docling 2.58.0__tar.gz → 2.60.0__tar.gz

{docling-2.58.0 → docling-2.60.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: docling
-Version: 2.58.0
+Version: 2.60.0
 Summary: SDK and CLI for parsing PDF, DOCX, HTML, and more, to a unified document representation for powering downstream workflows such as gen AI applications.
 Author-email: Christoph Auer <cau@zurich.ibm.com>, Michele Dolfi <dol@zurich.ibm.com>, Maxim Lysak <mly@zurich.ibm.com>, Nikos Livathinos <nli@zurich.ibm.com>, Ahmed Nassar <ahn@zurich.ibm.com>, Panos Vagenas <pva@zurich.ibm.com>, Peter Staar <taa@zurich.ibm.com>
 License-Expression: MIT
@@ -22,6 +22,7 @@ Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
 Requires-Python: <4.0,>=3.9
 Description-Content-Type: text/markdown
 License-File: LICENSE
@@ -45,7 +46,7 @@ Requires-Dist: beautifulsoup4<5.0.0,>=4.12.3
 Requires-Dist: pandas<3.0.0,>=2.1.4
 Requires-Dist: marko<3.0.0,>=2.1.2
 Requires-Dist: openpyxl<4.0.0,>=3.1.5
-Requires-Dist: lxml<6.0.0,>=4.0.0
+Requires-Dist: lxml<7.0.0,>=4.0.0
 Requires-Dist: pillow<12.0.0,>=10.0.0
 Requires-Dist: tqdm<5.0.0,>=4.65.0
 Requires-Dist: pluggy<2.0.0,>=1.0.0
@@ -62,15 +63,15 @@ Requires-Dist: ocrmac<2.0.0,>=1.0.0; sys_platform == "darwin" and extra == "ocrm
 Provides-Extra: vlm
 Requires-Dist: transformers<5.0.0,>=4.46.0; extra == "vlm"
 Requires-Dist: accelerate<2.0.0,>=1.2.1; extra == "vlm"
-Requires-Dist: mlx-vlm<1.0.0,>=0.3.0; (python_version >= "3.10" and sys_platform == "darwin" and platform_machine == "arm64") and extra == "vlm"
-Requires-Dist: vllm<1.0.0,>=0.10.0; (python_version >= "3.10" and sys_platform == "linux" and platform_machine == "x86_64") and extra == "vlm"
+Requires-Dist: mlx-vlm<1.0.0,>=0.3.0; (python_version >= "3.10" and python_version < "3.14" and sys_platform == "darwin" and platform_machine == "arm64") and extra == "vlm"
+Requires-Dist: vllm<1.0.0,>=0.10.0; (python_version >= "3.10" and python_version < "3.14" and sys_platform == "linux" and platform_machine == "x86_64") and extra == "vlm"
 Requires-Dist: qwen-vl-utils>=0.0.11; extra == "vlm"
 Provides-Extra: rapidocr
-Requires-Dist: rapidocr<4.0.0,>=3.3; python_version < "3.14" and extra == "rapidocr"
-Requires-Dist: onnxruntime<2.0.0,>=1.7.0; extra == "rapidocr"
+Requires-Dist: rapidocr<4.0.0,>=3.3; extra == "rapidocr"
+Requires-Dist: onnxruntime<2.0.0,>=1.7.0; python_version < "3.14" and extra == "rapidocr"
 Provides-Extra: asr
-Requires-Dist: mlx-whisper>=0.4.3; (python_version >= "3.10" and sys_platform == "darwin" and platform_machine == "arm64") and extra == "asr"
-Requires-Dist: openai-whisper>=20250625; extra == "asr"
+Requires-Dist: mlx-whisper>=0.4.3; (python_version >= "3.10" and python_version < "3.14" and sys_platform == "darwin" and platform_machine == "arm64") and extra == "asr"
+Requires-Dist: openai-whisper>=20250625; python_version < "3.14" and extra == "asr"
 Dynamic: license-file
 
