Litellm dev 02 10 2025 p1 (#8438)

* fix(azure/chat/gpt_transformation.py): fix str compare to use int - ensure correct api version check is done

Resolves https://github.com/BerriAI/litellm/issues/8241#issuecomment-2647142891

* test(test_azure_openai.py): add better testing

litellm/llms/azure/chat/gpt_transformation.py

@@ -113,6 +113,17 @@ class AzureOpenAIConfig(BaseConfig):
 
         return False
 
+    def _is_response_format_supported_api_version(
+        self, api_version_year: str, api_version_month: str
+    ) -> bool:
+        """
+        - check if api_version is supported for response_format
+        """
+
+        is_supported = int(api_version_year) <= 2024 and int(api_version_month) >= 8
+
+        return is_supported
+
     def map_openai_params(
         self,
         non_default_params: dict,
@@ -171,13 +182,20 @@ class AzureOpenAIConfig(BaseConfig):
                 _is_response_format_supported_model = (
                     self._is_response_format_supported_model(model)
                 )
-                should_convert_response_format_to_tool = (
-                    api_version_year <= "2024" and api_version_month < "08"
-                ) or not _is_response_format_supported_model
+
+                is_response_format_supported_api_version = (
+                    self._is_response_format_supported_api_version(
+                        api_version_year, api_version_month
+                    )
+                )
+                is_response_format_supported = (
+                    is_response_format_supported_api_version
+                    and _is_response_format_supported_model
+                )
                 optional_params = self._add_response_format_to_tools(
                     optional_params=optional_params,
                     value=value,
-                    should_convert_response_format_to_tool=should_convert_response_format_to_tool,
+                    is_response_format_supported=is_response_format_supported,
                 )
             elif param == "tools" and isinstance(value, list):
                 optional_params.setdefault("tools", [])

litellm/llms/base_llm/chat/transformation.py

@@ -20,6 +20,7 @@ from pydantic import BaseModel
 
 from litellm._logging import verbose_logger
 from litellm.constants import RESPONSE_FORMAT_TOOL_NAME
+from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
 from litellm.types.llms.openai import (
     AllMessageValues,
     ChatCompletionToolChoiceFunctionParam,
@@ -27,9 +28,6 @@ from litellm.types.llms.openai import (
     ChatCompletionToolParam,
     ChatCompletionToolParamFunctionChunk,
 )
-
-from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
-
 from litellm.types.utils import ModelResponse
 from litellm.utils import CustomStreamWrapper
 
@@ -163,7 +161,7 @@ class BaseConfig(ABC):
         self,
         optional_params: dict,
         value: dict,
-        should_convert_response_format_to_tool: bool,
+        is_response_format_supported: bool,
     ) -> dict:
         """
         Follow similar approach to anthropic - translate to a single tool call.
@@ -183,7 +181,8 @@ class BaseConfig(ABC):
         elif "json_schema" in value:
             json_schema = value["json_schema"]["schema"]
 
-        if json_schema and should_convert_response_format_to_tool:
+        if json_schema and not is_response_format_supported:
+
             _tool_choice = ChatCompletionToolChoiceObjectParam(
                 type="function",
                 function=ChatCompletionToolChoiceFunctionParam(

tests/llm_translation/test_azure_openai.py

@@ -285,12 +285,27 @@ def test_azure_openai_gpt_4o_naming(monkeypatch):
         assert "tool_calls" not in mock_post.call_args.kwargs
 
 
-def test_azure_gpt_4o_with_tool_call_and_response_format():
+@pytest.mark.parametrize(
+    "api_version",
+    [
+        "2024-10-21",
+        # "2024-02-15-preview",
+    ],
+)
+def test_azure_gpt_4o_with_tool_call_and_response_format(api_version):
     from litellm import completion
     from typing import Optional
     from pydantic import BaseModel
     import litellm
 
+    from openai import AzureOpenAI
+
+    client = AzureOpenAI(
+        api_key="fake-key",
+        base_url="https://fake-azure.openai.azure.com",
+        api_version=api_version,
+    )
+
     class InvestigationOutput(BaseModel):
         alert_explanation: Optional[str] = None
         investigation: Optional[str] = None
@@ -322,6 +337,7 @@ def test_azure_gpt_4o_with_tool_call_and_response_format():
         }
     ]
 
+    with patch.object(client.chat.completions.with_raw_response, "create") as mock_post:
         response = litellm.completion(
             model="azure/gpt-4o",
             messages=[
@@ -329,18 +345,26 @@ def test_azure_gpt_4o_with_tool_call_and_response_format():
                     "role": "system",
                     "content": "You are a tool-calling AI assist provided with common devops and IT tools that you can use to troubleshoot problems or answer questions.\nWhenever possible you MUST first use tools to investigate then answer the question.",
                 },
-            {"role": "user", "content": "What is the current date and time in NYC?"},
+                {
+                    "role": "user",
+                    "content": "What is the current date and time in NYC?",
+                },
             ],
             drop_params=True,
             temperature=0.00000001,
             tools=tools,
             tool_choice="auto",
             response_format=InvestigationOutput,  # commenting this line will cause the output to be correct
+            api_version=api_version,
+            client=client,
         )
 
-    assert response.choices[0].finish_reason == "tool_calls"
+        mock_post.assert_called_once()
 
-    print(response.to_json())
+        if api_version == "2024-10-21":
+            assert "response_format" in mock_post.call_args.kwargs
+        else:
+            assert "response_format" not in mock_post.call_args.kwargs
 
 
 def test_map_openai_params():