Observation for V2 Backend

[AdityaMannu1709]

Suppose I have 2 PDFs -
1st PDF has 460 pages and 2nd PDF has 17 pages
Docling V1
1st PDF takes ~8000 seconds and 2nd PDF takes ~100 seconds
Docling V2
1st PDF takes ~4000 seconds (50% reduced) and 2nd PDF takes ~600 seconds (increased drastically)

NOTE : All configurations are same (OCR on + TABLE accurate)

Do anyone have any explanation on this why time taken is being increased so much on PDFs having less than 100 pages in V2?

[dolfim-ibm]

@AdityaMannu1709 can you please provide the output of docling --version for the two versions you are comparing.

The latest version should actually have improved the performance considerably.

[AdityaMannu1709]

Environment:
OS: Windows
deepsearch-glm: 1.0.0
docling: 2.10.0
docling-core: 2.9.0
docling-ibm-models: 2.0.7
docling-parse: 3.0.0

Above, V1 refers to DoclingParseDocumentBackend and V2 refers to DoclingParseV2DocumentBackend.
As in latest release V2 was made as default backend for PDFs.
But above pattern in time was observed as it was performing great for large PDFs but not for small PDFs (instead increased parsing time)

[dolfim-ibm]

We won't be able to reproduce the same setup on Windows.

Would you be able to provide the profiling output like in this post?
#568 (comment)

We have then the tools to plot it and analyze.