#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/muse_glimmer/modular_muse_glimmer.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_muse_glimmer.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2026 the HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import itertools
import math

import torch
from torchvision.transforms.v2 import functional as tvF

from ...image_processing_backends import TorchvisionBackend
from ...image_processing_utils import BatchFeature
from ...image_transforms import group_images_by_shape, reorder_images
from ...image_utils import ImageInput, PILImageResampling, SizeDict
from ...processing_utils import ImagesKwargs, Unpack
from ...utils import TensorType, auto_docstring
from ...utils.constants import IMAGENET_STANDARD_MEAN, IMAGENET_STANDARD_STD


class MuseGlimmerImageProcessorKwargs(ImagesKwargs, total=False):
    """
    patch_size (`int`, *optional*, defaults to 14):
        The spatial patch size of the vision encoder.
    temporal_patch_size (`int`, *optional*, defaults to 2):
        The temporal patch size of the vision encoder.
    merge_size (`int`, *optional*, defaults to 2):
        The merge size of the vision encoder to llm encoder.
    max_image_tokens (`int`, *optional*, defaults to 4096):
        The maximum number of merged image tokens produced for one image.
    """

    patch_size: int
    temporal_patch_size: int
    merge_size: int

    max_image_tokens: int


def smart_resize(
    height: int,
    width: int,
    patch_size: int,
    max_tokens: int,
) -> tuple[int, int]:
    """Pick the integer patch grid closest to the input aspect ratio under the token cap.

    Returns the resize target ``(target_height, target_width)`` in pixels.
    """
    ideal_patches_height = height / patch_size
    ideal_patches_width = width / patch_size
    ratio = ideal_patches_width / ideal_patches_height if ideal_patches_height > 0 else 1.0
    if ideal_patches_height * ideal_patches_width > max_tokens:
        ideal_patches_height = (max_tokens / ratio) ** 0.5
        ideal_patches_width = ideal_patches_height * ratio
    candidates = list(
        set(
            itertools.product(
                [math.floor(ideal_patches_height), math.ceil(ideal_patches_height)],
                [math.floor(ideal_patches_width), math.ceil(ideal_patches_width)],
            )
        )
    )
    candidates = [
        (patches_height, patches_width)
        for patches_height, patches_width in candidates
        if patches_height >= 1 and patches_width >= 1 and patches_height * patches_width <= max_tokens
    ]
    if not candidates:
        candidates = [(max(1, round(ideal_patches_height)), max(1, round(ideal_patches_width)))]
    patches_height, patches_width = min(candidates, key=lambda grid: abs(grid[0] / grid[1] - height / width))
    return patches_height * patch_size, patches_width * patch_size


@auto_docstring
class MuseGlimmerImageProcessor(TorchvisionBackend):
    do_resize = True
    resample = PILImageResampling.LANCZOS
    size = None
    default_to_square = False
    do_rescale = True
    rescale_factor = 1 / 255
    do_normalize = True
    image_mean = IMAGENET_STANDARD_MEAN
    image_std = IMAGENET_STANDARD_STD
    do_convert_rgb = True
    patch_size = 14
    temporal_patch_size = 2
    merge_size = 2
    valid_kwargs = MuseGlimmerImageProcessorKwargs
    model_input_names = ["pixel_values", "image_grid_thw"]
    max_image_tokens = 4096

    @auto_docstring
    def preprocess(self, images: ImageInput, **kwargs: Unpack[MuseGlimmerImageProcessorKwargs]) -> BatchFeature:
        return super().preprocess(images, **kwargs)

    def resize(
        self,
        images: torch.Tensor,
        patch_size: int,
        merge_size: int,
        max_tokens: int,
        resample: PILImageResampling | tvF.InterpolationMode | int | None,
        **kwargs,
    ) -> torch.Tensor:
        """Resize dynamically based on input image aspect ratio."""
        height, width = images.shape[-2:]
        resized_height, resized_width = smart_resize(
            height=height,
            width=width,
            patch_size=patch_size * merge_size,
            max_tokens=max_tokens,
        )
        return super().resize(
            image=images,
            size=SizeDict(height=resized_height, width=resized_width),
            resample=resample,
            antialias=True,
        )

