fix: add missing extra_body to client.chat.completions.create() call

- test requires vLLM as provider, current is skipped in GH Action
- test:
>export VLLM_URL="http://localhost:8000"
>pytest tests/integration/inference/test_openai_completion.py::test_openai_chat_completion_extra_body -v --stack-config="inference=remote::vllm"

Signed-off-by: Wen Zhou <wenzhou@redhat.com>

llama_stack/providers/remote/inference/vllm/vllm.py

@@ -654,6 +654,7 @@ class VLLMInferenceAdapter(Inference, ModelsProtocolPrivate):
         top_logprobs: int | None = None,
         top_p: float | None = None,
         user: str | None = None,
+        extra_body: dict[str, Any] | None = None,
     ) -> OpenAIChatCompletion | AsyncIterator[OpenAIChatCompletionChunk]:
         self._lazy_initialize_client()
         model_obj = await self._get_model(model)
@@ -681,6 +682,7 @@ class VLLMInferenceAdapter(Inference, ModelsProtocolPrivate):
             top_logprobs=top_logprobs,
             top_p=top_p,
             user=user,
+            extra_body=extra_body,
         )
         return await self.client.chat.completions.create(**params)  # type: ignore
 

tests/integration/inference/test_openai_completion.py

@@ -5,6 +5,8 @@
 # the root directory of this source tree.
 
 
+from unittest.mock import patch
+
 import pytest
 from openai import OpenAI
 
@@ -209,6 +211,48 @@ def test_openai_completion_guided_choice(llama_stack_client, client_with_models,
 # Run the chat-completion tests with both the OpenAI client and the LlamaStack client
 
 
+def test_openai_chat_completion_extra_body(llama_stack_client, client_with_models, text_model_id):
+    skip_if_provider_isnt_vllm(client_with_models, text_model_id)
+
+    adapter = client_with_models._routing_table.get_provider_impl(text_model_id)
+
+    captured_params = None
+
+    async def mock_create(**params):
+        nonlocal captured_params
+        captured_params = params
+
+        from unittest.mock import AsyncMock
+
+        mock_response = AsyncMock()
+        mock_response.choices = [AsyncMock()]
+        mock_response.choices[0].message = AsyncMock()
+        mock_response.choices[0].message.content = "Test response"
+        return mock_response
+
+    # Need explicit test extra_body is passed
+    with patch.object(adapter.client.chat.completions, "create", side_effect=mock_create):
+        response = llama_stack_client.chat.completions.create(
+            model=text_model_id,
+            messages=[
+                {
+                    "role": "user",
+                    "content": "Hi how are you?",
+                }
+            ],
+            stream=False,
+            extra_body={"chat_template_kwargs": {"thinking": True}},
+        )
+    assert captured_params is not None
+    assert "extra_body" in captured_params
+    assert captured_params["extra_body"]["chat_template_kwargs"]["thinking"] is True
+
+    assert len(response.choices) > 0
+    choice = response.choices[0]
+    assert choice.message.content is not None
+    assert len(choice.message.content) > 0
+
+
 @pytest.mark.parametrize(
     "test_case",
     [