Merge pull request #41469 from BerriAI/litellm_bedrock_mantle_responses_drop_top_p

fix(responses): drop top_p for gpt-5 reasoning models when drop_params is set

litellm/llms/bedrock_mantle/responses/transformation.py

@@ -344,6 +344,10 @@ class BedrockMantleResponsesAPIConfig(BedrockMantleAuthMixin, OpenAIResponsesAPI
         kept: Final = [item for item, _ in normalized if item is not None]  # mutable-ok: ResponseInputParam is a list
         return kept  # pyright: ignore[reportReturnType]  # Codex passthrough items sit outside the OpenAI input union
 
+    @staticmethod
+    def _model_map_lookup_name(model: str) -> str:
+        return model.split("/")[-1].removeprefix("openai.")
+
     def map_openai_params(
         self,
         response_api_optional_params: ResponsesAPIOptionalRequestParams,

litellm/llms/openai/responses/transformation.py

@@ -125,6 +125,10 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
             return False
         return is_gpt_reasoning_series_name(model)
 
+    @staticmethod
+    def _model_map_lookup_name(model: str) -> str:
+        return model
+
     @staticmethod
     def _supports_reasoning_effort_none(model: str) -> bool:
         """Return True if the model supports reasoning.effort='none'."""
@@ -208,8 +212,9 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
     ) -> dict:
         """No mapping applied since inputs are in OpenAI spec already.
 
-        GPT-5 models have restrictions on temperature (only temperature=1
-        is accepted unless reasoning_effort='none' on models that support it).
+        GPT-5 models have restrictions on temperature and top_p (only temperature=1
+        is accepted, and top_p is rejected, unless reasoning.effort resolves to
+        'none' on models that support it).
         Apply the same validation used by the chat completions path.
         """
         params: Final = dict(response_api_optional_params)
@@ -234,13 +239,16 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
                     status_code=400,
                 )
 
-        if self._is_gpt_5_model(model=model):
+        lookup_name: Final = self._model_map_lookup_name(model)
+        if self._is_gpt_5_model(model=lookup_name):
+            reasoning: Final = params.get("reasoning") or {}
+            effort: Final = reasoning.get("effort") if isinstance(reasoning, dict) else None
+            supports_none: Final = self._supports_reasoning_effort_none(model=lookup_name)
+            effort_is_none: Final = supports_none and self._effort_resolves_to_none(lookup_name, effort)
+
             temperature: Final = params.get("temperature")
             if temperature is not None and temperature != 1:
-                reasoning: Final = params.get("reasoning") or {}
-                effort: Final = reasoning.get("effort") if isinstance(reasoning, dict) else None
-                supports_none: Final = self._supports_reasoning_effort_none(model=model)
-                if supports_none and self._effort_resolves_to_none(model, effort):
+                if effort_is_none:
                     pass  # flexible temperature allowed
                 elif drop_params or litellm.drop_params:
                     params.pop("temperature", None)
@@ -256,6 +264,20 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
                         status_code=400,
                     )
 
+            if "top_p" in params and not effort_is_none:
+                if drop_params or litellm.drop_params:
+                    params.pop("top_p", None)
+                else:
+                    raise litellm.UnsupportedParamsError(
+                        message=(
+                            f"{model} only supports top_p when reasoning.effort resolves to 'none', "
+                            "either set explicitly on the request or declared as the model's "
+                            "default_reasoning_effort. "
+                            "To drop unsupported params set `litellm.drop_params = True`"
+                        ),
+                        status_code=400,
+                    )
+
         return params
 
     def transform_responses_api_request(

tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py

@@ -369,6 +369,52 @@ class TestBedrockMantleResponsesTools:
         assert "file_search" in str(mock_warning.call_args)
 