 <p align="center">

{docling-2.58.0 → docling-2.60.0}/docling/backend/msexcel_backend.py

@@ -139,10 +139,14 @@ class MsExcelDocumentBackend(DeclarativeDocumentBackend, PaginatedDocumentBacken
         self.workbook = None
         try:
             if isinstance(self.path_or_stream, BytesIO):
-                self.workbook = load_workbook(filename=self.path_or_stream)
+                self.workbook = load_workbook(
+                    filename=self.path_or_stream, data_only=True
+                )
 
             elif isinstance(self.path_or_stream, Path):
-                self.workbook = load_workbook(filename=str(self.path_or_stream))
+                self.workbook = load_workbook(
+                    filename=str(self.path_or_stream), data_only=True
+                )
 
             self.valid = self.workbook is not None
         except Exception as e:

{docling-2.58.0 → docling-2.60.0}/docling/backend/pypdfium2_backend.py

@@ -229,10 +229,10 @@ class PyPdfiumPageBackend(PdfPageBackend):
                     b=max(cell.rect.to_bounding_box().b for cell in group),
                 )
 
-                assert self._ppage is not None
-                self.text_page = self._ppage.get_textpage()
+                assert self.text_page is not None
                 bbox = merged_bbox.to_bottom_left_origin(page_size.height)
-                merged_text = self.text_page.get_text_bounded(*bbox.as_tuple())
+                with pypdfium2_lock:
+                    merged_text = self.text_page.get_text_bounded(*bbox.as_tuple())
 
                 return TextCell(
                     index=group[0].index,
@@ -255,9 +255,9 @@ class PyPdfiumPageBackend(PdfPageBackend):
     def get_bitmap_rects(self, scale: float = 1) -> Iterable[BoundingBox]:
         AREA_THRESHOLD = 0  # 32 * 32
         page_size = self.get_size()
-        rotation = self._ppage.get_rotation()
 
         with pypdfium2_lock:
+            rotation = self._ppage.get_rotation()
             for obj in self._ppage.get_objects(filter=[pdfium_c.FPDF_PAGEOBJ_IMAGE]):
                 pos = obj.get_pos()
                 if rotation == 90:

{docling-2.58.0 → docling-2.60.0}/docling/cli/main.py

@@ -738,10 +738,15 @@ def convert(  # noqa: C901
 
                         pipeline_options.vlm_options = SMOLDOCLING_MLX
                     except ImportError:
-                        _log.warning(
-                            "To run SmolDocling faster, please install mlx-vlm:\n"
-                            "pip install mlx-vlm"
-                        )
+                        if sys.version_info < (3, 14):
+                            _log.warning(
+                                "To run SmolDocling faster, please install mlx-vlm:\n"
+                                "pip install mlx-vlm"
+                            )
+                        else:
+                            _log.warning(
+                                "You can run SmolDocling faster with MLX support, but it is unfortunately not yet available on Python 3.14."
+                            )
 
             elif vlm_model == VlmModelType.GRANITEDOCLING:
                 pipeline_options.vlm_options = GRANITEDOCLING_TRANSFORMERS
@@ -751,10 +756,16 @@ def convert(  # noqa: C901
 
                         pipeline_options.vlm_options = GRANITEDOCLING_MLX
                     except ImportError:
-                        _log.warning(
-                            "To run GraniteDocling faster, please install mlx-vlm:\n"
-                            "pip install mlx-vlm"
-                        )
+                        if sys.version_info < (3, 14):
+                            _log.warning(
+                                "To run GraniteDocling faster, please install mlx-vlm:\n"
+                                "pip install mlx-vlm"
+                            )
+                        else:
+                            _log.warning(
+                                "You can run GraniteDocling faster with MLX support, but it is unfortunately not yet available on Python 3.14."
+                            )
+
             elif vlm_model == VlmModelType.SMOLDOCLING_VLLM:
                 pipeline_options.vlm_options = SMOLDOCLING_VLLM
 

