processing_qwen3_vl.py

from typing import Any, Optional, Union

import torch
import numpy as np
from transformers.image_utils import ImageInput
from transformers.processing_utils import ImagesKwargs, MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack, VideosKwargs
from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
from transformers.utils import logging
from transformers.video_utils import VideoInput

from .chunk_utils import chunk_tokens


logger = logging.get_logger(__name__)


class MMFeature(dict):
    def __init__(self, data, tensor_type: str | None = None):
        super().__init__(data)
        self.tensor_type = tensor_type
        self.convert_to_tensor()

    def convert_to_tensor(self) -> "MMFeature":
        if self.tensor_type is None:
            return self

        match self.tensor_type:
            case "pt":
                for k, v in self.items():
                    if not isinstance(v, torch.Tensor):
                        try:
                            self[k] = torch.tensor(v)
                        except Exception:
                            pass
            case "np":
                for k, v in self.items():
                    if not isinstance(v, np.ndarray):
                        try:
                            self[k] = np.array(v)
                        except Exception:
                            pass
            case _:
                raise ValueError(f"Unsupported tensor type: {self.tensor_type}")

        return self

    def to(self, target: Any) -> "MMFeature":
        for k, v in self.items():
            if isinstance(v, torch.Tensor):
                self[k] = v.to(target)

        return self


class Qwen3VLVideosProcessorKwargs(VideosKwargs, total=False):
    focus_size: Optional[int]
    max_chunk_size: Optional[int]


class Qwen3VLImagesKwargs(ImagesKwargs):
    min_pixels: Optional[int]
    max_pixels: Optional[int]
    patch_size: Optional[int]
    temporal_patch_size: Optional[int]
    merge_size: Optional[int]
    focus_size: Optional[int]


class Qwen3VLProcessorKwargs(ProcessingKwargs, total=False):
    images_kwargs: Qwen3VLImagesKwargs # type: ignore
    videos_kwargs: Qwen3VLVideosProcessorKwargs # type: ignore
    _defaults = { # type: ignore
        "text_kwargs": {
            "padding": False,
            "return_token_type_ids": False,
            "return_mm_token_type_ids": False,
        },
        "videos_kwargs": {"return_metadata": True},
    }


class ZFQwen3VLProcessor(ProcessorMixin):
    r"""
    Constructs a Qwen3VL processor which wraps a Qwen3VL image processor and a Qwen2 tokenizer into a single processor.
    [`Qwen3VLProcessor`] offers all the functionalities of [`Qwen2VLImageProcessor`] and [`Qwen2TokenizerFast`]. See the
    [`~Qwen3VLProcessor.__call__`] and [`~Qwen3VLProcessor.decode`] for more information.
    Args:
        image_processor ([`Qwen2VLImageProcessor`], *optional*):
            The image processor is a required input.
        tokenizer ([`Qwen2TokenizerFast`], *optional*):
            The tokenizer is a required input.
        video_processor ([`Qwen3VLVideoProcessor`], *optional*):
            The video processor 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.
    """

    attributes = ["image_processor", "tokenizer", "video_processor"]
    image_processor_class = "AutoImageProcessor"
    video_processor_class = "AutoVideoProcessor"
    tokenizer_class = ("Qwen2Tokenizer", "Qwen2TokenizerFast")

    def __init__(self, image_processor=None, tokenizer=None, video_processor=None, chat_template=None, **kwargs):
        super().__init__(image_processor, tokenizer, video_processor, chat_template=chat_template)
        self.image_token = "<|image_pad|>" if not hasattr(tokenizer, "image_token") else tokenizer.image_token # type: ignore
        self.video_token = "<|video_pad|>" if not hasattr(tokenizer, "video_token") else tokenizer.video_token # type: ignore
        self.image_token_id = (
            tokenizer.image_token_id # type: ignore
            if getattr(tokenizer, "image_token_id", None)
            else tokenizer.convert_tokens_to_ids(self.image_token) # type: ignore
        )
        self.video_token_id = (
            tokenizer.video_token_id # type: ignore
            if getattr(tokenizer, "video_token_id", None)
            else tokenizer.convert_tokens_to_ids(self.video_token) # type: ignore
        )
        self.vision_start_token = (
            "<|vision_start|>" if not hasattr(tokenizer, "vision_start_token") else tokenizer.vision_start_token # type: ignore
        )
        self.vision_end_token = (
            "<|vision_end|>" if not hasattr(tokenizer, "vision_end_token") else tokenizer.vision_end_token # type: ignore
        )
        self.vision_start_token_id = (
            tokenizer.vision_start_token_id # type: ignore
            if getattr(tokenizer, "vision_start_token_id", None)
            else tokenizer.convert_tokens_to_ids(self.vision_start_token) # type: ignore
        )
        self.vision_end_token_id = (
            tokenizer.vision_end_token_id # type: ignore
            if getattr(tokenizer, "vision_end_token_id", None)
            else tokenizer.convert_tokens_to_ids(self.vision_end_token) # type: ignore
        )

    def __call__( # type: ignore
        self,
        images: ImageInput = None, # type: ignore
        text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]] = None, # type: ignore
        videos: VideoInput = None, # type: ignore
        **kwargs: Unpack[Qwen3VLProcessorKwargs],
    ) -> MMFeature:
        """
        Main method to prepare for the model one or several sequences(s) 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 `vision_infos` and `kwrags` arguments to
        Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`.

