chore: import upstream snapshot with attribution

python/sglang/srt/configs/deepseek_ocr.py

@@ -0,0 +1,905 @@
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+from PIL import Image, ImageOps
+from torchvision.transforms import InterpolationMode
+from torchvision.transforms import functional as TF
+from transformers import (
+    AutoProcessor,
+    LlamaTokenizerFast,
+    PretrainedConfig,
+    ProcessorMixin,
+)
+
+from sglang.srt.multimodal.customized_mm_processor_utils import (
+    register_customized_processor,
+)
+from sglang.srt.sampling.custom_logit_processor import (
+    DeepseekOCRNoRepeatNGramLogitProcessor,
+)
+
+DeepseekOCRImage = Union[Image.Image, torch.Tensor]
+
+BASE_SIZE = 1024
+IMAGE_SIZE = 640
+CROP_MODE = True
+MIN_CROPS = 2
+MAX_CROPS = 6  # max:9; If your GPU memory is small, it is recommended to set it to 6.
+MAX_CONCURRENCY = 100  # If you have limited GPU memory, lower the concurrency count.
+NUM_WORKERS = 64  # image pre-process (resize/padding) workers
+PRINT_NUM_VIS_TOKENS = False
+SKIP_REPEAT = True
+MODEL_PATH = "deepseek-ai/DeepSeek-OCR"  # change to your model path
+
+NGRAM_NO_REPEAT_SIZE = 30
+NGRAM_NO_REPEAT_WINDOW = 90
+# Whitelist `<td>` and `</td>` token ids to allow table structures.
+NGRAM_NO_REPEAT_WHITELIST = (128821, 128822)
+
+DEFAULT_CUSTOM_LOGIT_PROCESSOR = DeepseekOCRNoRepeatNGramLogitProcessor.to_str()
+
+
+def get_default_ngram_custom_params() -> Dict[str, Any]:
+    """Return default custom params for the DeepSeek-OCR n-gram no repeat processor."""
+
+    return {
+        "ngram_size": NGRAM_NO_REPEAT_SIZE,
+        "window_size": NGRAM_NO_REPEAT_WINDOW,
+        "whitelist_token_ids": list(NGRAM_NO_REPEAT_WHITELIST),
+    }
+
+
+PROMPT = "<image>\n<|grounding|>Convert the document to markdown."
+
+
+def get_image_size(img: DeepseekOCRImage) -> Tuple[int, int]:
+    """Return (width, height) for both PIL.Image and torch.Tensor (CHW)."""
+    if isinstance(img, Image.Image):
+        return img.size
+    if isinstance(img, torch.Tensor):
+        if img.ndim != 3:
+            raise TypeError(f"Expected CHW image tensor, got shape {tuple(img.shape)}")
+        return int(img.shape[-1]), int(img.shape[-2])
+    raise TypeError(f"Unsupported image type: {type(img)}")
+
+
+def resize_image(img: DeepseekOCRImage, size: Tuple[int, int]) -> DeepseekOCRImage:
+    """Resize image to (width, height) for both PIL and tensor."""
+    if isinstance(img, Image.Image):
+        return img.resize(size, Image.BICUBIC)
+    return TF.resize(
+        img,
+        [size[1], size[0]],
+        interpolation=InterpolationMode.BICUBIC,
+        antialias=True,
+    ).contiguous()
+
+
+def crop_image(
+    img: DeepseekOCRImage, box: Tuple[int, int, int, int]
+) -> DeepseekOCRImage:
+    """Crop image with box=(left, upper, right, lower) for both PIL and tensor."""
+    if isinstance(img, Image.Image):
+        return img.crop(box)
+    left, upper, right, lower = box
+    return img[:, upper:lower, left:right].contiguous()
+
+
+def pad_image(
+    img: DeepseekOCRImage,
+    target_size: Tuple[int, int],
+    fill_color: Tuple[int, int, int],
+) -> DeepseekOCRImage:
+    """Fit-and-center-pad image to target_size=(width, height).
+
+    Replaces ImageOps.pad for tensor inputs.