{docling-2.58.0 → docling-2.60.0}/docling/datamodel/base_models.py

@@ -207,6 +207,8 @@ class VlmPrediction(BaseModel):
     text: str = ""
     generated_tokens: list[VlmPredictionToken] = []
     generation_time: float = -1
+    num_tokens: Optional[int] = None
+    stop_reason: Optional[str] = None  # todo define an enum for possible stop reasons
 
 
 class ContainerElement(

{docling-2.58.0 → docling-2.60.0}/docling/datamodel/pipeline_options.py

@@ -361,15 +361,7 @@ class PdfPipelineOptions(PaginatedPipelineOptions):
 
     generate_parsed_pages: bool = False
 
-
-class ProcessingPipeline(str, Enum):
-    STANDARD = "standard"
-    VLM = "vlm"
-    ASR = "asr"
-
-
-class ThreadedPdfPipelineOptions(PdfPipelineOptions):
-    """Pipeline options for the threaded PDF pipeline with batching and backpressure control"""
+    ### Arguments for threaded PDF pipeline with batching and backpressure control
 
     # Batch sizes for different stages
     ocr_batch_size: int = 4
@@ -377,7 +369,18 @@ class ThreadedPdfPipelineOptions(PdfPipelineOptions):
     table_batch_size: int = 4
 
     # Timing control
-    batch_timeout_seconds: float = 2.0
+    batch_polling_interval_seconds: float = 0.5
 
     # Backpressure and queue control
     queue_max_size: int = 100
+
+
+class ProcessingPipeline(str, Enum):
+    LEGACY = "legacy"
+    STANDARD = "standard"
+    VLM = "vlm"
+    ASR = "asr"
+
+
+class ThreadedPdfPipelineOptions(PdfPipelineOptions):
+    """Pipeline options for the threaded PDF pipeline with batching and backpressure control"""

{docling-2.58.0 → docling-2.60.0}/docling/datamodel/pipeline_options_vlm_model.py

@@ -82,6 +82,7 @@ class InlineVlmOptions(BaseVlmOptions):
 
     use_kv_cache: bool = True
     max_new_tokens: int = 4096
+    track_generated_tokens: bool = False
 
     @property
     def repo_cache_folder(self) -> str:

{docling-2.58.0 → docling-2.60.0}/docling/models/api_vlm_model.py

@@ -73,7 +73,7 @@ class ApiVlmModel(BasePageModel):
                         # Skip non-GenerationStopper criteria (should have been caught in validation)
 
                     # Streaming path with early abort support
-                    page_tags = api_image_request_streaming(
+                    page_tags, num_tokens = api_image_request_streaming(
                         image=hi_res_image,
                         prompt=prompt,
                         url=self.vlm_options.url,
@@ -84,7 +84,7 @@ class ApiVlmModel(BasePageModel):
                     )
                 else:
                     # Non-streaming fallback (existing behavior)
-                    page_tags = api_image_request(
+                    page_tags, num_tokens = api_image_request(
                         image=hi_res_image,
                         prompt=prompt,
                         url=self.vlm_options.url,
@@ -94,7 +94,9 @@ class ApiVlmModel(BasePageModel):
                     )
 
                 page_tags = self.vlm_options.decode_response(page_tags)
-                page.predictions.vlm_response = VlmPrediction(text=page_tags)
+                page.predictions.vlm_response = VlmPrediction(
+                    text=page_tags, num_tokens=num_tokens
+                )
             return page
 
         with ThreadPoolExecutor(max_workers=self.concurrency) as executor:

{docling-2.58.0 → docling-2.60.0}/docling/models/layout_model.py

@@ -167,6 +167,10 @@ class LayoutModel(BasePageModel):
             valid_pages.append(page)
             valid_page_images.append(page_image)
 
