source_helpers.py

def _layer_hidden(output):
    """Return the hidden tensor from a decoder-layer output."""

    import torch

    if torch.is_tensor(output):
        return output
    if isinstance(output, (tuple, list)) and output and torch.is_tensor(output[0]):
        return output[0]
    raise TypeError(f"unsupported decoder-layer output type: {type(output)!r}")

def _closest_ratio(aspect, ratios, width, height, image_size):
    best = (1, 1)
    difference = float("inf")
    area = width * height
    for ratio in ratios:
        candidate = ratio[0] / ratio[1]
        current = abs(aspect - candidate)
        if current < difference or (
            current == difference
            and area > 0.5 * image_size * image_size * ratio[0] * ratio[1]
        ):
            difference = current
            best = ratio
    return best

def dynamic_tiles(image, *, image_size: int, max_tiles: int, thumbnail: bool):
    """Official InternVL dynamic tiling, kept local for reproducibility."""

    ratios = sorted(
        {
            (i, j)
            for n in range(1, max_tiles + 1)
            for i in range(1, n + 1)
            for j in range(1, n + 1)
            if 1 <= i * j <= max_tiles
        },
        key=lambda item: item[0] * item[1],
    )
    width, height = image.size
    columns, rows = _closest_ratio(
        width / height, ratios, width, height, image_size
    )
    resized = image.convert("RGB").resize(
        (image_size * columns, image_size * rows), resample=3
    )
    tiles = []
    for index in range(columns * rows):
        left = (index % columns) * image_size
        top = (index // columns) * image_size
        tiles.append(
            resized.crop((left, top, left + image_size, top + image_size))
        )
    if thumbnail and len(tiles) != 1:
        tiles.append(image.convert("RGB").resize((image_size, image_size), 3))
    return tiles

def _normalize_tiles(tiles):
    import numpy as np
    import torch

    mean = torch.tensor((0.485, 0.456, 0.406)).view(3, 1, 1)
    std = torch.tensor((0.229, 0.224, 0.225)).view(3, 1, 1)
    tensors = []
    for tile in tiles:
        array = np.asarray(tile, dtype=np.float32) / 255.0
        tensor = torch.from_numpy(array).permute(2, 0, 1)
        tensors.append((tensor - mean) / std)
    return torch.stack(tensors)