Docling MLX

Project-authored code is Apache-2.0 licensed. Individual source files adapted from third-party
code retain the notices and license stated below. Engines written after a reference
implementation, rather than copied from it, credit that reference below. Adaptation and reference
versions and commits are recorded in the component validation documents under docs/.

USGS-derived test fixtures:
  `tests/fixtures/tableformer_v2/basin_table_1.png`, `basin_table_2.png`, and
  `basin_table_3.png`, plus the deterministic wrappers
  `tests/fixtures/tableformer_v2/basin_table_1.pdf` and
  `tests/fixtures/tableformer_v1/basin_table_1_text.pdf`, derive from U.S.
  Geological Survey Open-File Report 1958-87, a U.S. Geological Survey publication in the public
  domain in the United States; credit to the U.S. Geological Survey is retained per USGS guidance.
  The report material is tracked as LicenseRef-USGS-Public-Domain (see
  LICENSES/LicenseRef-USGS-Public-Domain.txt). The deterministic PDF wrappers
  also contain Docling MLX contributors' Apache-2.0 material; see
  LICENSES/Apache-2.0.txt. See the provenance records at
  `tests/fixtures/tableformer_v1/UPSTREAM.md` and
  `tests/fixtures/tableformer_v2/UPSTREAM.md`.
  https://pubs.usgs.gov/of/1958/0087/

Native RT-DETR-v2 architecture and grid sampling:
  Adapted from mlx-vlm (mlx_vlm/models/rt_detr_v2).
  Copyright © 2025 Prince Canuma.
  Licensed under the MIT License; see LICENSES/MIT.txt.
  https://github.com/Blaizzy/mlx-vlm

Docling stage and parser behavior:
  The Granite table OTSL parser and chart CSV/code parsers are locally adapted
  from Docling. Copyright The Docling Contributors. Licensed under MIT;
  see LICENSES/MIT.txt.
  https://github.com/docling-project/docling

Docling test fixture:
  `tests/fixtures/document_figure/picture_classification.pdf` was copied from
  Docling; the two reference PNGs were derived from its image XObjects. Copyright The
  Docling Contributors. Licensed under MIT; see LICENSES/MIT.txt and
  tests/fixtures/document_figure/UPSTREAM.md.
  https://github.com/docling-project/docling

Reference implementations:

Native EfficientNet architecture:
  The MLX modules were written after the Hugging Face Transformers efficientnet
  reference implementation (Copyright 2023 Google Research, Inc. and The
  HuggingFace Inc. team, Apache License 2.0). They follow the reference module
  structure, parameter names, and forward-pass order.
  https://github.com/huggingface/transformers

Native D-FINE architecture:
  The MLX modules were written after the Hugging Face Transformers d_fine and
  hgnet_v2 reference implementations (Copyright 2018- The Hugging Face team,
  Apache License 2.0). They follow the reference module structure, parameter
  names, and forward-pass order.
  https://github.com/huggingface/transformers

Native TableFormer v1 architecture:
  The Accurate and Fast presets are covered by this entry. The MLX modules were
  written after the docling-ibm-models tableformer reference implementation
  (Copyright (c) 2024 International Business Machines, MIT) and, for the
  ResNet-18 feature stack, after Torchvision (Copyright (c) Soumith Chintala
  2016, BSD 3-Clause). They follow the reference module structure, parameter
  names, and forward-pass order.
  https://github.com/docling-project/docling-ibm-models
  https://github.com/pytorch/vision

Native TableFormerV2 architecture:
  TableFormerV2 has no published paper. The MLX modules were written after the
  docling-ibm-models tableformer_v2 reference implementation (Copyright (c) 2024
  International Business Machines, MIT) and, for the EfficientNetV2-S encoder,
  after Torchvision (Copyright (c) Soumith Chintala 2016, BSD 3-Clause). They
  follow the reference module structure, parameter names, and forward-pass
  order.
  https://github.com/docling-project/docling-ibm-models
  https://github.com/pytorch/vision