        Args:
            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`):
                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]`, `list[list[str]]`):
                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
            videos (`np.ndarray`, `torch.Tensor`, `list[np.ndarray]`, `list[torch.Tensor]`):
                The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
                tensor, or a nested list of 3D frames. Both channels-first and channels-last formats are supported.
            return_tensors (`str` or [`~utils.TensorType`], *optional*):
                If set, will return tensors of a particular framework. Acceptable values are:
                - `'tf'`: Return TensorFlow `tf.constant` objects.
                - `'pt'`: Return PyTorch `torch.Tensor` objects.
                - `'np'`: Return NumPy `np.ndarray` objects.
                - `'jax'`: Return JAX `jnp.ndarray` objects.

        Returns:
            [`MMFeature`]: A [`MMFeature`] 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 (when
              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
              `None`).
            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
            - **pixel_values_videos** -- Pixel values of videos to be fed to a model. Returned when `videos` is not `None`.
            - **image_grid_thw** -- List of image 3D grid in LLM. Returned when `images` is not `None`.
            - **video_grid_thw** -- List of video 3D grid in LLM. Returned when `videos` is not `None`.
        """
        output_kwargs = self._merge_kwargs(
            Qwen3VLProcessorKwargs, # type: ignore
            tokenizer_init_kwargs=self.tokenizer.init_kwargs, # type: ignore
            **kwargs,
        )
        if images is not None:
            image_inputs = self.image_processor(images=images, **output_kwargs["images_kwargs"]) # type: ignore
            image_grid_thw = image_inputs["image_grid_thw"]
        else:
            image_inputs = {}
            image_grid_thw = None

        if videos is not None:
            videos_inputs = self.video_processor(videos=videos, **output_kwargs["videos_kwargs"]) # type: ignore
            video_grid_thw = videos_inputs["video_grid_thw"]
            # If user has not requested video metadata, pop it
            if "return_metadata" not in kwargs:
                video_metadata = videos_inputs.pop("video_metadata")
            else:
                video_metadata = videos_inputs["video_metadata"]
            video_grid_thw = videos_inputs["video_grid_thw"]
        else:
            videos_inputs = {}
            video_grid_thw = None

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

        text = text.copy()  # below lines change text in-place
        if image_grid_thw is not None:
            merge_length = self.image_processor.merge_size**2 # type: ignore
            index = 0
            for i in range(len(text)):
                while self.image_token in text[i]:
                    num_image_tokens = image_grid_thw[index].prod() // merge_length
                    text[i] = text[i].replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1) # type: ignore
                    index += 1
                text[i] = text[i].replace("<|placeholder|>", self.image_token) # type: ignore

        if video_grid_thw is not None:
            merge_length = self.video_processor.merge_size**2 # type: ignore
            index = 0
            for i in range(len(text)):
                while self.video_token in text[i]:
                    metadata = video_metadata[index] # type: ignore
                    if metadata.fps is None:
                        logger.warning_once( # type: ignore
                            "Qwen3VL requires frame timestamps to construct prompts, but the `fps` of the input video could not be inferred. "
                            "Probably `video_metadata` was missing from inputs and you passed pre-sampled frames. "
                            "Defaulting to `fps=24`. Please provide `video_metadata` for more accurate results."
                        )
                        metadata.fps = 24 if metadata.fps is None else metadata.fps

                    # if timestamps are not provided, calculate them
                    curr_timestamp = self._calculate_timestamps(
                        metadata.frames_indices,
                        metadata.fps,
                        self.video_processor.merge_size, # type: ignore
                        self.video_processor.focus_size, # type: ignore
                    )

                    video_placeholder = ""
                    frame_seqlen = video_grid_thw[index][1:].prod() // merge_length
                    for frame_idx in range(video_grid_thw[index][0]):
                        curr_time = curr_timestamp[frame_idx]
                        video_placeholder += f"<{curr_time:.1f} seconds>"
                        video_placeholder += (
                            self.vision_start_token + "<|placeholder|>" * frame_seqlen + self.vision_end_token
                        )
                    if f"{self.vision_start_token}{self.video_token}{self.vision_end_token}" in text[i]:
                        text[i] = text[i].replace( # type: ignore
                            f"{self.vision_start_token}{self.video_token}{self.vision_end_token}", video_placeholder, 1
                        )
                    else:
                        # vllm may input video token directly
                        text[i] = text[i].replace(self.video_token, video_placeholder, 1) # type: ignore
                    index += 1

                text[i] = text[i].replace("<|placeholder|>", self.video_token) # type: ignore

        return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
        return_mm_token_type_ids = output_kwargs["text_kwargs"].pop("return_mm_token_type_ids", None)
        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"]) # type: ignore
        self._check_special_mm_tokens(text, text_inputs, modalities=["image", "video"]) # type: ignore
        array_ids = np.array(text_inputs["input_ids"])
        array_attention_mask = np.array(text_inputs["attention_mask"]) if "attention_mask" in text_inputs else None

