docling/docling/models/picture_description_vlm_model.py

import threading
from collections.abc import Iterable
from pathlib import Path
from typing import Optional, Type, Union

from PIL import Image
from transformers import AutoModelForImageTextToText

from docling.datamodel.accelerator_options import AcceleratorOptions
from docling.datamodel.pipeline_options import (
    PictureDescriptionBaseOptions,
    PictureDescriptionVlmOptions,
)
from docling.models.picture_description_base_model import PictureDescriptionBaseModel
from docling.models.utils.hf_model_download import (
    HuggingFaceModelDownloadMixin,
)
from docling.utils.accelerator_utils import decide_device

# Global lock for model initialization to prevent threading issues
_model_init_lock = threading.Lock()


class PictureDescriptionVlmModel(
    PictureDescriptionBaseModel, HuggingFaceModelDownloadMixin
):
    @classmethod
    def get_options_type(cls) -> Type[PictureDescriptionBaseOptions]:
        return PictureDescriptionVlmOptions

    def __init__(
        self,
        enabled: bool,
        enable_remote_services: bool,
        artifacts_path: Optional[Union[Path, str]],
        options: PictureDescriptionVlmOptions,
        accelerator_options: AcceleratorOptions,
    ):
        super().__init__(
            enabled=enabled,
            enable_remote_services=enable_remote_services,
            artifacts_path=artifacts_path,
            options=options,
            accelerator_options=accelerator_options,
        )
        self.options: PictureDescriptionVlmOptions

        if self.enabled:
            if artifacts_path is None:
                artifacts_path = self.download_models(repo_id=self.options.repo_id)
            else:
                artifacts_path = Path(artifacts_path) / self.options.repo_cache_folder

            self.device = decide_device(accelerator_options.device)

            try:
                import torch
                from transformers import AutoModelForVision2Seq, AutoProcessor
            except ImportError:
                raise ImportError(
                    "transformers >=4.46 is not installed. Please install Docling with the required extras `pip install docling[vlm]`."
                )

            # Initialize processor and model
            with _model_init_lock:
                self.processor = AutoProcessor.from_pretrained(artifacts_path)
                self.model = AutoModelForImageTextToText.from_pretrained(
                    artifacts_path,
                    device_map=self.device,
                    torch_dtype=torch.bfloat16,
                    _attn_implementation=(
                        "flash_attention_2"
                        if self.device.startswith("cuda")
                        and accelerator_options.cuda_use_flash_attention2
                        else "sdpa"
                    ),
                )
                self.model = torch.compile(self.model)  # type: ignore

            self.provenance = f"{self.options.repo_id}"

    def _annotate_images(self, images: Iterable[Image.Image]) -> Iterable[str]:
        from transformers import GenerationConfig

        # Create input messages
        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image"},
                    {"type": "text", "text": self.options.prompt},
                ],
            },
        ]

        # TODO: do batch generation

        for image in images:
            # Prepare inputs
            prompt = self.processor.apply_chat_template(
                messages, add_generation_prompt=True
            )
            inputs = self.processor(text=prompt, images=[image], return_tensors="pt")
            inputs = inputs.to(self.device)

            # Generate outputs
            generated_ids = self.model.generate(
                **inputs,
                generation_config=GenerationConfig(**self.options.generation_config),
            )
            generated_texts = self.processor.batch_decode(
                generated_ids[:, inputs["input_ids"].shape[1] :],
                skip_special_tokens=True,
            )

            yield generated_texts[0].strip()