Merge 8a7b51ca4e into aeede53112

litellm/llms/openai/responses/transformation.py

@@ -373,6 +373,20 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
     def _model_map_lookup_name(model: str) -> str:
         return model
 
+    @staticmethod
+    def _without_unsupported_o_series_temperature(
+        params: Mapping[str, object], model: str, drop_params: bool
+    ) -> Mapping[str, object]:
+        temperature: Final = params.get("temperature")
+        if (
+            litellm.openaiOSeriesConfig.is_model_o_series_model(model=model)
+            and temperature is not None
+            and temperature != 1
+            and (drop_params or litellm.drop_params)
+        ):
+            return {key: value for key, value in params.items() if key != "temperature"}
+        return params
+
     @staticmethod
     def _supports_reasoning_effort_none(model: str) -> bool:
         """Return True if the model supports reasoning.effort='none'."""
@@ -589,8 +603,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
                         ),
                         status_code=400,
                     )
-
-        return params
+        return self._without_unsupported_o_series_temperature(params, lookup_name, drop_params)
 
     def transform_responses_api_request(
         self,

tests/unit/llms/openai/responses/test_openai_responses_transformation.py

@@ -2393,6 +2393,33 @@ class TestPromptCacheOptionsOnResponsesPath:
         }
 
 
+@pytest.mark.parametrize(
+    ("model", "temperature", "drop_params", "global_drop_params", "expected"),
+    (
+        ("o3-mini", 0.5, True, False, {"max_output_tokens": 128}),
+        ("o3-mini", 1, True, False, {"temperature": 1, "max_output_tokens": 128}),
+        ("o3-mini", 0.5, False, True, {"max_output_tokens": 128}),
+        ("o3-mini", 0.5, False, False, {"temperature": 0.5, "max_output_tokens": 128}),
+        ("gpt-4o", 0.5, True, False, {"temperature": 0.5, "max_output_tokens": 128}),
+    ),
+)
+def test_o_series_responses_drops_only_unsupported_temperature_when_enabled(
+    model: str,
+    temperature: float,
+    drop_params: bool,
+    global_drop_params: bool,
+    expected: Mapping[str, object],
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setattr(litellm, "drop_params", global_drop_params)
+    mapped: Final = OpenAIResponsesAPIConfig().map_openai_params(
+        response_api_optional_params={"temperature": temperature, "max_output_tokens": 128},
+        model=model,
+        drop_params=drop_params,
+    )
+    assert mapped == expected
+
+
 class TestResponsesSurfaceSharesTheEffortRule:
     """The Responses API reaches the same gpt-5 models over a different wire, and the default
     /v1/messages bridge for openai models routes through it. It carried its own copy of the