docling 2.34.0__py3-none-any.whl → 2.36.0__py3-none-any.whl

docling/datamodel/vlm_model_specs.py

@@ -0,0 +1,144 @@
+import logging
+from enum import Enum
+
+from pydantic import (
+    AnyUrl,
+)
+
+from docling.datamodel.accelerator_options import AcceleratorDevice
+from docling.datamodel.pipeline_options_vlm_model import (
+    ApiVlmOptions,
+    InferenceFramework,
+    InlineVlmOptions,
+    ResponseFormat,
+    TransformersModelType,
+)
+
+_log = logging.getLogger(__name__)
+
+
+# SmolDocling
+SMOLDOCLING_MLX = InlineVlmOptions(
+    repo_id="ds4sd/SmolDocling-256M-preview-mlx-bf16",
+    prompt="Convert this page to docling.",
+    response_format=ResponseFormat.DOCTAGS,
+    inference_framework=InferenceFramework.MLX,
+    supported_devices=[AcceleratorDevice.MPS],
+    scale=2.0,
+    temperature=0.0,
+)
+
+SMOLDOCLING_TRANSFORMERS = InlineVlmOptions(
+    repo_id="ds4sd/SmolDocling-256M-preview",
+    prompt="Convert this page to docling.",
+    response_format=ResponseFormat.DOCTAGS,
+    inference_framework=InferenceFramework.TRANSFORMERS,
+    transformers_model_type=TransformersModelType.AUTOMODEL_VISION2SEQ,
+    supported_devices=[
+        AcceleratorDevice.CPU,
+        AcceleratorDevice.CUDA,
+        AcceleratorDevice.MPS,
+    ],
+    scale=2.0,
+    temperature=0.0,
+)
+
+# GraniteVision
+GRANITE_VISION_TRANSFORMERS = InlineVlmOptions(
+    repo_id="ibm-granite/granite-vision-3.2-2b",
+    prompt="Convert this page to markdown. Do not miss any text and only output the bare markdown!",
+    response_format=ResponseFormat.MARKDOWN,
+    inference_framework=InferenceFramework.TRANSFORMERS,
+    transformers_model_type=TransformersModelType.AUTOMODEL_VISION2SEQ,
+    supported_devices=[
+        AcceleratorDevice.CPU,
+        AcceleratorDevice.CUDA,
+        AcceleratorDevice.MPS,
+    ],
+    scale=2.0,
+    temperature=0.0,
+)
+
+GRANITE_VISION_OLLAMA = ApiVlmOptions(
+    url=AnyUrl("http://localhost:11434/v1/chat/completions"),
+    params={"model": "granite3.2-vision:2b"},
+    prompt="Convert this page to markdown. Do not miss any text and only output the bare markdown!",
+    scale=1.0,
+    timeout=120,
+    response_format=ResponseFormat.MARKDOWN,
+    temperature=0.0,
+)
+
+# Pixtral
+PIXTRAL_12B_TRANSFORMERS = InlineVlmOptions(
+    repo_id="mistral-community/pixtral-12b",
+    prompt="Convert this page to markdown. Do not miss any text and only output the bare markdown!",
+    response_format=ResponseFormat.MARKDOWN,
+    inference_framework=InferenceFramework.TRANSFORMERS,
+    transformers_model_type=TransformersModelType.AUTOMODEL_VISION2SEQ,
+    supported_devices=[AcceleratorDevice.CPU, AcceleratorDevice.CUDA],
+    scale=2.0,
+    temperature=0.0,
+)
+
+PIXTRAL_12B_MLX = InlineVlmOptions(
+    repo_id="mlx-community/pixtral-12b-bf16",
+    prompt="Convert this page to markdown. Do not miss any text and only output the bare markdown!",
+    response_format=ResponseFormat.MARKDOWN,
+    inference_framework=InferenceFramework.MLX,
+    supported_devices=[AcceleratorDevice.MPS],
+    scale=2.0,
+    temperature=0.0,
+)
+
+# Phi4
+PHI4_TRANSFORMERS = InlineVlmOptions(
+    repo_id="microsoft/Phi-4-multimodal-instruct",
+    prompt="Convert this page to MarkDown. Do not miss any text and only output the bare markdown",
+    trust_remote_code=True,
+    response_format=ResponseFormat.MARKDOWN,
+    inference_framework=InferenceFramework.TRANSFORMERS,
+    transformers_model_type=TransformersModelType.AUTOMODEL_CAUSALLM,
+    supported_devices=[AcceleratorDevice.CPU, AcceleratorDevice.CUDA],
+    scale=2.0,
+    temperature=0.0,
+    extra_generation_config=dict(num_logits_to_keep=0),
+)
+
+# Qwen
+QWEN25_VL_3B_MLX = InlineVlmOptions(
+    repo_id="mlx-community/Qwen2.5-VL-3B-Instruct-bf16",
+    prompt="Convert this page to markdown. Do not miss any text and only output the bare markdown!",
+    response_format=ResponseFormat.MARKDOWN,
+    inference_framework=InferenceFramework.MLX,
+    supported_devices=[AcceleratorDevice.MPS],
+    scale=2.0,
+    temperature=0.0,
+)
+
+# Gemma-3
+GEMMA3_12B_MLX = InlineVlmOptions(
+    repo_id="mlx-community/gemma-3-12b-it-bf16",
+    prompt="Convert this page to markdown. Do not miss any text and only output the bare markdown!",
+    response_format=ResponseFormat.MARKDOWN,
+    inference_framework=InferenceFramework.MLX,
+    supported_devices=[AcceleratorDevice.MPS],
+    scale=2.0,
+    temperature=0.0,
+)
+
+GEMMA3_27B_MLX = InlineVlmOptions(
+    repo_id="mlx-community/gemma-3-27b-it-bf16",
+    prompt="Convert this page to markdown. Do not miss any text and only output the bare markdown!",
+    response_format=ResponseFormat.MARKDOWN,
+    inference_framework=InferenceFramework.MLX,
+    supported_devices=[AcceleratorDevice.MPS],
+    scale=2.0,
+    temperature=0.0,
+)
+
+
+class VlmModelType(str, Enum):
+    SMOLDOCLING = "smoldocling"
+    GRANITE_VISION = "granite_vision"
+    GRANITE_VISION_OLLAMA = "granite_vision_ollama"

