Fix pixel attention mask padding in smolvlm

[ManuelFay]

What does this PR do?

The attention mask (binary) is initialized with zeros in float64, while in make_pixel_mask, the ones and zeros are in int64.

def make_pixel_mask(
    image: np.ndarray, output_size: Tuple[int, int], input_data_format: Optional[Union[str, ChannelDimension]] = None
) -> np.ndarray:
    """
    Make a pixel mask for the image, where 1 indicates a valid pixel and 0 indicates padding.
    Args:
        image (`np.ndarray`):
            Image to make the pixel mask for.
        output_size (`Tuple[int, int]`):
            Output size of the mask.
    """
    input_height, input_width = get_image_size(image, channel_dim=input_data_format)
    mask = np.zeros(output_size, dtype=np.int64)
    mask[:input_height, :input_width] = 1
    return mask

Fixes # (issue)

This causes either int64, or float64 mask values depending on if inputs are padded.
This makes MPS devices bug since they don't support float64.

By the way @andimarafioti , we could just cast as np.bool if masks are binary but I guess there is a reason it was done in np.int64 ?

Who can review?

@orrzohar @andimarafioti @zucchini-nlp

[github-actions]

Hi 👋, thank you for opening this pull request! The pull request is converted to draft by default. The CI will be paused while the PR is in draft mode. When it is ready for review, please click the Ready for review button (at the bottom of the PR page). This will assign reviewers and trigger CI.

[qubvel]

LGTM, thanks!

[zucchini-nlp]

Great, let's merge!

[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.