handle invalid images

[irexyc]

Motivation

Handle invalid image with invalid url, invalid base64 or truncated data.

related #2302

[lvhan028]

Is there a better way to validate the image?

[irexyc]

Currently all models will convert the image to RGB, and checking here will not increase the time consumption

[AllentDan]

Is there a better way to validate the image?

I'm not sure if load() function satisfied the need. Besides, if convert('RGB') is a must for all the model, we may do it here and remove them from forward function.

[AllentDan]

Did not pass the following UT:

def test_load_image_base64():
    im = PIL.Image.new(mode='RGB', size=(200, 200))
    encoded = encode_image_base64(im)
    url = f'data:image/jpeg;base64,{encoded}'
    load_image(url)
    import pytest
    with pytest.raises(ValueError):
        load_image(url[:-100])

Ref: #1615

Update, it is ok since we create dummy image now.

[AllentDan]

LGTM. We may validate the image outside generate function in the future.

[allen-ash]

Considering of VLM service, if the image provided by the request is invalid, why bother to generate a dummy image? Is this gonna return the wrong answer for the request?

[irexyc]

@allen-ash

A reasonable approach is to return an error code if the inputs is invalid. The reason for doing this is that some people have reported that the service will crash if the input is invalid and using a dummy image requires less code changes. We may use error code in the future.