docling/document_converter.py

@@ -186,6 +186,11 @@ class DocumentConverter:
             Tuple[Type[BasePipeline], str], BasePipeline
         ] = {}
 
+    def _get_initialized_pipelines(
+        self,
+    ) -> dict[tuple[Type[BasePipeline], str], BasePipeline]:
+        return self.initialized_pipelines
+
     def _get_pipeline_options_hash(self, pipeline_options: PipelineOptions) -> str:
         """Generate a hash of pipeline options to use as part of the cache key."""
         options_str = str(pipeline_options.model_dump())

docling/models/api_vlm_model.py

@@ -3,7 +3,7 @@ from concurrent.futures import ThreadPoolExecutor
 
 from docling.datamodel.base_models import Page, VlmPrediction
 from docling.datamodel.document import ConversionResult
-from docling.datamodel.pipeline_options import ApiVlmOptions
+from docling.datamodel.pipeline_options_vlm_model import ApiVlmOptions
 from docling.exceptions import OperationNotAllowed
 from docling.models.base_model import BasePageModel
 from docling.utils.api_image_request import api_image_request

docling/models/base_ocr_model.py

@@ -11,9 +11,10 @@ from PIL import Image, ImageDraw
 from rtree import index
 from scipy.ndimage import binary_dilation, find_objects, label
 
+from docling.datamodel.accelerator_options import AcceleratorOptions
 from docling.datamodel.base_models import Page
 from docling.datamodel.document import ConversionResult
-from docling.datamodel.pipeline_options import AcceleratorOptions, OcrOptions
+from docling.datamodel.pipeline_options import OcrOptions
 from docling.datamodel.settings import settings
 from docling.models.base_model import BaseModelWithOptions, BasePageModel
 

docling/models/code_formula_model.py

@@ -16,9 +16,10 @@ from docling_core.types.doc.labels import CodeLanguageLabel
 from PIL import Image, ImageOps
 from pydantic import BaseModel
 
+from docling.datamodel.accelerator_options import AcceleratorOptions
 from docling.datamodel.base_models import ItemAndImageEnrichmentElement
-from docling.datamodel.pipeline_options import AcceleratorOptions
 from docling.models.base_model import BaseItemAndImageEnrichmentModel
+from docling.models.utils.hf_model_download import download_hf_model
 from docling.utils.accelerator_utils import decide_device
 
