Is it possible to improve the batch process?

[goriholdings]

Currently, the batch process starts vlm ocr after the layout check.
At this point, we have confirmed that PDFs with a large number of pages have a higher pages/second rate.
(Conversely, PDFs with a small number of pages have a lower pages/second rate.)

When starting a batch job, is it possible to perform layout checks on all files at once
and then start VLM OCR on all files at once?

I hope that I can achieve a higher pages/sec by collecting all PDFs and processing them at once. (I'm not sure if this is correct.)

(Results from rtx 4090.)
Results when processing a 56-page PDF

Layout time: 6.2, image num: 56
Doc analyze time: 68.03, speed: 0.82 pages/second

Results when processing a 343-page PDF

Layout time: 31.44, image count: 343
Doc analyze time: 202.18, speed: 1.7 pages/second

[zenosai]

As we have already supported batch processing for images, it is indeed possible to extend this to batch process PDFs as well. We will add official support for this feature as soon as possible — stay tuned!

[zenosai]

Hello, we have added support for mixed batch processing of PDFs and images. You can now use parse.py with the -g argument to specify the maximum number of pages per batch, enabling simultaneous parsing of multiple PDF files and images (see PR #159).

Feel free to try it out and share your feedback!

[goriholdings]

Thank you so much,
you are absolutely incredible.
Thank you so much again

[echo840]

Thank you very much for sharing your testing speeds on the 4090. This information is incredibly helpful for us and the open-source community.

[goriholdings]

Thank you very much for sharing your testing speeds on the 4090. This information is incredibly helpful for us and the open-source community.

I'm grateful to the developers for providing such good software. (For reference, I've observed that I average 4-5 page/sec for a 300-page document on the H200 GPU).

[echo840]

Thank you once again for the valuable information!
We've just released a smaller model that runs faster while outperforming our previous 3B model on Chinese documents. You can give it a try! MonkeyOCR-pro-1.2B