+    """
+    if isinstance(img, Image.Image):
+        return ImageOps.pad(img, target_size, color=fill_color)
+    # tensor path: CHW format
+    _, h, w = img.shape
+    target_w, target_h = target_size
+    scale = min(target_w / w, target_h / h)
+    new_w = int(w * scale)
+    new_h = int(h * scale)
+    resized = TF.resize(
+        img,
+        [new_h, new_w],
+        interpolation=InterpolationMode.BICUBIC,
+        antialias=True,
+    )
+    pad_left = (target_w - new_w) // 2
+    pad_top = (target_h - new_h) // 2
+    if img.dtype == torch.uint8:
+        fill_tensor = torch.tensor(
+            list(fill_color), device=img.device, dtype=torch.uint8
+        ).view(3, 1, 1)
+    else:
+        fill_tensor = torch.tensor(
+            [c / 255.0 for c in fill_color], device=img.device, dtype=img.dtype
+        ).view(3, 1, 1)
+    result = fill_tensor.expand(3, target_h, target_w).clone()
+    result[:, pad_top : pad_top + new_h, pad_left : pad_left + new_w] = resized
+    return result.contiguous()
+
+
+class DictOutput(object):
+    def items(self):
+        return self.__dict__.items()
+
+    def keys(self):
+        return self.__dict__.keys()
+
+    def __getitem__(self, item):
+        return self.__dict__[item]
+
+    def __contains__(self, key):
+        return key in self.__dict__
+
+    def __setitem__(self, key, value):
+        self.__dict__[key] = value
+
+
+@dataclass
+class VLChatProcessorOutput(DictOutput):
+    input_ids: torch.LongTensor
+    target_ids: torch.LongTensor
+    images_crop: torch.LongTensor
+    pixel_values: (
+        torch.Tensor
+    )  # rename from "images" to "pixel_values" for compatibility
+    images_seq_mask: torch.BoolTensor
+    images_spatial_crop: torch.LongTensor
+
+    def __len__(self):
+        return len(self.input_ids)
+
+
+class ImageTransform(object):
+    def __init__(
+        self,
+        mean: Optional[Tuple[float, float, float]] = (0.5, 0.5, 0.5),
+        std: Optional[Tuple[float, float, float]] = (0.5, 0.5, 0.5),
+        normalize: bool = True,
+    ):
+        self.mean = mean
+        self.std = std
+        self.normalize = normalize
+
+        # only load torchvision.transforms when needed
+        try:
+            import torchvision.transforms as T
+
+            # FIXME: add version check for gguf
+        except ImportError as err:
+            raise ImportError(
+                "Please install torchvision via `pip install torchvision` to use Deepseek-VL2."
+            ) from err
+
+        transform_pipelines = [T.ToTensor()]
+
+        if normalize:
+            transform_pipelines.append(T.Normalize(mean, std))
+
+        self.transform = T.Compose(transform_pipelines)
+
+    def __call__(self, img):
+        if isinstance(img, torch.Tensor):
+            x = img
+            if x.dtype == torch.uint8:
+                x = x.to(torch.float32).div(255)
+            elif not x.is_floating_point():
+                x = x.to(torch.float32)
+            if self.normalize:
+
+                import torchvision.transforms as T
+
+                x = T.Normalize(self.mean, self.std)(x)
+            return x
+        x = self.transform(img)
+        return x
+
+
+def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
+    best_ratio_diff = float("inf")
+    best_ratio = (1, 1)
+    area = width * height
+    for ratio in target_ratios:
+        target_aspect_ratio = ratio[0] / ratio[1]
+        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
+        if ratio_diff < best_ratio_diff:
+            best_ratio_diff = ratio_diff
+            best_ratio = ratio
+        elif ratio_diff == best_ratio_diff:
+            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
+                best_ratio = ratio
+    return best_ratio
+
+
+def dynamic_preprocess(
+    image, min_num=MIN_CROPS, max_num=MAX_CROPS, image_size=640, use_thumbnail=False
+):
+    orig_width, orig_height = get_image_size(image)
+    aspect_ratio = orig_width / orig_height
+
+    # calculate the existing image aspect ratio
+    target_ratios = set(
+        (i, j)
+        for n in range(min_num, max_num + 1)
+        for i in range(1, n + 1)
+        for j in range(1, n + 1)
+        if i * j <= max_num and i * j >= min_num
+    )
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    # find the closest aspect ratio to the target
+    target_aspect_ratio = find_closest_aspect_ratio(
+        aspect_ratio, target_ratios, orig_width, orig_height, image_size
+    )
+
+    # calculate the target width and height
+    target_width = image_size * target_aspect_ratio[0]
+    target_height = image_size * target_aspect_ratio[1]
+    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+
+    # resize the image
