls=True,
+        drop_params=True,
+    )

tests/llm_translation/test_optional_params.py

@@ -1069,7 +1069,6 @@ def test_gemini_frequency_penalty():
     assert optional_params["frequency_penalty"] == 0.5
 
 
-
 def test_azure_prediction_param():
     optional_params = get_optional_params(
         model="chatgpt-v2",
@@ -1084,6 +1083,7 @@ def test_azure_prediction_param():
         "content": "LiteLLM is a very useful way to connect to a variety of LLMs.",
     }
 
+
 def test_vertex_ai_ft_llama():
     optional_params = get_optional_params(
         model="1984786713414729728",
@@ -1093,3 +1093,24 @@ def test_vertex_ai_ft_llama():
     )
     assert optional_params["frequency_penalty"] == 0.5
     assert "max_retries" not in optional_params
+
+
+@pytest.mark.parametrize(
+    "model, expected_thinking",
+    [
+        ("claude-3-5-sonnet", False),
+        ("claude-3-7-sonnet", True),
+        ("gpt-3.5-turbo", False),
+    ],
+)
+def test_anthropic_thinking_param(model, expected_thinking):
+    optional_params = get_optional_params(
+        model=model,
+        custom_llm_provider="anthropic",
+        thinking={"type": "enabled", "budget_tokens": 1024},
+        drop_params=True,
+    )
+    if expected_thinking:
+        assert "thinking" in optional_params
+    else:
+        assert "thinking" not in optional_params

tests/local_testing/test_streaming.py

@@ -4072,7 +4072,7 @@ def test_mock_response_iterator_tool_use():
         "anthropic/claude-3-7-sonnet-20250219",
     ],
 )
-def test_deepseek_reasoning_content_completion(model):
+def test_reasoning_content_completion(model):
     # litellm.set_verbose = True
     try:
         # litellm._turn_on_debug()
@@ -4081,7 +4081,6 @@ def test_deepseek_reasoning_content_completion(model):
             messages=[{"role": "user", "content": "Tell me a joke."}],
             stream=True,
             thinking={"type": "enabled", "budget_tokens": 1024},
-            timeout=5,
         )
 
         reasoning_content_exists = False