bug: Unable to request PILImage with Content-Type image/jpeg with new Service decorator

[adrian-arapiles]

Describe the bug

Unable to get an endpoint with PILImage without Content-Type:multipart/form-data with new Service decorator

To reproduce

I already have working code with bentoml 1.1.11, runners, yolo/pytorch model.
I'm trying to migrate to new decorators @bentoml.service and @bentoml.api to use concurrency feature introduced.

Bentoml version tested: [1.2.9, 1.2.10].

All examples running with bentoml serve -p 8000 service.py:Service.

My working code with 1.1.11 and 1.2.10 with Runner/Runnable looks like

from bentoml.io import Image, JSON


my_runnable = bentoml.Runner(
    MyRunnable,
    name="runnable"
)
svc = bentoml.Service("runnable", runners=[my_runnable])


@svc.api(input=Image(), output=JSON(), route="/v1/inference")
async def inference(image: Image):
        #ofuscated code for prepare image, inference, build response and return
        return output_json

An example curl to before code is

curl -X 'POST' \
        'http://localhost:8000/v1/inference' \
        -H 'accept: application/json' \
           -H 'Content-Type: image/jpeg' \
              --data-binary '@/path/to/my/image.jpg'

Now, I'm trying to migrate to new decorators and my code looks like

from src.Service import Service
from PIL.Image import Image as PILImage


@bentoml.service()
class Service:
    # this class is a sustitute of MyRunnable class. 
    my_new_service = bentoml.depends(Service)

    @bentoml.api(
        route="/v1/debug"
    )
    async def debug(
            self,
            image: PILImage = Field(description="Image to be processed")
    ):
        return {"width": image.width, "height": image.height}

Now my curl for this code (extracted from swagger) is

curl -X 'POST' \
  'http://localhost:8000/v1/inference' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'image=@/path/to/my/image.jpg'

The code is working, but I don't want to change the endpoint call from Content-Type: image/jpeg to Content-Type: multipart/form-data.

I tried this:

# This is the using the same Image() that `@svc.api(input=Image()` old decorator uses

from bentoml.io import Image
from PIL.Image import Image as PILImage
from src.Service import Service

@bentoml.service()
class Service:
    # this class is a sustitute of MyRunnable class. 
    my_new_service = bentoml.depends(Service)

    @bentoml.api(
        input_spec=Image(),
        route="/v1/debug"
    )
    async def debug(
            self,
            image: PILImage = Field(description="Image to be processed")
    ):
        return {"width": image.width, "height": image.height}

With the following errors:

  File "/.../service.py", line 34, in <module>
    class Service:
  File "/.../service.py", line 37, in Service
    @bentoml.api(
     ^^^^^^^^^^^^
  File "/.../_bentoml_sdk/decorators.py", line 101, in wrapper
    return APIMethod(func, **params)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "<attrs generated init _bentoml_sdk.method.APIMethod>", line 13, in __init__
  File "/.../_bentoml_sdk/method.py", line 36, in _io_descriptor_converter
    raise ValueError(f"{it} must be a class type")
ValueError: Image must be a class type

I already tried this:

# This is the using the Annotated example in documentation

from typing import Annotated
from bentoml.validators import ContentType
from PIL.Image import Image as PILImage
from src.Service import Service

@bentoml.service()
class Service:
    # this class is a sustitute of MyRunnable class. 
    my_new_service = bentoml.depends(Service)

    @bentoml.api(
        route="/v1/debug"
    )
    async def debug(
            self,
            image: Annotated[PILImage, ContentType('image/jpeg')]
    ):
        return {"width": image.width, "height": image.height}

This code is working too but the same mentioned problem, changed Content-Type to multipart.

I'm not sure if it is a bug or not. It's possible just I was missing something.

The motivation to upgrade to new decorators is the use of concurrency feature introduced.
Is there any way to use with old Runners way?

I really appreciated some help there.

Thanks in advance,
Adrián.

Expected behavior

Using PILImage or bentoml.io.Image in the same way and with the same endpoint result.

Environment

bentoml: 1.2.9 and 1.2.10
python: 3.11.8
platform: Macos Sonoma 14.4.1. M1 PRO cpu.