 
+class TestBedrockMantleSamplingParams:
+    """Mantle serves OpenAI's gpt-5 models under their OpenAI sampling rule: top_p and a
+    non-default temperature are accepted only when reasoning.effort resolves to none, so
+    the `openai.` catalogue name (region-prefixed on GovCloud) must answer from the OpenAI
+    model's map entry instead of dropping both params on every request."""
+
+    @pytest.mark.parametrize(
+        "model, effort, survives",
+        [
+            ("openai.gpt-5.4", None, True),
+            ("openai.gpt-5.5", None, False),
+            ("openai.gpt-5.6-luna", None, False),
+            ("openai.gpt-5.6-luna", "none", True),
+            ("openai.gpt-5.6-luna", "low", False),
+            ("us-gov-west-1/openai.gpt-5.4", None, True),
+            ("us-gov-west-1/openai.gpt-5.6-luna", None, False),
+        ],
+    )
+    def test_top_p_and_temperature_follow_the_resolved_effort(self, local_cost_map, model, effort, survives):
+        params = {"top_p": 0.9, "temperature": 0.2}
+        if effort is not None:
+            params["reasoning"] = {"effort": effort}
+        mapped = BedrockMantleResponsesAPIConfig().map_openai_params(
+            response_api_optional_params=params,
+            model=model,
+            drop_params=True,
+        )
+        assert ("top_p" in mapped) is survives
+        assert ("temperature" in mapped) is survives
+
+    def test_top_p_without_drop_params_raises_only_while_reasoning_is_active(self, local_cost_map):
+        with pytest.raises(litellm.UnsupportedParamsError):
+            BedrockMantleResponsesAPIConfig().map_openai_params(
+                response_api_optional_params={"top_p": 0.9},
+                model="openai.gpt-5.6-luna",
+                drop_params=False,
+            )
+
+        mapped = BedrockMantleResponsesAPIConfig().map_openai_params(
+            response_api_optional_params={"top_p": 0.9},
+            model="openai.gpt-5.4",
+            drop_params=False,
+        )
+        assert mapped["top_p"] == 0.9
+
+
 class TestBedrockMantleResponsesWebSearch:
     """Web Search on Amazon Bedrock is a server-side built-in tool that Mantle runs
     itself when the caller passes {"type": "web_search"} on the Responses path, so

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

@@ -1835,6 +1835,46 @@ class TestResponsesSurfaceSharesTheEffortRule:
         )
         assert ("temperature" in mapped) is temperature_survives
 
+    @pytest.mark.parametrize(
+        "model, effort, top_p_survives",
+        [
+            ("gpt-5.1", None, True),
+            ("gpt-5.4", None, True),
+            ("gpt-5.5", None, False),
+            ("gpt-5.6-terra", None, False),
+            ("gpt-5.6-sol", None, False),
+            ("gpt-5.6-terra", "none", True),
+            ("gpt-5.6-terra", "medium", False),
+            ("gpt-6-astra", None, False),
+            ("gpt-6-astra", "low", False),
+        ],
+    )
+    def test_top_p_follows_the_resolved_effort(self, local_model_cost_map, model, effort, top_p_survives):
+        params = {"top_p": 0.9}
+        if effort is not None:
+            params["reasoning"] = {"effort": effort}
+        mapped = OpenAIResponsesAPIConfig().map_openai_params(
+            response_api_optional_params=params,
+            model=model,
+            drop_params=True,
+        )
+        assert ("top_p" in mapped) is top_p_survives
+
+    def test_top_p_raises_without_drop_params(self, local_model_cost_map):
+        with pytest.raises(litellm.UnsupportedParamsError):
+            OpenAIResponsesAPIConfig().map_openai_params(
+                response_api_optional_params={"top_p": 0.9},
+                model="gpt-5.5",
+                drop_params=False,
+            )
+
+        mapped = OpenAIResponsesAPIConfig().map_openai_params(
+            response_api_optional_params={"top_p": 0.9, "reasoning": {"effort": "none"}},
+            model="gpt-5.6-terra",
+            drop_params=False,
+        )
+        assert mapped["top_p"] == 0.9
+
 
 class TestFlattenToolSchemaCombinatorsWiring:
     """Regression tests for MCP tools with a top-level anyOf schema (Codex Desktop).