Speedups, bug fixes

• Fix some edge case OCR bugs
• ~20% end to end speedup by improving layout and text detection

Fix OCR bugs

• Fix bbox issue with OCR and resizing
• Fix issue with layout bboxes missing after OCR

Fix misc bugs

• Ensure we don't have 0 area table boxes
• Ensure fullymergedblock gets a valid input

Fix layout bugs

• Improve layout, which improves output quality
• Fix header level detection bugs

Fix OOM errors

• Add batch size for table rec model to avoid OOM
• Enable configuring batch size
• Fix error with debugging

Bugfixes, output quality improvement

• Fix MPS bug with torch 2.5
• Fix heading bug with zero line blocks
• Improve output quality when visual boxes and text boxes are offset

Better tables, improved output quality, header levels

Tables!

• Integrate custom table model for better table rendering - this uses a new state of the art open table model

Markdown output

• Adjust block detection to improve markdown output globally
• Assign layout labels to blocks in a better way - will improve quality globally
• Better line spacing in markdown output
• Push footnotes to end of page

Header levels

• Add detection for header levels like #, ##, etc.
• Add computed table of contents

Bugfixes/misc

• Fix bug with pagination not working
• Much better debugging with debug image output
• Python 3.13 support

OCR and misc improvements; demo app

• Language no longer needs to be specified
• Fix OCR memory leak
• Add marker GUI demo app to test out conversion
• Add progress for equation detection
• Improve table recognition slightly
• Add table benchmark

Significant speedup

This release has a 15% GPU speedup, 3x CPU, 7x MPS. The speedup comes from new surya models for layout and text detection that are a lot more efficient.

This is a "best case" speedup, if you need to OCR or do equation recognition, the speedup will be lower. But it will still be a lot faster.

Fix transformers bugs

• New transformers version introduces a new kwarg in donut models. Handle this case by ignoring it.
• New transformers version breaks MPS compatibility by using torch .isin to do a comparison. Handle this by setting the pytorch mps fallback setting.