+        print(f"{len(pages)=}, {pages[0].page_no}-{pages[-1].page_no}")
+        print(f"{len(valid_pages)=}")
+        print(f"{len(valid_page_images)=}")
+
         # Process all valid pages with batch prediction
         batch_predictions = []
         if valid_page_images:

{docling-2.58.0 → docling-2.60.0}/docling/models/picture_description_vlm_model.py

@@ -1,3 +1,4 @@
+import sys
 import threading
 from collections.abc import Iterable
 from pathlib import Path
@@ -75,7 +76,10 @@ class PictureDescriptionVlmModel(
                         else "sdpa"
                     ),
                 )
-                self.model = torch.compile(self.model)  # type: ignore
+                if sys.version_info < (3, 14):
+                    self.model = torch.compile(self.model)  # type: ignore
+                else:
+                    self.model.eval()
 
             self.provenance = f"{self.options.repo_id}"
 

{docling-2.58.0 → docling-2.60.0}/docling/models/vlm_models_inline/hf_transformers_model.py

@@ -1,5 +1,6 @@
 import importlib.metadata
 import logging
+import sys
 import time
 from collections.abc import Iterable
 from pathlib import Path
@@ -129,7 +130,10 @@ class HuggingFaceTransformersVlmModel(BaseVlmPageModel, HuggingFaceModelDownload
                 trust_remote_code=vlm_options.trust_remote_code,
                 revision=vlm_options.revision,
             )
-            self.vlm_model = torch.compile(self.vlm_model)  # type: ignore
+            if sys.version_info < (3, 14):
+                self.vlm_model = torch.compile(self.vlm_model)  # type: ignore
+            else:
+                self.vlm_model.eval()
 
             # Load generation config
             self.generation_config = GenerationConfig.from_pretrained(
@@ -363,13 +367,19 @@ class HuggingFaceTransformersVlmModel(BaseVlmPageModel, HuggingFaceModelDownload
             decoded_texts = [text.rstrip(pad_token) for text in decoded_texts]
 
         # -- Optional logging
+        num_tokens = None
         if generated_ids.shape[0] > 0:
+            num_tokens = int(generated_ids[0].shape[0])
             _log.debug(
-                f"Generated {int(generated_ids[0].shape[0])} tokens in {generation_time:.2f}s "
+                f"Generated {num_tokens} tokens in {generation_time:.2f}s "
                 f"for batch size {generated_ids.shape[0]}."
             )
 
         for text in decoded_texts:
             # Apply decode_response to the output text
             decoded_text = self.vlm_options.decode_response(text)
-            yield VlmPrediction(text=decoded_text, generation_time=generation_time)
+            yield VlmPrediction(
+                text=decoded_text,
+                generation_time=generation_time,
+                num_tokens=num_tokens,
+            )

{docling-2.58.0 → docling-2.60.0}/docling/models/vlm_models_inline/mlx_model.py

@@ -50,9 +50,14 @@ class HuggingFaceMlxModel(BaseVlmPageModel, HuggingFaceModelDownloadMixin):
                 from mlx_vlm.prompt_utils import apply_chat_template  # type: ignore
                 from mlx_vlm.utils import load_config  # type: ignore
             except ImportError:
-                raise ImportError(
-                    "mlx-vlm is not installed. Please install it via `pip install mlx-vlm` to use MLX VLM models."
-                )
+                if sys.version_info < (3, 14):
+                    raise ImportError(
+                        "mlx-vlm is not installed. Please install it via `pip install mlx-vlm` to use MLX VLM models."
+                    )
+                else:
+                    raise ImportError(
+                        "mlx-vlm is not installed. It is not yet available on Python 3.14."
+                    )
 
             repo_cache_folder = vlm_options.repo_id.replace("/", "--")
 
@@ -313,5 +318,6 @@ class HuggingFaceMlxModel(BaseVlmPageModel, HuggingFaceModelDownloadMixin):
                     text=decoded_output,
                     generation_time=generation_time,
                     generated_tokens=tokens,
+                    num_tokens=len(tokens),
                 )
             _log.debug("MLX model: Released global lock")