+    resized_img = resize_image(image, (target_width, target_height))
+    processed_images = []
+    for i in range(blocks):
+        box = (
+            (i % (target_width // image_size)) * image_size,
+            (i // (target_width // image_size)) * image_size,
+            ((i % (target_width // image_size)) + 1) * image_size,
+            ((i // (target_width // image_size)) + 1) * image_size,
+        )
+        # split the image
+        split_img = crop_image(resized_img, box)
+        processed_images.append(split_img)
+    assert len(processed_images) == blocks
+    if use_thumbnail and len(processed_images) != 1:
+        thumbnail_img = resize_image(image, (image_size, image_size))
+        processed_images.append(thumbnail_img)
+    return processed_images, target_aspect_ratio
+
+
+class DeepseekOCRProcessor(ProcessorMixin):
+    tokenizer_class = ("LlamaTokenizer", "LlamaTokenizerFast")
+    attributes = ["tokenizer"]
+
+    def __init__(
+        self,
+        tokenizer: LlamaTokenizerFast,
+        candidate_resolutions: Tuple[Tuple[int, int]],
+        patch_size: int,
+        downsample_ratio: int,
+        image_mean: Tuple[float, float, float] = (0.5, 0.5, 0.5),
+        image_std: Tuple[float, float, float] = (0.5, 0.5, 0.5),
+        normalize: bool = True,
+        image_token: str = "<image>",
+        pad_token: str = "<｜▁pad▁｜>",
+        add_special_token: bool = False,
+        sft_format: str = "deepseek",
+        mask_prompt: bool = True,
+        ignore_id: int = -100,
+        ocr2_mode: bool = False,
+        **kwargs,
+    ):
+
+        self.candidate_resolutions = candidate_resolutions
+        self.image_size = candidate_resolutions[0][0]
+        self.patch_size = patch_size
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.normalize = normalize
+        self.downsample_ratio = downsample_ratio
+        self.base_size = BASE_SIZE
+        self.image_transform = ImageTransform(
+            mean=image_mean, std=image_std, normalize=normalize
+        )
+        self.tokenizer = tokenizer
+        # must set this，padding side with make a difference in batch inference
+        self.tokenizer.padding_side = "left"
+
+        # add the pad_token as special token to use 'tokenizer.pad_token' and 'tokenizer.pad_token_id'
+        if tokenizer.pad_token is None:
+            self.tokenizer.add_special_tokens({"pad_token": pad_token})
+
+        # add image token
+        image_token_id = self.tokenizer.vocab.get(image_token)
+        if image_token_id is None:
+            special_tokens = [image_token]
+            special_tokens_dict = {"additional_special_tokens": special_tokens}
+            self.tokenizer.add_special_tokens(special_tokens_dict)
+        self.image_token_id = self.tokenizer.vocab.get(image_token)
+
+        # add five special tokens for grounding-related tasks
+        # <|ref|>, <|/ref|>, <|det|>, <|/det|>, <|grounding|>
+        special_tokens = ["<|ref|>", "<|/ref|>", "<|det|>", "<|/det|>", "<|grounding|>"]
+        special_tokens_dict = {"additional_special_tokens": special_tokens}
+        self.tokenizer.add_special_tokens(special_tokens_dict)
+
+        # add special tokens for SFT data
+        special_tokens = ["<|User|>", "<|Assistant|>"]
+        special_tokens_dict = {"additional_special_tokens": special_tokens}
+        self.tokenizer.add_special_tokens(special_tokens_dict)
+
+        self.image_token = image_token
+        self.pad_token = pad_token
+        self.add_special_token = add_special_token
+        self.sft_format = sft_format
+        self.mask_prompt = mask_prompt
+        self.ignore_id = ignore_id
+        self.ocr2_mode = ocr2_mode
+
+        super().__init__(
+            tokenizer,
+            **kwargs,
+        )
+
+    def format_messages_v2(self, messages: str, pil_images, max_req_input_len=-1):
+        """play the role of format_messages_v2 and get_images_info in the last version"""
+        tokenized_data = []
+        masked_tokenized_data = []  # labels
+        images_list = []
+        images_seq_mask = []
+        images_spatial_crop = []
+
+        image_index = 0
+        image_token_cnt = messages.count(self.image_token)
+        (
+            input_ids,
+            images,
+            images_crop,
+            seq_mask,
+            spatial_crop,
+            num_image_tokens,
+            image_shapes,
+        ) = self.tokenize_with_images(
+            messages,
+            pil_images[image_index : image_index + image_token_cnt],
+            bos=True,
+            eos=True,
+            cropping=len(pil_images) <= 2,
+        )
+
+        image_index = image_token_cnt
+        images_list += images
+        images_seq_mask += seq_mask
+        images_spatial_crop = spatial_crop
+
+        return (
+            input_ids,
+            masked_tokenized_data,
+            images_list,
+            images_seq_mask,
+            images_spatial_crop,
+            images_crop,
+        )
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+
+    @property
+    def pad_id(self):
+        return self.tokenizer.pad_token_id
+
+    def encode(self, text: str, bos: bool = True, eos: bool = False):
+        t = self.tokenizer.encode(text, add_special_tokens=False)
+
+        if bos:
+            t = [self.bos_id] + t
+        if eos:
+            t = t + [self.eos_id]
+
+        return t
+
+    def decode(self, t: List[int], **kwargs) -> str:
+        return self.tokenizer.decode(t, **kwargs)
+
+    def process_one(
+        self,
+        prompt: str = None,
+        conversations: List[Dict[str, str]] = None,
+        images: List[Image.Image] = None,
+        apply_sft_format: bool = False,
+        inference_mode: bool = True,
+        system_prompt: str = "",
+        max_req_input_len: int = -1,
+        cropping: bool = True,
+        **kwargs,
+    ):
+        """
+
+        Args:
+            prompt (str): the formatted prompt;
+            conversations (List[Dict]): conversations with a list of messages;
+            images (List[ImageType]): the list of images;
+            apply_sft_format (bool): if prompt is not None, then apply the SFT format to prompt;
+                if conversations is not None, then it will always apply the SFT format to conversations;
+            inference_mode (bool): if True, then remove the last eos token;
+            system_prompt (str): the system prompt;
+            **kwargs:
+
+        Returns:
+            outputs (BaseProcessorOutput): the output of the processor,
+                - input_ids (torch.LongTensor): [N + image tokens]
+                - target_ids (torch.LongTensor): [N + image tokens]
+                - images (torch.FloatTensor): [n_images, 3, H, W]
+                - image_id (int): the id of the image token
+                - num_image_tokens (List[int]): the number of image tokens
+        """
+
+        prompt = conversations or prompt
+        (
+            input_ids,
+            masked_tokenized_str,
+            images_list,
+            images_seq_mask,
+            images_spatial_crop,
+            images_crop,
+        ) = self.format_messages_v2(prompt, images, max_req_input_len)
+
+        target_ids = torch.LongTensor(masked_tokenized_str)
+
+        has_images = len(images_list) > 0
+        has_local_crops = False
+        if len(images_spatial_crop) > 0:
+            has_local_crops = any(
+                crop[0] > 1 or crop[1] > 1 for crop in images_spatial_crop
+            )
+
+        if len(images_list) == 0:
+            images = torch.zeros((1, 3, self.image_size, self.image_size))
+        else:
+            images = torch.stack(images_list, dim=0)
+
+        images_spatial_crop = torch.stack(
+            [images_spatial_crop], dim=0
+        )  # stack the tensor to make it a batch of 1
+
+        prepare = VLChatProcessorOutput(
+            input_ids=input_ids,
+            target_ids=target_ids,
+            images_crop=images_crop,
+            pixel_values=images,
+            images_seq_mask=images_seq_mask,
+            images_spatial_crop=images_spatial_crop,
+        )
+        prepare.has_images = has_images
+        prepare.has_local_crops = has_local_crops
+
+        return prepare
+
+    def __call__(
+        self,
+        *,
+        prompt: str = None,
+        conversations: List[Dict[str, str]] = None,
+        images: List[Image.Image] = None,
+        apply_sft_format: bool = False,
+        inference_mode: bool = True,
+        system_prompt: str = "",
+        max_req_input_len: int = -1,
+        text: list[str] = None,
+        **kwargs,
+    ):
+        assert text is None or isinstance(text, list)
+        if text is not None:
+            text = text[0]
+        prepare = self.process_one(
+            prompt=prompt or text,
+            conversations=conversations,
+            images=images,
+            apply_sft_format=apply_sft_format,
+            inference_mode=inference_mode,
+            system_prompt=system_prompt,
+            max_req_input_len=max_req_input_len,
+        )
+
+        return prepare
+
+    def find_all_indices(self, messages, target_value):
+        indices = []
+        for index, item in enumerate(messages):
+            if item == target_value:
+                indices.append(index)
+        return indices
+
+    def tokenize_with_images(
+        self,
+        conversation: str,
+        images: List[Image.Image],
+        bos: bool = True,
+        eos: bool = True,
+        cropping: bool = True,
+    ):
+        """Tokenize text with <image> tags."""
+
+        conversation = conversation
+        assert conversation.count(self.image_token) == len(images)
+        text_splits = conversation.split(self.image_token)
+        images_list, images_crop_list, images_seq_mask, images_spatial_crop = (
+            [],
+            [],
+            [],
+            [],
+        )
+        image_shapes = []
+        num_image_tokens = []
+        tokenized_str = []
+        for text_sep, image in zip(text_splits, images):
+            """encode text_sep"""
+            tokenized_sep = self.encode(text_sep, bos=False, eos=False)
+
+            tokenized_str += tokenized_sep
+            images_seq_mask += [False] * len(tokenized_sep)
+
+            img_w, img_h = get_image_size(image)
+            image_shapes.append((img_w, img_h))
+
+            if img_w <= 640 and img_h <= 640:
+                crop_ratio = [1, 1]
+            else:
+                if cropping:
+                    images_crop_raw, crop_ratio = dynamic_preprocess(
+                        image, image_size=IMAGE_SIZE
+                    )
+                else:
+                    crop_ratio = [1, 1]
+
+            """process the global view"""
+            if self.image_size <= 640 and not cropping:
+                image = resize_image(image, (self.image_size, self.image_size))
+
+            global_view = pad_image(
+                image,
+                (self.base_size, self.base_size),
+                tuple(int(x * 255) for x in self.image_transform.mean),
+            )
+            images_list.append(self.image_transform(global_view))
+
+            num_width_tiles, num_height_tiles = crop_ratio
+            images_spatial_crop.append([num_width_tiles, num_height_tiles])
+
+            if num_width_tiles > 1 or num_height_tiles > 1:
+                for i in range(len(images_crop_raw)):
+                    images_crop_list.append(self.image_transform(images_crop_raw[i]))
+
+            """add image tokens"""
+            num_queries = math.ceil(
+                (self.image_size // self.patch_size) / self.downsample_ratio
+            )
+            num_queries_base = math.ceil(
+                (self.base_size // self.patch_size) / self.downsample_ratio
+            )
+
+            if self.ocr2_mode:
+                tokenized_image = []
+                if num_width_tiles > 1 or num_height_tiles > 1:
+                    tokenized_image += [self.image_token_id] * (
+                        num_queries * num_width_tiles * num_queries * num_height_tiles
+                    )
+                tokenized_image += [self.image_token_id] * (
+                    num_queries_base * num_queries_base
+                )
+                # One extra token for the view separator.
+                tokenized_image += [self.image_token_id]
+            else:
+                tokenized_image = (
+                    [self.image_token_id] * num_queries_base + [self.image_token_id]
+                ) * num_queries_base
+                tokenized_image += [self.image_token_id]
+                if num_width_tiles > 1 or num_height_tiles > 1:
+                    tokenized_image += (
+                        [self.image_token_id] * (num_queries * num_width_tiles)
+                        + [self.image_token_id]
+                    ) * (num_queries * num_height_tiles)
+            tokenized_str += tokenized_image
+
+            images_seq_mask += [True] * len(tokenized_image)
+            num_image_tokens.append(len(tokenized_image))
+
+        """process the last text split"""
+        tokenized_sep = self.encode(text_splits[-1], bos=False, eos=False)
+
+        tokenized_str += tokenized_sep
+        images_seq_mask += [False] * len(tokenized_sep)
+
+        """add the bos and eos tokens"""
+        if bos:
+            tokenized_str = [self.bos_id] + tokenized_str
+            images_seq_mask = [False] + images_seq_mask
+        if eos:
+            tokenized_str = tokenized_str + [self.eos_id]
+            images_seq_mask = images_seq_mask + [False]
+
+        assert len(tokenized_str) == len(
+            images_seq_mask
+        ), f"tokenize_with_images func: tokenized_str's length {len(tokenized_str)} is not equal to imags_seq_mask's length {len(images_seq_mask)}"
+
+        masked_tokenized_str = []
+        for token_index in tokenized_str:
+            if token_index != self.image_token_id:
+                masked_tokenized_str.append(token_index)
+            else:
+                masked_tokenized_str.append(self.ignore_id)
+
+        assert (
+            len(tokenized_str) == len(images_seq_mask) == len(masked_tokenized_str)
+        ), (
+            f"tokenized_str's length {len(tokenized_str)}, input_ids' length {len(masked_tokenized_str)}, "
+            f"imags_seq_mask's length {len(images_seq_mask)}, are not equal"
+        )
+        input_ids = torch.LongTensor(tokenized_str)
+        target_ids = torch.LongTensor(masked_tokenized_str)
+        images_seq_mask = torch.tensor(images_seq_mask, dtype=torch.bool)
+
+        # set input_ids < 0 | input_ids == self.image_token_id as ignore_id
+        target_ids[(input_ids < 0) | (input_ids == self.image_token_id)] = (
+            self.ignore_id
+        )
+        input_ids[input_ids < 0] = self.pad_id
+
+        inference_mode = True
+
+        if inference_mode:
+            # Remove the ending eos token
+            assert input_ids[-1] == self.eos_id
+            input_ids = input_ids[:-1]
+            target_ids = target_ids[:-1]
+            images_seq_mask = images_seq_mask[:-1]
+
+        if len(images_list) == 0:
+            pixel_values = torch.zeros((1, 3, self.base_size, self.base_size))
+            images_spatial_crop = torch.zeros((1, 1), dtype=torch.long)
+            images_crop = torch.zeros(
+                (1, 3, self.image_size, self.image_size)
+            ).unsqueeze(0)
+        else:
+            pixel_values = torch.stack(images_list, dim=0)
+            images_spatial_crop = torch.tensor(images_spatial_crop, dtype=torch.long)
+            if images_crop_list:
+                images_crop = torch.stack(images_crop_list, dim=0).unsqueeze(0)
+            else:
+                images_crop = torch.zeros(
+                    (1, 3, self.image_size, self.image_size)
+                ).unsqueeze(0)
+
+        input_ids = input_ids.unsqueeze(0)
+        return (
+            input_ids,
+            pixel_values,
+            images_crop,
+            images_seq_mask,
+            images_spatial_crop,
+            num_image_tokens,
+            image_shapes,
+        )
+
+
+class VisionEncoderConfig(PretrainedConfig):
+    model_type: str = "vision"
+
+    model_name: str = "vit_so400m_patch14_siglip_384.webli"
+    image_size: int = 384
+    patch_size: int = 16
+    width: int = 1024
+    layers: int = 24
+    heads: int = 16
+    mlp_ratio: int = 4
+    global_pool: str = "map"
+    ignore_head: bool = True
+    class_token: bool = False
+    num_classes: int = 0
+    use_checkpoint: bool = False
+    weight_init: str = "skip"
+    deterministic: bool = False
+    num_recomputing_layers: int = 0
+
+    def __init__(
+        self,
+        model_name: str = "vit_so400m_patch14_siglip_384.webli",
+        image_size: int = 384,
+        patch_size: int = 16,
+        width: int = 1024,
+        layers: int = 24,
+        heads: int = 16,
+        mlp_ratio: int = 4,
+        global_pool: str = "map",
+        ignore_head: bool = True,
+        class_token: bool = False,