        if return_mm_token_type_ids:
            mm_token_type_ids = np.zeros_like(text_inputs["input_ids"])
            mm_token_type_ids[array_ids == self.image_token_id] = 1
            text_inputs["mm_token_type_ids"] = mm_token_type_ids.tolist()

        chunks = chunk_tokens(
            max_chunk_size=self.video_processor.max_chunk_size, # type: ignore
            input_ids=array_ids,
            image_token_id=self.image_token_id,
            video_token_id=self.video_token_id,
            merge_size=self.image_processor.merge_size, # type: ignore
            focus_size=self.video_processor.focus_size, # type: ignore
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
        )
        image_token_mask = (array_ids == self.image_token_id).astype(int)
        video_token_mask = (array_ids == self.video_token_id).astype(int)
        text_token_mask = np.ones_like(image_token_mask) - image_token_mask - video_token_mask
        if array_attention_mask is not None:
            text_token_mask = text_token_mask * array_attention_mask
            image_token_mask = image_token_mask * array_attention_mask
            video_token_mask = video_token_mask * array_attention_mask

        return MMFeature(data={
            **text_inputs,
            **image_inputs,
            **videos_inputs,
            "token_chunks": chunks,
            "text_token_mask": text_token_mask,
            "image_token_mask": image_token_mask,
            "video_token_mask": video_token_mask,
        }, tensor_type=return_tensors)

    def _get_num_multimodal_tokens(self, image_sizes=None, video_sizes=None, **kwargs):
        """
        Computes the number of placeholder tokens needed for multimodal inputs with the given sizes.
        Args:
            image_sizes (`list[list[int]]`, *optional*):
                The input sizes formatted as (height, width) per each image.
            video_sizes (`list[list[int]]`, *optional*):
                The input sizes formatted as (num_frames, height, width) per each video.
        Returns:
            `MultiModalData`: A `MultiModalData` object holding number of tokens per each of the provided
            input modalities, along with other useful data.
        """

        vision_data = {}
        if image_sizes is not None:
            images_kwargs = Qwen3VLProcessorKwargs._defaults.get("images_kwargs", {})
            images_kwargs.update(kwargs)
            merge_size = images_kwargs.get("merge_size", None) or self.image_processor.merge_size # type: ignore

            num_image_patches = [
                self.image_processor.get_number_of_image_patches(*image_size, images_kwargs) # type: ignore
                for image_size in image_sizes
            ]
            num_image_tokens = [(num_patches // merge_size**2) for num_patches in num_image_patches]
            vision_data.update({"num_image_tokens": num_image_tokens, "num_image_patches": num_image_patches})

        if video_sizes is not None:
            videos_kwargs = Qwen3VLProcessorKwargs._defaults.get("videos_kwargs", {})
            videos_kwargs.update(kwargs)
            num_video_patches = [
                self.video_processor.get_number_of_video_patches(*video_size, videos_kwargs) # type: ignore
                for video_size in video_sizes
            ]
            num_video_tokens = [(num_patches // merge_size**2) for num_patches in num_video_patches] # type: ignore
            vision_data["num_video_tokens"] = num_video_tokens

        return MultiModalData(**vision_data)

    def post_process_image_text_to_text(
        self, generated_outputs, skip_special_tokens=True, clean_up_tokenization_spaces=False, **kwargs
    ):
        """
        Post-process the output of the model to decode the text.

        Args:
            generated_outputs (`torch.Tensor` or `np.ndarray`):
                The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)`
                or `(sequence_length,)`.
            skip_special_tokens (`bool`, *optional*, defaults to `True`):
                Whether or not to remove special tokens in the output. Argument passed to the tokenizer's `batch_decode` method.
            clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
                Whether or not to clean up the tokenization spaces. Argument passed to the tokenizer's `batch_decode` method.
            **kwargs:
                Additional arguments to be passed to the tokenizer's `batch_decode method`.

        Returns:
            `list[str]`: The decoded text.
        """
        return self.tokenizer.batch_decode( # type: ignore
            generated_outputs,
            skip_special_tokens=skip_special_tokens,
            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
            **kwargs,
        )

    def _calculate_timestamps(
        self,
        indices: Union[list[int], np.ndarray],
        video_fps: float,
        merge_size: int = 2,
        focus_size: int = 2,
    ):
        if not isinstance(indices, list):
            indices = indices.tolist()
        b_size = merge_size * focus_size
        if len(indices) % b_size != 0:
            indices.extend(indices[-1] for _ in range(b_size - len(indices) % b_size)) # type: ignore
        timestamps = [idx / video_fps for idx in indices]
        # @JJJYmmm frames are merged by self.merge_size, \
        # so we need to average the timestamps between the first/last frame within the temporal patch
        timestamps = [
            (timestamps[i] + timestamps[i + merge_size - 1]) / 2 for i in range(0, len(timestamps), merge_size)
        ]
        return timestamps


__all__ = ["ZFQwen3VLProcessor"]