Process inputs directly in apply_chat_template in image-text-to-text pipeline

[yonigozlan]

What does this PR do?

Follows #34275
Process inputs directly in apply_chat_template instead of calling apply_chat_template then the processor.
This also means that a small part of the pipeline logic needed to change, but I think it's better now :).
The pipeline also supports passing images with the images arg even when using chat template, where the corresponding image is represented with a {"type": "image"} in the chat.

In the previous behavior, when the input was:

image_ny = "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
image_chicago = "https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg"
messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "What’s the difference between these two images?"},
            {"type": "image"},
            {"type": "image"},
        ],
    }
]
outputs = pipe([image_ny, image_chicago], text=messages)

The output would be:

[
    {
        "input_text": [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": "What’s the difference between these two images?"},
                    {"type": "image"},
                    {"type": "image"},
                ],
            }
        ],
        "generated_text": [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": "What’s the difference between these two images?"},
                    {"type": "image"},
                    {"type": "image"},
                ],
            },
            {
                "role": "assistant",
                "content": "The first image shows a statue of Liberty in the foreground, while the second image shows a city skyline",
            },
        ],
    }
]

With no mention of the actual input images.
Now the output is:

[
    {
        "input_text": [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": "What’s the difference between these two images?"},
                    {
                        "type": "image",
                        "image": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg",
                    },
                    {
                        "type": "image",
                        "image": "https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg",
                    },
                ],
            }
        ],
        "generated_text": [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": "What’s the difference between these two images?"},
                    {
                        "type": "image",
                        "image": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg",
                    },
                    {
                        "type": "image",
                        "image": "https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg",
                    },
                ],
            },
            {
                "role": "assistant",
                "content": "The first image shows a statue of Liberty in the foreground, while the second image shows a city skyline",
            },
        ],
    }
]

Who can review?

@zucchini-nlp @Rocketknight1

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[zucchini-nlp]

Nice, thanks for updating the pipeline. Left a couple comments

[zucchini-nlp]

This should be 1 to work correctly with different ViT backbones. Was it causing any test failures?

[yonigozlan]

Without this change, I'm getting errors on pipeline tests that use to work with llava-interleave. For example:

pipe = pipeline("image-text-to-text", model="llava-hf/llava-interleave-qwen-0.5b-hf")
        image = "./tests/fixtures/tests_samples/COCO/000000039769.png"
        text = "<image> What this is? Assistant: This is"

        outputs = pipe(image, text=text)
        self.assertEqual(
            outputs,
            [
                {
                    "input_text": "<image> What this is? Assistant: This is",
                    "generated_text": "<image> What this is? Assistant: This is a photo of two cats lying on a pink blanket. The cats are sleeping and appear to be comfortable",
                }
            ],
        )

returns:

ValueError: Image features and image tokens do not match: tokens: 728, features 729

[zucchini-nlp]

in case of llava-interleave-qwen-0.5b-hf I see a mismatch in vision_feature_select_strategy for the model config and for processor. Will fix that on the hub :)

[qubvel]

Thanks! A few comments on my side

[ArthurZucker]

Missing doc / examples but nice otherwise!