{docling-2.58.0 → docling-2.60.0}/docling/models/vlm_models_inline/nuextract_transformers_model.py

@@ -1,4 +1,5 @@
 import logging
+import sys
 import time
 from collections.abc import Iterable
 from pathlib import Path
@@ -153,7 +154,10 @@ class NuExtractTransformersModel(BaseVlmModel, HuggingFaceModelDownloadMixin):
                 ),
                 trust_remote_code=vlm_options.trust_remote_code,
             )
-            self.vlm_model = torch.compile(self.vlm_model)  # type: ignore
+            if sys.version_info < (3, 14):
+                self.vlm_model = torch.compile(self.vlm_model)  # type: ignore
+            else:
+                self.vlm_model.eval()
 
             # Load generation config
             self.generation_config = GenerationConfig.from_pretrained(artifacts_path)
@@ -278,13 +282,19 @@ class NuExtractTransformersModel(BaseVlmModel, HuggingFaceModelDownloadMixin):
         )
 
         # Optional logging
+        num_tokens = None
         if generated_ids.shape[0] > 0:  # type: ignore
+            num_tokens = int(generated_ids[0].shape[0])
             _log.debug(
-                f"Generated {int(generated_ids[0].shape[0])} tokens in {generation_time:.2f}s "
+                f"Generated {num_tokens} tokens in {generation_time:.2f}s "
                 f"for batch size {generated_ids.shape[0]}."  # type: ignore
             )
 
         for text in decoded_texts:
             # Apply decode_response to the output text
             decoded_text = self.vlm_options.decode_response(text)
-            yield VlmPrediction(text=decoded_text, generation_time=generation_time)
+            yield VlmPrediction(
+                text=decoded_text,
+                generation_time=generation_time,
+                num_tokens=num_tokens,
+            )

{docling-2.58.0 → docling-2.60.0}/docling/models/vlm_models_inline/vllm_model.py

@@ -1,4 +1,5 @@
 import logging
+import sys
 import time
 from collections.abc import Iterable
 from pathlib import Path
@@ -8,7 +9,7 @@ import numpy as np
 from PIL.Image import Image
 
 from docling.datamodel.accelerator_options import AcceleratorOptions
-from docling.datamodel.base_models import Page, VlmPrediction
+from docling.datamodel.base_models import Page, VlmPrediction, VlmPredictionToken
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options_vlm_model import (
     InlineVlmOptions,
@@ -87,7 +88,7 @@ class VllmVlmModel(BaseVlmPageModel, HuggingFaceModelDownloadMixin):
         vlm_options: InlineVlmOptions,
     ):
         self.enabled = enabled
-        self.vlm_options = vlm_options
+        self.vlm_options: InlineVlmOptions = vlm_options
 
         self.llm = None
         self.sampling_params = None
@@ -100,7 +101,18 @@ class VllmVlmModel(BaseVlmPageModel, HuggingFaceModelDownloadMixin):
             return
 
         from transformers import AutoProcessor
-        from vllm import LLM, SamplingParams
+
+        try:
+            from vllm import LLM, SamplingParams
+        except ImportError:
+            if sys.version_info < (3, 14):
+                raise ImportError(
+                    "vllm is not installed. Please install it via `pip install vllm`."
+                )
+            else:
+                raise ImportError(
+                    "vllm is not installed. It is not yet available on Python 3.14."
+                )
 
         # Device selection
         self.device = decide_device(
@@ -222,7 +234,8 @@ class VllmVlmModel(BaseVlmPageModel, HuggingFaceModelDownloadMixin):
                     pages_with_images.append(page)
 
                 if images:
-                    predictions = list(self.process_images(images, user_prompts))
+                    with TimeRecorder(conv_res, "vlm_inference"):
+                        predictions = list(self.process_images(images, user_prompts))
                     for page, prediction in zip(pages_with_images, predictions):
                         page.predictions.vlm_response = prediction
 