    def patchify(
        self,
        images: torch.Tensor,
        patch_size: int,
        temporal_patch_size: int,
    ) -> tuple[torch.Tensor, int, int]:
        """Patchifies each image into flat layout of shape (`seq_len`, `patch_dim`) so we can concat dynamically shaped pixels."""
        batch_size, channel, resized_height, resized_width = images.shape
        grid_h, grid_w = resized_height // patch_size, resized_width // patch_size
        patches = images.view(
            batch_size,
            channel,
            grid_h,
            patch_size,
            grid_w,
            patch_size,
        )
        # Unlike Glm4v, each flattened patch is laid out (temporal, channel), not (channel, temporal).
        patches = patches.permute(0, 2, 4, 1, 3, 5)
        flatten_patches = (
            patches.unsqueeze(3)
            .expand(-1, -1, -1, temporal_patch_size, -1, -1, -1)
            .reshape(
                batch_size,
                grid_h * grid_w,
                temporal_patch_size * channel * patch_size * patch_size,
            )
        )
        return flatten_patches, grid_h, grid_w

    def _preprocess(
        self,
        images: list[torch.Tensor],
        do_resize: bool,
        resample: PILImageResampling | tvF.InterpolationMode | int | None,
        do_rescale: bool,
        rescale_factor: float,
        do_normalize: bool,
        image_mean: float | list[float] | None,
        image_std: float | list[float] | None,
        return_tensors: str | TensorType | None,
        patch_size: int,
        temporal_patch_size: int,
        max_image_tokens: int,
        merge_size: int,
        disable_grouping: bool = False,
        **kwargs,
    ) -> BatchFeature:
        """
        Preprocess an image or batch of images.
        """
        grouped_images, grouped_images_index = group_images_by_shape(images, disable_grouping=disable_grouping)
        resized_images_grouped = {}
        for shape, stacked_images in grouped_images.items():
            if do_resize:
                # Unlike Glm4v's `smart_resize`, the target size keeps aspect ratio under a token cap.
                stacked_images = self.resize(
                    stacked_images,
                    patch_size=patch_size,
                    merge_size=merge_size,
                    max_tokens=max_image_tokens,
                    resample=resample,
                )
            resized_images_grouped[shape] = stacked_images
        resized_images = reorder_images(resized_images_grouped, grouped_images_index)

        grouped_images, grouped_images_index = group_images_by_shape(resized_images, disable_grouping=disable_grouping)
        processed_images_grouped = {}
        processed_grids = {}
        for shape, stacked_images in grouped_images.items():
            stacked_images = self.rescale_and_normalize(
                stacked_images, do_rescale, rescale_factor, do_normalize, image_mean, image_std
            )
            patches, grid_h, grid_w = self.patchify(
                stacked_images,
                patch_size=patch_size,
                temporal_patch_size=temporal_patch_size,
            )

            processed_images_grouped[shape] = patches
            processed_grids[shape] = [[1, grid_h, grid_w]] * len(stacked_images)

        processed_images = reorder_images(processed_images_grouped, grouped_images_index)
        processed_grids = reorder_images(processed_grids, grouped_images_index)
        pixel_values = torch.cat(processed_images, dim=0)
        image_grid_thw = torch.tensor(processed_grids)

        return BatchFeature(
            data={"pixel_values": pixel_values, "image_grid_thw": image_grid_thw}, tensor_type=return_tensors
        )

    def get_number_of_image_patches(self, height: int, width: int, images_kwargs=None) -> int:
        """
        A utility that returns number of image patches for a given image size.

        Note: Do not remove this method! It is used by vLLM to infer the number of patches and placeholders
        without an image input.

        Args:
            height (`int`):
                Height of the input image.
            width (`int`):
                Width of the input image.
            images_kwargs (`dict`, *optional*)
                Any kwargs to override defaults of the image processor.
        Returns:
            `int`: Number of image patches per image.
        """
        patch_size = images_kwargs.get("patch_size", self.patch_size)
        merge_size = images_kwargs.get("merge_size", self.merge_size)
        max_image_tokens = images_kwargs.get("max_image_tokens", self.max_image_tokens)

        resized_height, resized_width = smart_resize(
            height=height,
            width=width,
            patch_size=patch_size * merge_size,
            max_tokens=max_image_tokens,
        )
        grid_h, grid_w = resized_height // patch_size, resized_width // patch_size
        return grid_h * grid_w

    def _validate_preprocess_kwargs(self, **kwargs):
        # MuseGlimmer uses aspect_ratio_preserving_resize driven by patch_size,
        # not the standard `size` parameter. Temporarily disable do_resize so
        # the base validation doesn't raise an error
        kwargs["do_resize"] = False
        super()._validate_preprocess_kwargs(**kwargs)


__all__ = ["MuseGlimmerImageProcessor"]