Revise pipeline

Signed-off-by: Christoph Auer <cau@zurich.ibm.com>

docling/document_converter.py

@@ -1,8 +1,10 @@
 import hashlib
 import logging
 import sys
+import threading
 import time
 from collections.abc import Iterable, Iterator
+from concurrent.futures import ThreadPoolExecutor
 from functools import partial
 from pathlib import Path
 from typing import Dict, List, Optional, Tuple, Type, Union
@@ -50,6 +52,9 @@ from docling.utils.utils import chunkify
 
 _log = logging.getLogger(__name__)
 
+# Module-level lock for pipeline cache
+_pipeline_cache_lock = threading.Lock()
+
 
 class FormatOption(BaseModel):
     pipeline_cls: Type[BasePipeline]
@@ -287,7 +292,10 @@ class DocumentConverter:
             #with ThreadPoolExecutor(
             #    max_workers=settings.perf.doc_batch_concurrency
             #) as pool:
-            #   yield from pool.map(self.process_document, input_batch)
+            #    yield from pool.map(
+            #        partial(self._process_document, raises_on_error=raises_on_error),
+            #        input_batch,
+            #    )
             # Note: PDF backends are not thread-safe, thread pool usage was disabled.
 
             for item in map(
@@ -315,6 +323,7 @@ class DocumentConverter:
         # Use a composite key to cache pipelines
         cache_key = (pipeline_class, options_hash)
 
+        with _pipeline_cache_lock:
             if cache_key not in self.initialized_pipelines:
                 _log.info(
                     f"Initializing pipeline for {pipeline_class.__name__} with options hash {options_hash}"

tests/test_threaded_pipeline.py

@@ -1,50 +1,98 @@
+import logging
 import time
 from pathlib import Path
 from typing import List
 
 import pytest
 
-from docling.document_converter import DocumentConverter, PdfFormatOption
+from docling.datamodel.base_models import ConversionStatus, InputFormat
-from docling.datamodel.base_models import InputFormat, ConversionStatus
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options import (
     PdfPipelineOptions,
-    ThreadedPdfPipelineOptions
+    ThreadedPdfPipelineOptions,
 )
+from docling.document_converter import DocumentConverter, PdfFormatOption
 from docling.pipeline.standard_pdf_pipeline import StandardPdfPipeline
 from docling.pipeline.threaded_standard_pdf_pipeline import ThreadedStandardPdfPipeline
 
+
 def test_threaded_pipeline_multiple_documents():
-    """Test threaded pipeline with multiple documents"""
+    """Test threaded pipeline with multiple documents and compare with standard pipeline"""
-    converter = DocumentConverter(
+    test_files = [
+        "tests/data/pdf/2203.01017v2.pdf",
+        "tests/data/pdf/2206.01062.pdf",
+        "tests/data/pdf/2305.03393v1.pdf"
+    ]
+
+    # Standard pipeline
+    standard_converter = DocumentConverter(
         format_options={
             InputFormat.PDF: PdfFormatOption(
-                pipeline_cls=ThreadedStandardPdfPipeline,
+                pipeline_cls=StandardPdfPipeline,
-                pipeline_options=ThreadedPdfPipelineOptions(
+                pipeline_options=PdfPipelineOptions(
-                    layout_batch_size=48,
+                    do_table_structure=True,
-                    ocr_batch_size=24,
+                    do_ocr=True,
-                    batch_timeout_seconds=1.0,
+                ),
-                )
             )
         }
     )
 
-    # Test threaded pipeline with multiple documents
+    # Threaded pipeline
-    results = []
+    threaded_converter = DocumentConverter(
+        format_options={
+            InputFormat.PDF: PdfFormatOption(
+                pipeline_cls=ThreadedStandardPdfPipeline,
+                pipeline_options=ThreadedPdfPipelineOptions(
+                    layout_batch_size=1,
+                    table_batch_size=1,
+                    ocr_batch_size=1,
+                    batch_timeout_seconds=1.0,
+                    do_table_structure=True,
+                    do_ocr=True,
+                ),
+            )
+        }
+    )
+
+    # Test standard pipeline
+    standard_results = []
     start_time = time.perf_counter()
-    for result in converter.convert_all([
+    for result in standard_converter.convert_all(test_files, raises_on_error=True):
-        "tests/data/pdf/2206.01062.pdf", 
+        print("Finished converting document with standard pipeline:", result.input.file.name)
-        "tests/data/pdf/2305.03393v1.pdf"
+        standard_results.append(result)
-    ]):
+    standard_time = time.perf_counter() - start_time
-        results.append(result)
-    end_time = time.perf_counter()
 
-    conversion_duration = end_time - start_time
+    del standard_converter
-    print(f"Threaded multi-doc conversion took {conversion_duration:.2f} seconds")
 
-    assert len(results) == 2
+    # Test threaded pipeline
-    for result in results:
+    threaded_results = []
+    start_time = time.perf_counter()
+    for result in threaded_converter.convert_all(test_files, raises_on_error=True):
+        print("Finished converting document with threaded pipeline:", result.input.file.name)
+        threaded_results.append(result)
+    threaded_time = time.perf_counter() - start_time
+
+    del threaded_converter
+
+    print("\nMulti-document Pipeline Comparison:")
+    print(f"Standard pipeline:  {standard_time:.2f} seconds")
+    print(f"Threaded pipeline:  {threaded_time:.2f} seconds")
+    print(f"Speedup:            {standard_time / threaded_time:.2f}x")
+
+    # Verify results
+    assert len(standard_results) == len(threaded_results)
+    for result in standard_results:
         assert result.status == ConversionStatus.SUCCESS
+    for result in threaded_results:
+        assert result.status == ConversionStatus.SUCCESS
+
+    # Basic content comparison
+    for i, (standard_result, threaded_result) in enumerate(zip(standard_results, threaded_results)):
+        standard_doc = standard_result.document
+        threaded_doc = threaded_result.document
+        
+        assert len(standard_doc.pages) == len(threaded_doc.pages), f"Document {i} page count mismatch"
+        assert len(standard_doc.texts) == len(threaded_doc.texts), f"Document {i} text count mismatch"
 
 
 def test_pipeline_comparison():
@@ -73,7 +121,7 @@ def test_pipeline_comparison():
                     layout_batch_size=1,
                     ocr_batch_size=1,
                     table_batch_size=1,
-                )
+                ),
             )
         }
     )
@@ -82,14 +130,16 @@ def test_pipeline_comparison():
     threaded_results = list(threaded_converter.convert_all([test_file]))
     threaded_time = time.perf_counter() - start_time
 
-    print(f"\nPipeline Comparison:")
+    print("\nPipeline Comparison:")
     print(f"Sync pipeline:     {sync_time:.2f} seconds")
     print(f"Threaded pipeline: {threaded_time:.2f} seconds")
     print(f"Speedup:           {sync_time / threaded_time:.2f}x")
 
     # Verify results are equivalent
     assert len(sync_results) == len(threaded_results) == 1
-    assert sync_results[0].status == threaded_results[0].status == ConversionStatus.SUCCESS
+    assert (
+        sync_results[0].status == threaded_results[0].status == ConversionStatus.SUCCESS
+    )
 
     # Basic content comparison
     sync_doc = sync_results[0].document
@@ -99,9 +149,6 @@ def test_pipeline_comparison():
     assert len(sync_doc.texts) == len(threaded_doc.texts)
 
 
-
-
-
 if __name__ == "__main__":
     # Run basic performance test
     test_pipeline_comparison()