BUG: Compat with openai extra body (#2759)

xinference/core/tests/test_types.py

@@ -51,8 +51,9 @@ def test_create_completion_types():
     with pytest.raises(ValidationError):
         CreateCompletion()
 
-    with pytest.raises(ValidationError):
-        CreateCompletion(model="abc", prompt="def", not_exist="jdk")
+    # In order to support extra body for openai client, we allow non exist key values.
+    # with pytest.raises(ValidationError):
+    #     CreateCompletion(model="abc", prompt="def", not_exist="jdk")
 
     CreateCompletion(model="abc", prompt="def")
 
@@ -88,3 +89,47 @@ def test_create_chat_completion_types():
     # These chat and generate share the same type.
     assert CreateChatCompletionLlamaCpp is CreateCompletionLlamaCpp
     assert CreateChatCompletionTorch is CreateCompletionTorch
+
+
+def test_openai_requests():
+    # https://github.com/xorbitsai/inference/pull/2673
+    # The following request body is generated by this:
+    # client = openai.Client(api_key="not empty", base_url="http://127.0.0.1:9997/v1")
+    #
+    # class TSchema(BaseModel):
+    #     test: str
+    #     num: int
+    #
+    # schema = TSchema.model_json_schema()
+    # messages = [
+    #     {"role": "system", "content": "you are a helpful assistant"},
+    #     {"role": "user", "content": "give me an example json, fill a random number"}
+    # ]
+    # completion = client.chat.completions.create(
+    #     model="qwen2-instruct",
+    #     messages=messages,
+    #     extra_body={
+    #         "guided_json": schema,
+    #         "guided_decoding_backend": "outlines"
+    #     })
+    obj = {
+        "messages": [
+            {"role": "system", "content": "you are a helpful assistant"},
+            {
+                "role": "user",
+                "content": "give me an example json, fill a random number",
+            },
+        ],
+        "model": "qwen2-instruct",
+        "guided_json": {
+            "properties": {
+                "test": {"title": "Test", "type": "string"},
+                "num": {"title": "Num", "type": "integer"},
+            },
+            "required": ["test", "num"],
+            "title": "TSchema",
+            "type": "object",
+        },
+        "guided_decoding_backend": "outlines",
+    }
+    CreateChatCompletion.parse_obj(obj)

xinference/types.py

@@ -335,8 +335,10 @@ def get_pydantic_model_from_method(
     exclude_fields: Optional[Iterable[str]] = None,
     include_fields: Optional[Dict[str, Any]] = None,
 ) -> BaseModel:
+    # The validate_arguments set Config.extra = "forbid" by default.
     f = validate_arguments(meth, config={"arbitrary_types_allowed": True})
     model = f.model
+    model.Config.extra = "ignore"
     model.__fields__.pop("self", None)
     model.__fields__.pop("args", None)
     model.__fields__.pop("kwargs", None)