From 88befe027b5c4efc834da5be8dff4de2eea25d72 Mon Sep 17 00:00:00 2001
From: HarshDubey23 <191850210+HarshDubey23@users.noreply.github.com>
Date: Tue, 27 Oct 2026 20:48:03 +0530
Subject: [PATCH] fix(hosted_vllm): drop unneeded allowed_openai_params change,
 test extra_body merge

The provider params reach the request through HostedVLLMResponsesAPIConfig
alone, so the change to get_requested_response_api_optional_param is not
needed for this fix. Also cover a request that sends its own extra_body:
both reach vLLM and extra_body wins on a shared key.
---
 litellm/responses/utils.py                    | 14 ++--------
 .../responses/test_hosted_vllm_responses.py   | 27 +++++++++++++++++++
 2 files changed, 29 insertions(+), 12 deletions(-)

diff --git a/litellm/responses/utils.py b/litellm/responses/utils.py
index 3fab03cc3b0..dc75e77d685 100644
--- a/litellm/responses/utils.py
+++ b/litellm/responses/utils.py
@@ -334,20 +334,10 @@ class ResponsesAPIRequestUtils:
         """
         from litellm.utils import PreProcessNonDefaultParams
 
-        valid_keys: Final[frozenset[str]] = frozenset(get_type_hints(ResponsesAPIOptionalRequestParams).keys())
+        valid_keys: Final = get_type_hints(ResponsesAPIOptionalRequestParams).keys()
         custom_llm_provider: Final = params.pop("custom_llm_provider", None)
         special_params: Final[dict[str, object]] = params.pop("kwargs", {})
 
-        allowed_openai_params: Final[object] = params.get("allowed_openai_params") or special_params.get(
-            "allowed_openai_params"
-        )
-        allowed_set: Final[frozenset[str]] = (
-            frozenset(p for p in allowed_openai_params if isinstance(p, str))  # pyright: ignore[reportUnknownVariableType]  # element type is dynamic
-            if isinstance(allowed_openai_params, (list, tuple, set, frozenset))
-            else frozenset()
-        )
-        default_keys: Final[frozenset[str]] = valid_keys | allowed_set
-
         additional_drop_params: Final[list[str] | None] = params.pop("additional_drop_params", None)
         non_default_params: Final = _apply_nested_drop_params(
             PreProcessNonDefaultParams.base_pre_process_non_default_params(
@@ -355,7 +345,7 @@ class ResponsesAPIRequestUtils:
                 special_params=special_params,
                 custom_llm_provider=custom_llm_provider,
                 additional_drop_params=additional_drop_params,
-                default_param_values={k: None for k in default_keys},
+                default_param_values={k: None for k in valid_keys},
                 additional_endpoint_specific_params=["input"],
             ),
             additional_drop_params,
diff --git a/tests/unit/llms/hosted_vllm/responses/test_hosted_vllm_responses.py b/tests/unit/llms/hosted_vllm/responses/test_hosted_vllm_responses.py
index 2f0302004de..a83ad492bdd 100644
--- a/tests/unit/llms/hosted_vllm/responses/test_hosted_vllm_responses.py
+++ b/tests/unit/llms/hosted_vllm/responses/test_hosted_vllm_responses.py
@@ -272,3 +272,30 @@ def test_hosted_vllm_responses_transform_request_from_optional_request_params():
 
     assert request["chat_template_kwargs"] == {"enable_thinking": False}
     assert request["top_k"] == 30
+
+
+def test_hosted_vllm_responses_provider_params_merge_with_request_extra_body():
+    """
+    A request-level extra_body must not drop the forwarded provider params:
+    unrelated keys from both reach vLLM, and extra_body wins on a shared key.
+    """
+    mock_client = _make_mock_http_client(_make_mock_responses_api_response("Test output"))
+
+    with patch(
+        "litellm.llms.custom_httpx.llm_http_handler.get_httpx_client",
+        return_value=mock_client,
+    ):
+        litellm.responses(
+            model="hosted_vllm/Qwen/Qwen3-8B",
+            input="hi",
+            api_base="https://test-vllm.example.com/v1",
+            api_key="test-key",
+            chat_template_kwargs={"enable_thinking": False},
+            top_k=20,
+            extra_body={"min_p": 0.1, "top_k": 5},
+        )
+
+    call_json = mock_client.post.call_args[1]["json"]
+    assert call_json["chat_template_kwargs"] == {"enable_thinking": False}
+    assert call_json["min_p"] == 0.1
+    assert call_json["top_k"] == 5