fix: allow legacy runnable methods when calling API (#4573)

* fix: allow legacy runnable methods when calling API

Signed-off-by: Frost Ming <me@frostming.com>

* fix: output content type

Signed-off-by: Frost Ming <me@frostming.com>

* fix compatibility with python 3.9

Signed-off-by: Frost Ming <me@frostming.com>

src/_bentoml_impl/client/proxy.py

@@ -7,7 +7,6 @@
 import typing as t
 
 from _bentoml_sdk import Service
-from _bentoml_sdk.method import APIMethod
 from bentoml.exceptions import BentoMLException
 
 from .http import AbstractClient
@@ -68,7 +67,7 @@ def as_service(self) -> T:
 
     def call(self, __name: str, /, *args: t.Any, **kwargs: t.Any) -> t.Any:
         original_func = getattr(self._inner, __name)
-        if not isinstance(original_func, APIMethod):
+        if not hasattr(original_func, "func"):
             raise BentoMLException(f"calling non-api method {__name} is not allowed")
         original_func = original_func.func
         while isinstance(original_func, functools.partial):

src/_bentoml_sdk/io_models.py

@@ -141,8 +141,6 @@ async def to_http_response(cls, obj: t.Any, serde: Serde) -> Response:
 
         from _bentoml_impl.serde import JSONSerde
 
-        structured_media_type = cls.media_type or serde.media_type
-
         if inspect.isasyncgen(obj):
 
             async def async_stream() -> t.AsyncGenerator[str | bytes, None]:
@@ -173,7 +171,7 @@ def content_stream() -> t.Generator[str | bytes, None, None]:
                 )
             return Response(
                 content=serde.serialize_model(t.cast(IODescriptor, obj)),
-                media_type=structured_media_type,
+                media_type=cls.mime_type(),
             )
         else:
             if is_file_type(type(obj)) and isinstance(serde, JSONSerde):
@@ -208,7 +206,7 @@ def content_stream() -> t.Generator[str | bytes, None, None]:
                 return Response(content=rendered, media_type=cls.mime_type())
             else:
                 return Response(
-                    content=serde.serialize_model(ins), media_type=structured_media_type
+                    content=serde.serialize_model(ins), media_type=cls.mime_type()
                 )
 
 
@@ -307,23 +305,25 @@ def from_output(cls, func: t.Callable[..., t.Any]) -> type[IODescriptor]:
             ) from e
 
 
-def ensure_io_descriptor(output_type: type) -> type[IODescriptor]:
+def ensure_io_descriptor(typ_: type) -> type[IODescriptor]:
     from pydantic._internal._utils import lenient_issubclass
 
-    if inspect.isclass(output_type) and lenient_issubclass(output_type, BaseModel):
-        if not issubclass(output_type, IODescriptor):
+    type_name = getattr(typ_, "__name__", "")
+
+    if inspect.isclass(typ_) and lenient_issubclass(typ_, BaseModel):
+        if not issubclass(typ_, IODescriptor):
             return t.cast(
                 t.Type[IODescriptor],
                 create_model(
-                    f"{output_type.__name__}IODescriptor",
-                    __base__=(IOMixin, output_type),
+                    f"{type_name}IODescriptor",
+                    __base__=(IOMixin, typ_),
                 ),
             )
-        return output_type
+        return typ_
     return t.cast(
         t.Type[IODescriptor],
         create_model(
-            f"{output_type.__name__}IODescriptor",
-            __base__=(IOMixin, RootModel[output_type]),
+            f"{type_name}IODescriptor",
+            __base__=(IOMixin, RootModel[typ_]),
         ),
     )

src/bentoml/_internal/runner/runnable.py

@@ -140,9 +140,12 @@ class RunnableMethod(t.Generic[T, P, R]):
     config: RunnableMethodConfig
     _bentoml_runnable_method: None = None
 
-    def __get__(self, obj: T, _: t.Type[T] | None = None) -> t.Callable[P, R]:
+    def __get__(self, obj: T | None, _: t.Type[T] | None = None) -> t.Callable[P, R]:
         from ..utils import is_async_callable
 
+        if obj is None:
+            return self
+
         if is_async_callable(self.func):
 
             async def method(*args: P.args, **kwargs: P.kwargs) -> R: