processing_diffusionvl_qwen2_5_vl.py

# coding=utf-8
# Copyright 2025 The HustVL Team and The HuggingFace Inc. team. All rights reserved.
#
# This code is based on Qwen2.5-VL, which is derived from EleutherAI's GPT-NeoX library
# and the GPT-NeoX and OPT implementations. It has been modified to create DiffusionVL.
#
# 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.
"""
DiffusionVL Processor - Combines image processor and tokenizer.
"""

import re
from typing import List, Optional, Union

import torch

from transformers.feature_extraction_utils import BatchFeature
from transformers.image_utils import ImageInput
from transformers.processing_utils import ProcessingKwargs, ProcessorMixin, Unpack
from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
from transformers.video_utils import VideoInput


IMAGE_TOKEN_INDEX = -200
DEFAULT_IMAGE_TOKEN = "<image>"


class DiffusionVL_Qwen2_5_VL_ProcessorKwargs(ProcessingKwargs, total=False):
    """Keyword arguments for DiffusionVL_Qwen2_5_VL_Processor."""

    _defaults = {
        "text_kwargs": {
            "padding": False,
        },
    }


def tokenizer_image_token(
    prompt: str,
    tokenizer,
    image_token_index: int = IMAGE_TOKEN_INDEX,
    return_tensors: Optional[str] = None,
) -> Union[List[int], torch.Tensor]:
    """
    Tokenize text with image placeholders, replacing <image> with IMAGE_TOKEN_INDEX.

    This implementation matches the training code (llava/mm_utils.py::tokenizer_image_token).

    Args:
        prompt: Input text containing <image> placeholders.
        tokenizer: The tokenizer to use for encoding text.
        image_token_index: The token index to use for image placeholders.
        return_tensors: If "pt", return a PyTorch tensor.

    Returns:
        List of token IDs or a PyTorch tensor.
    """
    # Tokenize each chunk (matching training code behavior)
    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split(DEFAULT_IMAGE_TOKEN)]

    def insert_separator(X, sep):
        return [ele for sublist in zip(X, [sep] * len(X)) for ele in sublist][:-1]

    input_ids = []
    offset = 0

    # Handle BOS token if present (matching training code)
    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
        offset = 1
        input_ids.append(prompt_chunks[0][0])

    for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
        input_ids.extend(x[offset:])

    if return_tensors is not None:
        if return_tensors == "pt":
            return torch.tensor(input_ids, dtype=torch.long)
        raise ValueError(f"Unsupported tensor type: {return_tensors}")
    return input_ids


class DiffusionVL_Qwen2_5_VL_Processor(ProcessorMixin):
    r"""
    Constructs a DiffusionVL processor which wraps an image processor and a tokenizer into a single processor.

    [`DiffusionVL_Qwen2_5_VL_Processor`] offers all the functionalities of [`Qwen2VLImageProcessor`] and [`Qwen2TokenizerFast`].
    See the [`~DiffusionVL_Qwen2_5_VL_Processor.__call__`] and [`~DiffusionVL_Qwen2_5_VL_Processor.decode`] for more information.

    This processor uses LLaVA-style image token handling:
    - `<image>` in text is replaced with `IMAGE_TOKEN_INDEX` (-200) in input_ids
    - The model's `prepare_inputs_labels_for_multimodal` replaces -200 with actual image features

    Args:
        image_processor ([`Qwen2VLImageProcessor`], *optional*):
            The image processor is a required input.
        tokenizer ([`Qwen2TokenizerFast`], *optional*):
            The tokenizer is a required input.
        chat_template (`str`, *optional*):
            A Jinja template which will be used to convert lists of messages in a chat into a tokenizable string.

    Example:

    ```python
    >>> from transformers import AutoProcessor
    >>> from PIL import Image

    >>> processor = AutoProcessor.from_pretrained("path/to/model", trust_remote_code=True)

    >>> # Prepare text with image placeholder
    >>> messages = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "Describe this image."}]}]
    >>> text = processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

    >>> # Process image and text
    >>> image = Image.open("image.jpg")
    >>> inputs = processor(text=[text], images=[image], return_tensors="pt")
    ```
    """

    attributes = ["image_processor", "tokenizer"]
    image_processor_class = "Qwen2VLImageProcessor"
    tokenizer_class = ("Qwen2Tokenizer", "Qwen2TokenizerFast")

    def __init__(
        self,
        image_processor=None,
        tokenizer=None,
        chat_template: Optional[str] = None,
        **kwargs,
    ):
        self.image_token = DEFAULT_IMAGE_TOKEN
        self.image_token_index = IMAGE_TOKEN_INDEX

        super().__init__(image_processor, tokenizer, chat_template=chat_template)

    def __call__(
        self,
        images: Optional[ImageInput] = None,
        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
        videos: Optional[VideoInput] = None,
        **kwargs: Unpack[DiffusionVL_Qwen2_5_VL_ProcessorKwargs],
    ) -> BatchFeature:
        """
        Main method to prepare for the model one or several sequences and image(s).

        This method forwards the `text` and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`]
        if `text` is not `None` to encode the text. To prepare the vision inputs, this method forwards the `images`
        and `kwargs` arguments to Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `images` is not `None`.

        The text should contain `<image>` placeholders where images should be inserted.
        These will be replaced with `IMAGE_TOKEN_INDEX` (-200) in the output input_ids.

        Args:
            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, *optional*):
                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array, or PyTorch
                tensor. Both channels-first and channels-last formats are supported.
            text (`str`, `List[str]`, *optional*):
                The sequence or batch of sequences to be encoded. Each sequence should be a string containing
                `<image>` placeholders where images will be inserted.
            videos (`np.ndarray`, `torch.Tensor`, `List[np.ndarray]`, *optional*):
                The video or batch of videos to be prepared. Currently not fully supported.
            return_tensors (`str` or [`~utils.TensorType`], *optional*):
                If set, will return tensors of a particular framework. Acceptable values are:
                - `'pt'`: Return PyTorch `torch.Tensor` objects.
                - `'np'`: Return NumPy `np.ndarray` objects.

        Returns:
            [`BatchFeature`]: A [`BatchFeature`] with the following fields:

            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model.
            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
            - **image_grid_thw** -- List of image 3D grid dimensions. Returned when `images` is not `None`.
        """
        output_kwargs = self._merge_kwargs(
            DiffusionVL_Qwen2_5_VL_ProcessorKwargs,
            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
            **kwargs,
        )

        # Process images
        image_inputs = {}
        if images is not None:
            image_inputs = self.image_processor(
                images=images, **output_kwargs.get("images_kwargs", {})
            )

        # Handle text input
        if text is None:
            return BatchFeature(data=image_inputs)

        if not isinstance(text, list):
            text = [text]

        # Tokenize with LLaVA-style image token handling
        return_tensors = output_kwargs.get("text_kwargs", {}).pop("return_tensors", None)

        all_input_ids = []
        for t in text:
            input_ids = tokenizer_image_token(
                t, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors=None
            )
            all_input_ids.append(input_ids)

        # Pad sequences
        max_len = max(len(ids) for ids in all_input_ids)
        padded_input_ids = []
        attention_masks = []

        pad_token_id = (
            self.tokenizer.pad_token_id if self.tokenizer.pad_token_id is not None else 0
        )

        for ids in all_input_ids:
            padding_length = max_len - len(ids)
            padded_ids = ids + [pad_token_id] * padding_length
            mask = [1] * len(ids) + [0] * padding_length
            padded_input_ids.append(padded_ids)
            attention_masks.append(mask)

        text_inputs = {
            "input_ids": padded_input_ids,
            "attention_mask": attention_masks,
        }

        return BatchFeature(data={**text_inputs, **image_inputs}, tensor_type=return_tensors)

    def build_conversation_input_ids(
        self,
        messages: List[dict],
        images: Optional[List] = None,
        add_generation_prompt: bool = True,
    ) -> dict:
        """
        Build input_ids from conversation messages in LLaVA format.

        This method converts a list of messages into a prompt string with `<image>` placeholders.
        Uses LLaVA-style chat template format (trained format).

        Args:
            messages: List of message dicts with 'role' and 'content' keys.
                Content can be a string or a list of dicts with 'type' key ('text' or 'image').
            images: Optional list of images (used for validation).
            add_generation_prompt: Whether to add generation prompt at the end.

        Returns:
            dict with 'text' key containing the prompt string with `<image>` placeholders.
        """
        # Build LLaVA-style prompt directly
        # Format: <|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<image>\nPrompt<|im_end|>\n<|im_start|>assistant\n

        text_parts = []

        for message in messages:
            role = message.get("role", "user")
            content = message.get("content", "")

            text_parts.append(f"<|im_start|>{role}\n")

            # Handle content - can be string or list of content items
            if isinstance(content, str):
                text_parts.append(content)
            elif isinstance(content, list):
                for item in content:
                    if isinstance(item, dict):
                        if item.get("type") == "image":
                            text_parts.append(DEFAULT_IMAGE_TOKEN)
                        elif item.get("type") == "text":
                            text_parts.append(item.get("text", ""))
                    else:
                        text_parts.append(str(item))

            text_parts.append("<|im_end|>\n")

        if add_generation_prompt:
            text_parts.append("<|im_start|>assistant\n")

        text = "".join(text_parts)
        return {"text": text}

    def batch_decode(self, *args, **kwargs):
        """
        Decode a batch of token IDs to text.

        This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.batch_decode`].
        Please refer to the docstring of this method for more information.
        """
        return self.tokenizer.batch_decode(*args, **kwargs)

    def decode(self, *args, **kwargs):
        """
        Decode token IDs to text.

        This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.decode`].
        Please refer to the docstring of this method for more information.
        """
        return self.tokenizer.decode(*args, **kwargs)

    @property
    def model_input_names(self) -> List[str]:
        """Return the list of model input names."""
        tokenizer_names = self.tokenizer.model_input_names
        image_processor_names = self.image_processor.model_input_names
        return list(dict.fromkeys(tokenizer_names + image_processor_names))


__all__ = ["DiffusionVL_Qwen2_5_VL_Processor", "tokenizer_image_token"]