Docling 2.10.0: Performance Degradation When Reading Large PDF Files

[langzichai]

Describe the problem
While using the latest version of Docling (2.10.0), the following issues were observed with PDF file read speed performance:

For small files (e.g. 3M), the performance improvement is significant and the read speed does increase by about 30%.
However, for large files (e.g. more than 60M), the read speed has not improved and still maintains a slower processing time, the same as in the previous version (2.9.0).

gpu ： V100 32g

A solution or guidance on how to tune performance for large file scenarios would be appreciated!

< v2.10.0
3M
default ocr 66s
63M
default ocr 1059.538206 seconds

v2.10.0
3M
default ocr 43.967382

63M
default ocr 1053s

def convert(filepath): start_time = datetime.now() print(f"start_time : {start_time}") input_doc = Path(filepath) pipeline_options = PdfPipelineOptions() pipeline_options.do_ocr = True pipeline_options.ocr_options.use_gpu = True pipeline_options.do_table_structure = False pipeline_options.table_structure_options.do_cell_matching = False converter = DocumentConverter( format_options={ InputFormat.PDF: PdfFormatOption( pipeline_options=pipeline_options, ) } ) doc = converter.convert(input_doc).document md = doc.export_to_text() end_time = datetime.now() print(f"end_time : {end_time}") print(f"Total processing time: {end_time - start_time}")

[dolfim-ibm]

Please note that proper GPU support is not yet released, but it is coming very soon.

To understand your timings, let us propose another experiment.

Please add the following in your python code.

from docling.datamodel.settings import settings
settings.debug.profile_pipeline_timings=True

Then you will get internal profiling output in

result = converter.convert(input_doc)
doc = result.document
print(result.timings)

[PeterStaar-IBM]

@langzichai Is there any chance you could share the bigger file?

[langzichai]

Thanks for the reply, I will test it

[langzichai]

@langzichai Is there any chance you could share the bigger file?

Sorry, the file can't be shared at the moment

[langzichai]

Please note that proper GPU support is not yet released, but it is coming very soon.

To understand your timings, let us propose another experiment.

Please add the following in your python code.

from docling.datamodel.settings import settings
settings.debug.profile_pipeline_timings=True

Then you will get internal profiling output in

result = converter.convert(input_doc)
doc = result.document
print(result.timings)

@dolfim-ibm
Attached is the information output by timings
timings.txt

[cau-git]

@langzichai Several improvements, especially for GPU acceleration and layout processing, were released since you last reported, would you mind checking again with docling==2.14.0?

[langzichai]

Ok, I will test again, thanks for the help!