processing_laguna.py

# Copyright 2026 WebBrain and the HuggingFace Inc. team. All rights reserved.
#
# Adapted from moonshotai/Kimi-VL-A3B-Instruct's processing_kimi_vl.py
# (Apache-2.0), itself based on the Qwen2-VL processor.
#
# 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.
"""Processor for Laguna XS 2.1 Vision: wraps LagunaImageProcessor + the Laguna tokenizer."""

from typing import List, Union

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.utils import logging

logger = logging.get_logger(__name__)


class LagunaProcessorKwargs(ProcessingKwargs, total=False):
    _defaults = {
        "text_kwargs": {"padding": False},
        "images_kwargs": {},
    }


class LagunaProcessor(ProcessorMixin):
    r"""
    Constructs a Laguna processor which wraps [`LagunaImageProcessor`] and the Laguna
    tokenizer into a single processor.

    The chat template emits one literal `〈|SPECIAL_10|〉` placeholder per image (wrapped
    in `〈|SPECIAL_8|〉image〈|SPECIAL_9|〉...〈|SPECIAL_11|〉` — four previously-unused
    reserved special-token ids repurposed as media start/content/pad/end markers, see
    ``config.json``'s ``media_placeholder_token_id`` and the model card). This processor
    expands that single placeholder into ``grid_h * grid_w / merge_length`` repeats — the
    number of tokens the vision tower + projector will actually produce for that image's
    resolution — before tokenization.
    """

    attributes = ["image_processor", "tokenizer"]
    valid_kwargs = ["chat_template"]
    image_processor_class = "AutoImageProcessor"
    tokenizer_class = "AutoTokenizer"

    def __init__(self, image_processor=None, tokenizer=None, chat_template=None, **kwargs):
        self.image_token = "〈|SPECIAL_10|〉"
        super().__init__(image_processor, tokenizer, chat_template=chat_template)

    def __call__(
        self,
        images: ImageInput = None,
        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
        **kwargs: Unpack[LagunaProcessorKwargs],
    ) -> BatchFeature:
        if images is None and text is None:
            raise ValueError("You have to specify at least one of `images` or `text`.")

        output_kwargs = self._merge_kwargs(
            LagunaProcessorKwargs,
            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
            **kwargs,
        )

        if images is not None:
            image_inputs = self.image_processor(images, **output_kwargs["images_kwargs"])
            image_grid_hws = image_inputs["image_grid_hws"]
        else:
            image_inputs = {}
            image_grid_hws = None

        if isinstance(text, str):
            text = [text]
        elif not isinstance(text, list) and not isinstance(text[0], str):
            raise ValueError("Invalid input text. Please provide a string, or a list of strings")

        if image_grid_hws is not None:
            merge_length = self.image_processor.merge_kernel_size[0] * self.image_processor.merge_kernel_size[1]
            index = 0
            for i in range(len(text)):
                while self.image_token in text[i]:
                    num_tokens = int(image_grid_hws[index].prod() // merge_length)
                    text[i] = text[i].replace(self.image_token, "〈|PLACEHOLDER|〉" * num_tokens, 1)
                    index += 1
                text[i] = text[i].replace("〈|PLACEHOLDER|〉", self.image_token)

        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
        return BatchFeature(data={**text_inputs, **image_inputs})

    def batch_decode(self, *args, **kwargs):
        return self.tokenizer.batch_decode(*args, **kwargs)

    def decode(self, *args, **kwargs):
        return self.tokenizer.decode(*args, **kwargs)

    @property
    def model_input_names(self):
        tokenizer_input_names = self.tokenizer.model_input_names
        image_processor_input_names = self.image_processor.model_input_names
        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))


__all__ = ["LagunaProcessor", "LagunaProcessorKwargs"]