@@ -288,13 +301,34 @@ class VllmVlmModel(BaseVlmPageModel, HuggingFaceModelDownloadMixin):
         # Optional debug
         if outputs:
             try:
-                num_tokens = len(outputs[0].outputs[0].token_ids)
-                _log.debug(f"Generated {num_tokens} tokens in {generation_time:.2f}s.")
+                num_tokens_within_batch = len(outputs[0].outputs[0].token_ids)
+                _log.debug(
+                    f"Generated {num_tokens_within_batch} tokens for batch in {generation_time:.2f}s."
+                )
             except Exception:
-                pass
+                num_tokens_within_batch = 0
 
         # Emit predictions
         for output in outputs:
             text = output.outputs[0].text if output.outputs else ""
+            stop_reason = output.outputs[0].stop_reason if output.outputs else ""
+            generated_tokens = [
+                VlmPredictionToken(token=int(p)) for p in output.outputs[0].token_ids
+            ]
+            num_tokens = len(generated_tokens)
             decoded_text = self.vlm_options.decode_response(text)
-            yield VlmPrediction(text=decoded_text, generation_time=generation_time)
+            if self.vlm_options.track_generated_tokens:
+                yield VlmPrediction(
+                    text=decoded_text,
+                    generation_time=generation_time,
+                    num_tokens=num_tokens,
+                    stop_reason=stop_reason,
+                    generated_tokens=generated_tokens,
+                )
+            else:
+                yield VlmPrediction(
+                    text=decoded_text,
+                    generation_time=generation_time,
+                    num_tokens=num_tokens,
+                    stop_reason=stop_reason,
+                )

{docling-2.58.0 → docling-2.60.0}/docling/pipeline/asr_pipeline.py

@@ -1,6 +1,7 @@
 import logging
 import os
 import re
+import sys
 import tempfile
 from io import BytesIO
 from pathlib import Path
@@ -117,9 +118,15 @@ class _NativeWhisperModel:
             try:
                 import whisper  # type: ignore
             except ImportError:
-                raise ImportError(
-                    "whisper is not installed. Please install it via `pip install openai-whisper` or do `uv sync --extra asr`."
-                )
+                if sys.version_info < (3, 14):
+                    raise ImportError(
+                        "whisper is not installed. Please install it via `pip install openai-whisper` or do `uv sync --extra asr`."
+                    )
+                else:
+                    raise ImportError(
+                        "whisper is not installed. Unfortunately its dependencies are not yet available for Python 3.14."
+                    )
+
             self.asr_options = asr_options
             self.max_tokens = asr_options.max_new_tokens
             self.temperature = asr_options.temperature

docling-2.58.0/docling/pipeline/standard_pdf_pipeline.py → docling-2.60.0/docling/pipeline/legacy_standard_pdf_pipeline.py

@@ -31,7 +31,7 @@ from docling.utils.profiling import ProfilingScope, TimeRecorder
 _log = logging.getLogger(__name__)
 
 
-class StandardPdfPipeline(PaginatedPipeline):
+class LegacyStandardPdfPipeline(PaginatedPipeline):
     def __init__(self, pipeline_options: PdfPipelineOptions):
         super().__init__(pipeline_options)
         self.pipeline_options: PdfPipelineOptions
@@ -102,7 +102,7 @@ class StandardPdfPipeline(PaginatedPipeline):
         local_dir: Optional[Path] = None, force: bool = False
     ) -> Path:
         warnings.warn(
-            "The usage of StandardPdfPipeline.download_models_hf() is deprecated "
+            "The usage of LegacyStandardPdfPipeline.download_models_hf() is deprecated "
             "use instead the utility `docling-tools models download`, or "
             "the upstream method docling.utils.models_downloader.download_all()",
             DeprecationWarning,