+        num_classes: int = 0,
+        use_checkpoint: bool = False,
+        **kwargs,
+    ):
+        self.model_name = model_name
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.width = width
+        self.layers = layers
+        self.heads = heads
+        self.mlp_ratio = mlp_ratio
+        self.global_pool = global_pool
+        self.ignore_head = ignore_head
+        self.class_token = class_token
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+
+        super().__init__(**kwargs)
+
+
+class MlpProjectorConfig(PretrainedConfig):
+    model_type = "mlp_projector"
+    projector_type: str = "downsample_mlp_gelu"
+    input_dim: int = 1152
+    n_embed: int = 2048
+    depth: int = 2
+    mlp_ratio: int = 1
+    downsample_ratio: int = 2
+    token_pooling: bool = False
+
+    def __init__(
+        self,
+        projector_type: str = "downsample_mlp_gelu",
+        input_dim: int = 1152,
+        n_embed: int = 2048,
+        depth: int = 2,
+        mlp_ratio: int = 1,
+        downsample_ratio: int = 2,
+        **kwargs,
+    ):
+        self.projector_type = projector_type
+        self.input_dim = input_dim
+        self.n_embed = n_embed
+        self.depth = depth
+        self.mlp_ratio = mlp_ratio
+        self.downsample_ratio = downsample_ratio
+
+        super().__init__(**kwargs)
+
+
+class DeepseekV2Config(PretrainedConfig):
+    model_type = "deepseek_v2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=102400,
+        hidden_size=4096,
+        intermediate_size=11008,
+        moe_intermediate_size=1407,
+        num_hidden_layers=30,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        n_shared_experts=None,
+        n_routed_experts=None,
+        ep_size=1,
+        routed_scaling_factor=1.0,
+        kv_lora_rank=512,
+        q_lora_rank=1536,
+        qk_rope_head_dim=64,
+        v_head_dim=128,
+        qk_nope_head_dim=128,
+        topk_method="gready",
+        n_group=None,
+        topk_group=None,
+        num_experts_per_tok=None,
+        moe_layer_freq=1,
+        first_k_dense_replace=0,
+        norm_topk_prob=False,
+        scoring_func="softmax",
+        aux_loss_alpha=0.001,
+        seq_aux=True,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=100000,
+        eos_token_id=100001,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        use_mla=True,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.n_shared_experts = n_shared_experts
+        self.n_routed_experts = n_routed_experts
+        self.ep_size = ep_size
+        self.routed_scaling_factor = routed_scaling_factor
+        self.kv_lora_rank = kv_lora_rank
+        self.q_lora_rank = q_lora_rank
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.topk_method = topk_method
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.num_experts_per_tok = num_experts_per_tok
+        self.moe_layer_freq = moe_layer_freq
+        self.first_k_dense_replace = first_k_dense_replace
+        self.norm_topk_prob = norm_topk_prob
+        self.scoring_func = scoring_func
+        self.aux_loss_alpha = aux_loss_alpha
+        self.seq_aux = seq_aux
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = float(rms_norm_eps)
+        self.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.use_mla = use_mla
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+@register_customized_processor(processor_class=DeepseekOCRProcessor)
+class DeepseekVLV2Config(PretrainedConfig):
+    # model_type = "deepseek_vl_v2"
+    model_type = "deepseek-ocr"
+    vision_config: VisionEncoderConfig = None
+    projector_config: MlpProjectorConfig = None
+
+    tile_tag: str = "2D"
+    global_view_pos: str = "head"
+    candidate_resolutions: tuple[tuple[int, int]] = ((384, 384),)
+    customized_processor_type: type[Any] = DeepseekOCRProcessor
+
+    def __init__(
+        self,
+        tile_tag: str = "tile_tag",
+        global_view_pos: str = "head",
+        candidate_resolutions: tuple[tuple[int, int]] = ((384, 384),),
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        vision_config = kwargs.get("vision_config", {})
+        self.vision_config = VisionEncoderConfig(**vision_config)
+
+        projector_config = kwargs.get("projector_config", {})
+        self.projector_config = MlpProjectorConfig(**projector_config)
+
+        language_config = kwargs.get("language_config", {})
+        self.text_config = DeepseekV2Config(**language_config)
+
+        self.tile_tag = tile_tag
+        self.global_view_pos = global_view_pos
+        self.candidate_resolutions = candidate_resolutions
+        self.vocab_size = self.text_config.vocab_size
+        self.hidden_size = self.text_config.hidden_size
+
+
+AutoProcessor.register(DeepseekVLV2Config, DeepseekOCRProcessor)