featu: support passing "extra body" throught to providers

# What does this PR do?
Allows passing through extra_body parameters to inference providers.


closes #2720

## Test Plan
CI and added new test

llama_stack/providers/utils/inference/litellm_openai_mixin.py

@@ -230,6 +230,9 @@ class LiteLLMOpenAIMixin(
     ) -> OpenAICompletion:
         model_obj = await self.model_store.get_model(params.model)
 
+        # Extract extra fields
+        extra_body = dict(params.__pydantic_extra__ or {})
+
         request_params = await prepare_openai_completion_params(
             model=self.get_litellm_model_name(model_obj.provider_resource_id),
             prompt=params.prompt,
@@ -253,6 +256,7 @@ class LiteLLMOpenAIMixin(
             suffix=params.suffix,
             api_key=self.get_api_key(),
             api_base=self.api_base,
+            **extra_body,
         )
         return await litellm.atext_completion(**request_params)
 
@@ -272,6 +276,9 @@ class LiteLLMOpenAIMixin(
 
         model_obj = await self.model_store.get_model(params.model)
 
+        # Extract extra fields
+        extra_body = dict(params.__pydantic_extra__ or {})
+
         request_params = await prepare_openai_completion_params(
             model=self.get_litellm_model_name(model_obj.provider_resource_id),
             messages=params.messages,
@@ -298,6 +305,7 @@ class LiteLLMOpenAIMixin(
             user=params.user,
             api_key=self.get_api_key(),
             api_base=self.api_base,
+            **extra_body,
         )
         return await litellm.acompletion(**request_params)
 

llama_stack/providers/utils/inference/openai_mixin.py

@@ -228,6 +228,9 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
         """
         Direct OpenAI completion API call.
         """
+        # Extract extra fields using Pydantic's built-in __pydantic_extra__
+        extra_body = dict(params.__pydantic_extra__ or {})
+
         # Handle parameters that are not supported by OpenAI API, but may be by the provider
         #  prompt_logprobs is supported by vLLM
         #  guided_choice is supported by vLLM
@@ -316,6 +319,9 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
             user=params.user,
         )
 
+        # Extract any additional provider-specific parameters using Pydantic's __pydantic_extra__
+        if extra_body := dict(params.__pydantic_extra__ or {}):
+            request_params["extra_body"] = extra_body
         resp = await self.client.chat.completions.create(**request_params)
 
         return await self._maybe_overwrite_id(resp, params.stream)  # type: ignore[no-any-return]

tests/unit/providers/inference/test_remote_vllm.py

@@ -191,3 +191,48 @@ async def test_openai_chat_completion_is_async(vllm_inference_adapter):
 
         assert mock_create_client.call_count == 4  # no cheating
         assert total_time < (sleep_time * 2), f"Total time taken: {total_time}s exceeded expected max"
+
+
+async def test_extra_body_forwarding(vllm_inference_adapter):
+    """
+    Test that extra_body parameters (e.g., chat_template_kwargs) are correctly
+    forwarded to the underlying OpenAI client.
+    """
+    mock_model = Model(identifier="mock-model", provider_resource_id="mock-model", provider_id="vllm-inference")
+    vllm_inference_adapter.model_store.get_model.return_value = mock_model
+
+    with patch.object(VLLMInferenceAdapter, "client", new_callable=PropertyMock) as mock_client_property:
+        mock_client = MagicMock()
+        mock_client.chat.completions.create = AsyncMock(
+            return_value=OpenAIChatCompletion(
+                id="chatcmpl-abc123",
+                created=1,
+                model="mock-model",
+                choices=[
+                    OpenAIChoice(
+                        message=OpenAIAssistantMessageParam(
+                            content="test response",
+                        ),
+                        finish_reason="stop",
+                        index=0,
+                    )
+                ],
+            )
+        )
+        mock_client_property.return_value = mock_client
+
+        # Test with chat_template_kwargs for Granite thinking mode
+        params = OpenaiChatCompletionRequest(
+            model="mock-model",
+            messages=[{"role": "user", "content": "test"}],
+            stream=False,
+            chat_template_kwargs={"thinking": True},
+        )
+        await vllm_inference_adapter.openai_chat_completion(params)
+
+        # Verify that the client was called with extra_body containing chat_template_kwargs
+        mock_client.chat.completions.create.assert_called_once()
+        call_kwargs = mock_client.chat.completions.create.call_args.kwargs
+        assert "extra_body" in call_kwargs
+        assert "chat_template_kwargs" in call_kwargs["extra_body"]
+        assert call_kwargs["extra_body"]["chat_template_kwargs"] == {"thinking": True}