 
@@ -117,20 +118,14 @@ class CodeFormulaModel(BaseItemAndImageEnrichmentModel):
         force: bool = False,
         progress: bool = False,
     ) -> Path:
-        from huggingface_hub import snapshot_download
-        from huggingface_hub.utils import disable_progress_bars
-
-        if not progress:
-            disable_progress_bars()
-        download_path = snapshot_download(
+        return download_hf_model(
             repo_id="ds4sd/CodeFormula",
-            force_download=force,
-            local_dir=local_dir,
             revision="v1.0.2",
+            local_dir=local_dir,
+            force=force,
+            progress=progress,
         )
 
-        return Path(download_path)
-
     def is_processable(self, doc: DoclingDocument, element: NodeItem) -> bool:
         """
         Determines if a given element in a document can be processed by the model.

docling/models/document_picture_classifier.py

@@ -13,8 +13,9 @@ from docling_core.types.doc import (
 from PIL import Image
 from pydantic import BaseModel
 
-from docling.datamodel.pipeline_options import AcceleratorOptions
+from docling.datamodel.accelerator_options import AcceleratorOptions
 from docling.models.base_model import BaseEnrichmentModel
+from docling.models.utils.hf_model_download import download_hf_model
 from docling.utils.accelerator_utils import decide_device
 
 
@@ -105,20 +106,14 @@ class DocumentPictureClassifier(BaseEnrichmentModel):
     def download_models(
         local_dir: Optional[Path] = None, force: bool = False, progress: bool = False
     ) -> Path:
-        from huggingface_hub import snapshot_download
-        from huggingface_hub.utils import disable_progress_bars
-
-        if not progress:
-            disable_progress_bars()
-        download_path = snapshot_download(
+        return download_hf_model(
             repo_id="ds4sd/DocumentFigureClassifier",
-            force_download=force,
-            local_dir=local_dir,
             revision="v1.0.1",
+            local_dir=local_dir,
+            force=force,
+            progress=progress,
         )
 
-        return Path(download_path)
-
     def is_processable(self, doc: DoclingDocument, element: NodeItem) -> bool:
         """
         Determines if the given element can be processed by the classifier.

docling/models/easyocr_model.py

@@ -9,11 +9,10 @@ import numpy
 from docling_core.types.doc import BoundingBox, CoordOrigin
 from docling_core.types.doc.page import BoundingRectangle, TextCell
 
+from docling.datamodel.accelerator_options import AcceleratorDevice, AcceleratorOptions
 from docling.datamodel.base_models import Page
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options import (
-    AcceleratorDevice,
-    AcceleratorOptions,
     EasyOcrOptions,
     OcrOptions,
 )

docling/models/layout_model.py

@@ -10,11 +10,12 @@ from docling_core.types.doc import DocItemLabel
 from docling_ibm_models.layoutmodel.layout_predictor import LayoutPredictor
 from PIL import Image
 
+from docling.datamodel.accelerator_options import AcceleratorOptions
 from docling.datamodel.base_models import BoundingBox, Cluster, LayoutPrediction, Page
 from docling.datamodel.document import ConversionResult
-from docling.datamodel.pipeline_options import AcceleratorOptions
 from docling.datamodel.settings import settings
 from docling.models.base_model import BasePageModel
+from docling.models.utils.hf_model_download import download_hf_model
 from docling.utils.accelerator_utils import decide_device
 from docling.utils.layout_postprocessor import LayoutPostprocessor
 from docling.utils.profiling import TimeRecorder
@@ -83,20 +84,14 @@ class LayoutModel(BasePageModel):
         force: bool = False,
         progress: bool = False,
     ) -> Path:
-        from huggingface_hub import snapshot_download
-        from huggingface_hub.utils import disable_progress_bars
-
-        if not progress:
-            disable_progress_bars()
-        download_path = snapshot_download(
+        return download_hf_model(
             repo_id="ds4sd/docling-models",
-            force_download=force,
+            revision="v2.2.0",
             local_dir=local_dir,
-            revision="v2.1.0",
+            force=force,
+            progress=progress,
         )
 
-        return Path(download_path)
-
     def draw_clusters_and_cells_side_by_side(
         self, conv_res, page, clusters, mode_prefix: str, show: bool = False
     ):
@@ -185,13 +180,23 @@ class LayoutModel(BasePageModel):
                     ).postprocess()
                     # processed_clusters, processed_cells = clusters, page.cells
 
-                    conv_res.confidence.pages[page.page_no].layout_score = float(
-                        np.mean([c.confidence for c in processed_clusters])
-                    )
+                    with warnings.catch_warnings():
+                        warnings.filterwarnings(
+                            "ignore",
+                            "Mean of empty slice|invalid value encountered in scalar divide",
+                            RuntimeWarning,
+                            "numpy",
+                        )
 
