MiniMaxAI/MiniMax-M3

1# Copyright 2023-2024 SGLang Team2# Licensed under the Apache License, Version 2.0 (the "License");3"""4MiniMax VL family HuggingFace-compatible Processor, ImageProcessor, VideoProcessor.5"""6import math7from typing import List, Tuple8 9import torch10from torchvision.transforms import InterpolationMode11from transformers import BatchFeature12from transformers.image_processing_utils_fast import (13    BaseImageProcessorFast,14    group_images_by_shape,15    reorder_images,16)17from transformers.image_utils import PILImageResampling, SizeDict18from transformers.processing_utils import (19    ImagesKwargs,20    Unpack,21)22from transformers.utils import TensorType23 24MAX_RATIO = 20025 26 27def round_by_factor(number: int, factor: int) -> int:28    return round(number / factor) * factor29 30 31def ceil_by_factor(number: int, factor: int) -> int:32    return math.ceil(number / factor) * factor33 34 35def floor_by_factor(number: int, factor: int) -> int:36    return math.floor(number / factor) * factor37 38 39def smart_resize(40    height: int,41    width: int,42    factor: int = 28,43    min_pixels: int = 4 * 28 * 28,44    max_pixels: int = 451584,45) -> tuple[int, int]:46    if max(height, width) / min(height, width) > MAX_RATIO:47        raise ValueError(48            f"absolute aspect ratio must be smaller than {MAX_RATIO}, "49            f"got {max(height, width) / min(height, width)}"50        )51    h_bar = max(factor, round_by_factor(height, factor))52    w_bar = max(factor, round_by_factor(width, factor))53    if h_bar * w_bar > max_pixels:54        beta = math.sqrt((height * width) / max_pixels)55        h_bar = floor_by_factor(height / beta, factor)56        w_bar = floor_by_factor(width / beta, factor)57    elif h_bar * w_bar < min_pixels:58        beta = math.sqrt(min_pixels / (height * width))59        h_bar = ceil_by_factor(height * beta, factor)60        w_bar = ceil_by_factor(width * beta, factor)61    return h_bar, w_bar62 63 64# ==============================================================================65# MiniMax M3 VL Image Processor Fast (Fast Mode - Torch based)66# ==============================================================================67 68 69class MiniMaxM3VLImageProcessorKwargs(ImagesKwargs, total=False):70    patch_size: int71    temporal_patch_size: int72    merge_size: int73    max_pixels: int74 75 76class MiniMaxM3VLImageProcessor(BaseImageProcessorFast):77    do_resize = True78    resample = PILImageResampling.BICUBIC79    size = {"height": 672, "width": 672}  # required by base class validation, not used as resize bound80    default_to_square = False81    do_rescale = True82    rescale_factor = 1 / 25583    do_normalize = True84    image_mean = [0.48145466, 0.4578275, 0.40821073]85    image_std = [0.26862954, 0.26130258, 0.27577711]86    do_convert_rgb = True87    patch_size = 1488    temporal_patch_size = 289    merge_size = 290    max_pixels = 451584             # 672*67291    valid_kwargs = MiniMaxM3VLImageProcessorKwargs92    model_input_names = ["pixel_values", "image_grid_thw"]93 94    def __init__(self, **kwargs: Unpack[MiniMaxM3VLImageProcessorKwargs]):95        super().__init__(**kwargs)96 97    def preprocess(98        self, images, **kwargs: Unpack[MiniMaxM3VLImageProcessorKwargs]99    ) -> BatchFeature:100        return super().preprocess(images, **kwargs)101 102    def _preprocess(103        self,104        images: List[torch.Tensor],105        do_resize: bool,106        size: SizeDict,107        resample: PILImageResampling | InterpolationMode | int | None,108        do_rescale: bool,109        rescale_factor: float,110        do_normalize: bool,111        image_mean: float | List[float] | None,112        image_std: float | List[float] | None,113        patch_size: int,114        temporal_patch_size: int,115        merge_size: int,116        max_pixels: int,117        disable_grouping: bool | None,118        return_tensors: str | TensorType | None,119        **kwargs,120    ) -> BatchFeature:121        grouped_images, grouped_images_index = group_images_by_shape(122            images, disable_grouping=disable_grouping123        )124        resized_images_grouped = {}125        factor = patch_size * merge_size126        for shape, stacked_images in grouped_images.items():127            height, width = stacked_images.shape[-2:]128            if do_resize:129                resized_height, resized_width = smart_resize(130                    height, width, factor=factor,131                    max_pixels=max_pixels,132                )133                stacked_images = self.resize(134                    stacked_images,135                    size=SizeDict(height=resized_height, width=resized_width),136                    resample=resample,137                )138            resized_images_grouped[shape] = stacked_images139 140        resized_images = reorder_images(resized_images_grouped, grouped_images_index)141 142        grouped_images, grouped_images_index = group_images_by_shape(143            resized_images, disable_grouping=disable_grouping144        )145        processed_images_grouped = {}146        processed_grids = {}147 148        for shape, stacked_images in grouped_images.items():149            resized_height, resized_width = stacked_images.shape[-2:]150 151            patches = self.rescale_and_normalize(152                stacked_images,153                do_rescale,154                rescale_factor,155                do_normalize,156                image_mean,157                image_std,158            )159            if patches.ndim == 4:160                patches = patches.unsqueeze(1)161 162            if patches.shape[1] % temporal_patch_size != 0:163                repeats = patches[:, -1:].repeat(164                    1,165                    temporal_patch_size - (patches.shape[1] % temporal_patch_size),166                    1,167                    1,168                    1,169                )170                patches = torch.cat([patches, repeats], dim=1)171 172            batch_size, grid_t, channel = patches.shape[:3]173            grid_t = grid_t // temporal_patch_size174            grid_h, grid_w = resized_height // patch_size, resized_width // patch_size175 176            patches = patches.view(177                batch_size,178                grid_t,179                temporal_patch_size,180                channel,181                grid_h // merge_size,182                merge_size,183                patch_size,184                grid_w // merge_size,185                merge_size,186                patch_size,187            )188            patches = patches.permute(0, 1, 4, 7, 5, 8, 3, 2, 6, 9)189 190            flatten_patches = patches.reshape(191                batch_size,192                grid_t * grid_h * grid_w,193                channel * temporal_patch_size * patch_size * patch_size,194            )195 196            processed_images_grouped[shape] = flatten_patches197            processed_grids[shape] = [[grid_t, grid_h, grid_w]] * batch_size198 199        processed_images = reorder_images(200            processed_images_grouped, grouped_images_index201        )202        processed_grids = reorder_images(processed_grids, grouped_images_index)203 204        pixel_values = torch.cat(processed_images, dim=0)205        image_grid_thw = torch.tensor(processed_grids, dtype=torch.long)206 207        return BatchFeature(208            data={"pixel_values": pixel_values, "image_grid_thw": image_grid_thw},209            tensor_type=return_tensors,210        )211 212    def get_number_of_image_patches(self, height: int, width: int, images_kwargs=None):213        images_kwargs = images_kwargs or {}214        patch_size = images_kwargs.get("patch_size", self.patch_size)215        merge_size = images_kwargs.get("merge_size", self.merge_size)216        max_pixels = images_kwargs.get("max_pixels", self.max_pixels)217 218        resized_height, resized_width = smart_resize(219            height, width, factor=patch_size * merge_size,220            max_pixels=max_pixels,221        )222        grid_h, grid_w = resized_height // patch_size, resized_width // patch_size223        return grid_h * grid_w224