New Layout Model

Releasing new layout model

• Moving to a new architecture for layout, trained from scratch. Significant improvements across many domains

What's Changed

• feat: new unified tokenizer by @VikParuchuri in #442
• Layout Model Release by @tarun-menta in #461
• Dev by @tarun-menta in #463

Full Changelog: v0.16.7...v0.17.0

Move flash imports

What's Changed

• Move flash attention funcs by @VikParuchuri in #457

Full Changelog: v0.16.6...v0.16.7

Enable setting attention method

What's Changed

• Get rid of attention method checks by @VikParuchuri in #455
• Enable setting attention method by @VikParuchuri in #456

Full Changelog: v0.16.5...v0.16.6

Minor init update

What's Changed

• Foundation predictor init by @VikParuchuri in #454

Full Changelog: v0.16.4...v0.16.5

Speed improvements; SDPA fix

• Fix attention mask for SDPA
• Improve performance 20-30%, more so with marker
• Add some more model loader options

What's Changed

• Backport by @VikParuchuri in #449
• Foundation Model Performance Improvements by @tarun-menta in #451
• Dev by @VikParuchuri in #452
• Bump version by @VikParuchuri in #453

Full Changelog: v0.16.3...v0.16.4

v0.16.3

Revert checkpoint

Full Changelog: v0.16.2...v0.16.3

Multi-token inference + Better Math/Tables

What's Changed

• Update README by @u-ashish in #447
• feat: multi-token decoding by @zanussbaum in #446
• Dev by @zanussbaum in #448

New Contributors

• @u-ashish made their first contribution in #447

Full Changelog: v0.16.1...v0.16.2

Transformers fix

Hotfix to be compatible with transformers 4.56.0.

What's Changed

• Remove unused import by @VikParuchuri in #444

Full Changelog: v0.16.0...v0.16.1

New OCR Model

General OCR Improvements

• Update to a better OCR model. Uses less vocab size, and a more performant vision encoder.
• Improved math performance, with less spurious use of math tags when not needed

Misc

• Fix table model by pinning to CPU on MPS devices

What's Changed

• update commercial license description in README and update LICENSE to… by @sandy0kwon in #434
• Dev by @VikParuchuri in #436
• Model Update: New Tokenizer and Encoder by @tarun-menta in #440
• Dev by @tarun-menta in #441

New Contributors

• @sandy0kwon made their first contribution in #434

Full Changelog: v0.15.4...v0.16.0

Improved Math OCR Model

Math OCR improvements

• Bump the OCR model with a version that improves general performance, and significantly improves math performance

What's Changed

• Improve model performance on math by @tarun-menta in #429

Full Changelog: v0.15.3...v0.15.4