-                    conv_res.confidence.pages[page.page_no].ocr_score = float(
-                        np.mean([c.confidence for c in processed_cells if c.from_ocr])
-                    )
+                        conv_res.confidence.pages[page.page_no].layout_score = float(
+                            np.mean([c.confidence for c in processed_clusters])
+                        )
+
+                        conv_res.confidence.pages[page.page_no].ocr_score = float(
+                            np.mean(
+                                [c.confidence for c in processed_cells if c.from_ocr]
+                            )
+                        )
 
                     page.cells = processed_cells
                     page.predictions.layout = LayoutPrediction(

docling/models/ocr_mac_model.py

@@ -8,10 +8,10 @@ from typing import Optional, Type
 from docling_core.types.doc import BoundingBox, CoordOrigin
 from docling_core.types.doc.page import BoundingRectangle, TextCell
 
+from docling.datamodel.accelerator_options import AcceleratorOptions
 from docling.datamodel.base_models import Page
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options import (
-    AcceleratorOptions,
     OcrMacOptions,
     OcrOptions,
 )

docling/models/page_preprocessing_model.py

@@ -1,4 +1,5 @@
 import re
+import warnings
 from collections.abc import Iterable
 from pathlib import Path
 from typing import Optional
@@ -7,7 +8,7 @@ import numpy as np
 from PIL import ImageDraw
 from pydantic import BaseModel
 
-from docling.datamodel.base_models import Page, ScoreValue
+from docling.datamodel.base_models import Page
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.settings import settings
 from docling.models.base_model import BasePageModel
@@ -76,11 +77,15 @@ class PagePreprocessingModel(BasePageModel):
             score = self.rate_text_quality(c.text)
             text_scores.append(score)
 
-        conv_res.confidence.pages[page.page_no].parse_score = float(
-            np.nanquantile(
-                text_scores, q=0.10
-            )  # To emphasise problems in the parse_score, we take the 10% percentile score of all text cells.
-        )
+        with warnings.catch_warnings():
+            warnings.filterwarnings(
+                "ignore", "Mean of empty slice", RuntimeWarning, "numpy"
+            )
+            conv_res.confidence.pages[page.page_no].parse_score = float(
+                np.nanquantile(
+                    text_scores, q=0.10
+                )  # To emphasise problems in the parse_score, we take the 10% percentile score of all text cells.
+            )
 
         # DEBUG code:
         def draw_text_boxes(image, cells, show: bool = False):

docling/models/picture_description_api_model.py

@@ -5,8 +5,8 @@ from typing import Optional, Type, Union
 
 from PIL import Image
 
+from docling.datamodel.accelerator_options import AcceleratorOptions
 from docling.datamodel.pipeline_options import (
-    AcceleratorOptions,
     PictureDescriptionApiOptions,
     PictureDescriptionBaseOptions,
 )

docling/models/picture_description_base_model.py

@@ -13,8 +13,8 @@ from docling_core.types.doc.document import (  # TODO: move import to docling_co
 )
 from PIL import Image
 
+from docling.datamodel.accelerator_options import AcceleratorOptions
 from docling.datamodel.pipeline_options import (
-    AcceleratorOptions,
     PictureDescriptionBaseOptions,
 )
 from docling.models.base_model import (

docling/models/picture_description_vlm_model.py

@@ -4,16 +4,21 @@ from typing import Optional, Type, Union
 
 from PIL import Image
 
+from docling.datamodel.accelerator_options import AcceleratorOptions
 from docling.datamodel.pipeline_options import (
-    AcceleratorOptions,
     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
 
 
-class PictureDescriptionVlmModel(PictureDescriptionBaseModel):
+class PictureDescriptionVlmModel(
+    PictureDescriptionBaseModel, HuggingFaceModelDownloadMixin
+):
     @classmethod
     def get_options_type(cls) -> Type[PictureDescriptionBaseOptions]:
         return PictureDescriptionVlmOptions
@@ -66,26 +71,6 @@ class PictureDescriptionVlmModel(PictureDescriptionBaseModel):
 
             self.provenance = f"{self.options.repo_id}"
 
-    @staticmethod
-    def download_models(
-        repo_id: str,
-        local_dir: Optional[Path] = None,
-        force: bool = False,
-        progress: bool = False,
-    ) -> Path:
-        from huggingface_hub import snapshot_download
-        from huggingface_hub.utils import disable_progress_bars
-
-        if not progress:
-            disable_progress_bars()
-        download_path = snapshot_download(
-            repo_id=repo_id,
-            force_download=force,
-            local_dir=local_dir,
-        )
-
-        return Path(download_path)
-
     def _annotate_images(self, images: Iterable[Image.Image]) -> Iterable[str]:
         from transformers import GenerationConfig
 

docling/models/rapid_ocr_model.py

@@ -7,11 +7,10 @@ import numpy
 from docling_core.types.doc import BoundingBox, CoordOrigin
 from docling_core.types.doc.page import BoundingRectangle, TextCell
 
+from docling.datamodel.accelerator_options import AcceleratorDevice, AcceleratorOptions
 from docling.datamodel.base_models import Page
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options import (
-    AcceleratorDevice,
-    AcceleratorOptions,
     OcrOptions,
     RapidOcrOptions,
 )

docling/models/table_structure_model.py

@@ -13,16 +13,16 @@ from docling_core.types.doc.page import (
 from docling_ibm_models.tableformer.data_management.tf_predictor import TFPredictor
 from PIL import ImageDraw
 
+from docling.datamodel.accelerator_options import AcceleratorDevice, AcceleratorOptions
 from docling.datamodel.base_models import Page, Table, TableStructurePrediction
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options import (
-    AcceleratorDevice,
-    AcceleratorOptions,
     TableFormerMode,
     TableStructureOptions,
 )
 from docling.datamodel.settings import settings
 from docling.models.base_model import BasePageModel
+from docling.models.utils.hf_model_download import download_hf_model
 from docling.utils.accelerator_utils import decide_device
 from docling.utils.profiling import TimeRecorder
 
@@ -90,20 +90,14 @@ class TableStructureModel(BasePageModel):
     def download_models(
         local_dir: Optional[Path] = None, force: bool = False, progress: bool = False
     ) -> Path:
-        from huggingface_hub import snapshot_download
-        from huggingface_hub.utils import disable_progress_bars
-
-        if not progress:
-            disable_progress_bars()
-        download_path = snapshot_download(
+        return download_hf_model(
             repo_id="ds4sd/docling-models",
-            force_download=force,
-            local_dir=local_dir,
             revision="v2.2.0",
+            local_dir=local_dir,
+            force=force,
+            progress=progress,
         )
 
-        return Path(download_path)
-
     def draw_table_and_cells(
         self,
         conv_res: ConversionResult,

docling/models/tesseract_ocr_cli_model.py

@@ -13,10 +13,10 @@ import pandas as pd
 from docling_core.types.doc import BoundingBox, CoordOrigin
 from docling_core.types.doc.page import TextCell
 
+from docling.datamodel.accelerator_options import AcceleratorOptions
 from docling.datamodel.base_models import Page
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options import (
-    AcceleratorOptions,
     OcrOptions,
     TesseractCliOcrOptions,
 )

docling/models/tesseract_ocr_model.py

@@ -7,10 +7,10 @@ from typing import Iterable, Optional, Type
 from docling_core.types.doc import BoundingBox, CoordOrigin
 from docling_core.types.doc.page import TextCell
 
+from docling.datamodel.accelerator_options import AcceleratorOptions
 from docling.datamodel.base_models import Page
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options import (
-    AcceleratorOptions,
     OcrOptions,
     TesseractOcrOptions,
 )

docling/models/utils/hf_model_download.py

@@ -0,0 +1,40 @@
+import logging
+from pathlib import Path
+from typing import Optional
+
+_log = logging.getLogger(__name__)
+
+
+def download_hf_model(
+    repo_id: str,
+    local_dir: Optional[Path] = None,
+    force: bool = False,
+    progress: bool = False,
+    revision: Optional[str] = None,
+) -> Path:
+    from huggingface_hub import snapshot_download
+    from huggingface_hub.utils import disable_progress_bars
+
+    if not progress:
+        disable_progress_bars()
+    download_path = snapshot_download(
+        repo_id=repo_id,
+        force_download=force,
+        local_dir=local_dir,
+        revision=revision,
+    )
+
+    return Path(download_path)
+
+
+class HuggingFaceModelDownloadMixin:
+    @staticmethod
+    def download_models(
+        repo_id: str,
+        local_dir: Optional[Path] = None,
+        force: bool = False,
+        progress: bool = False,
+    ) -> Path:
+        return download_hf_model(
+            repo_id=repo_id, local_dir=local_dir, force=force, progress=progress
+        )