docling-layout-heron-coreml

docling-layout-heron for Core ML

Core ML conversion of docling-project/docling-layout-heron, Docling's RT-DETRv2 document-layout model. The weights were converted, not retrained or fine-tuned.

This repository contains two equivalent artifacts:

• layout_heron.mlpackage — portable Core ML ML Program (recommended for redistribution and inspection).
• layout_heron.mlmodelc — precompiled copy for direct loading by docling.rs and other native Core ML clients. It was compiled with Xcode 26.6.

The model requires macOS 15+, iOS/iPadOS 18+, macCatalyst 18+, tvOS 18+, visionOS 2+, or watchOS 11+. It is an ML Program with FP16 weight and compute precision, a fixed batch size of one, and raw detector outputs.

Input and output contract

Input:

• pixel_values: Float32 [1, 3, 640, 640], NCHW RGB.
• Stretch the page image to exactly 640 × 640 using bilinear resampling.
• Rescale bytes to [0, 1] by multiplying by 1/255.
• Do not letterbox, pad, or apply mean/std normalization.

Outputs:

• logits: Float16 [1, 300, 17] raw class logits.
• pred_boxes: Float16 [1, 300, 4] normalized (center_x, center_y, width, height) boxes.

Those are the model-schema dtypes. coremltools may expose both arrays as Float32 NumPy values after prediction.

The output is intentionally raw. Apply sigmoid to logits, choose the highest query/class scores, threshold them, convert boxes from normalized CXCYWH to XYXY, and scale them back to the original page dimensions. The 17 labels and all machine-readable shape metadata are in coreml_config.json.

Use with docling.rs

Place either artifact at the default path:

mkdir -p .models
cp -R layout_heron.mlmodelc .models/layout_heron.mlmodelc

Then run a normal macOS build of docling.rs. Its default coreml-native feature automatically prefers the native Core ML layout backend when the model is present. You can also set an explicit path:

export DOCLING_RS_LAYOUT_ENGINE=coreml
export DOCLING_LAYOUT_COREML="$PWD/layout_heron.mlpackage"
cargo run -p docling-cli -- document.pdf

DOCLING_RS_LAYOUT_ENGINE=coreml means the native Core ML backend. It is different from the ONNX Runtime CoreML execution-provider feature.

Use from Python

Core ML execution requires macOS:

import coremltools as ct
import numpy as np
from PIL import Image

model = ct.models.MLModel("layout_heron.mlpackage", compute_units=ct.ComputeUnit.ALL)

image = Image.open("page.png").convert("RGB").resize((640, 640), Image.Resampling.BILINEAR)
pixel_values = np.asarray(image, dtype=np.float32).transpose(2, 0, 1)[None] / 255.0

output = model.predict({"pixel_values": pixel_values})
logits = output["logits"]
pred_boxes = output["pred_boxes"]

Validation

The conversion was checked against the FP32 ONNX export on 12 inputs: nine real document pages and three uniform-random tensors. At a 0.3 score threshold, greedy same-class matching with IoU > 0.5 produced:

On the conversion machine (Apple Silicon, macOS 26.5), median batch-1 latency over 20 warm runs was 12.31 ms with ComputeUnit.ALL, compared with 237.06 ms for ONNX FP32 CPU and 161.73 ms for ONNX INT8 CPU. These are conversion checks, not broad benchmark claims; performance varies by device and OS.

The Core ML compute plan placed all 2,427 non-constant compute operations on the GPU with ComputeUnit.ALL. CPU_AND_NE was slower for this graph.

Reproduce the conversion

The conversion requires macOS and Python 3.12:

python3.12 -m venv .venv
.venv/bin/pip install -r conversion/requirements.txt
.venv/bin/python conversion/convert.py
xcrun coremlcompiler compile layout_heron.mlpackage .

conversion/convert.py contains the three compatibility rewrites needed by coremltools 8.3 for RT-DETRv2: bilinear grid_sample decomposition, scalar casting for a one-element traced tensor, and an out-of-place anchor-generation rewrite. The rewrites are numerically equivalent for this fixed inference contract.

Limitations

• Fixed batch size 1 and fixed 640 × 640 input.
• Raw detector outputs; preprocessing and postprocessing are not embedded.
• The precompiled .mlmodelc was built by Xcode 26.6. If it does not load on a target system, use the .mlpackage so Core ML can compile it locally.
• The validation set is small and is intended to catch conversion drift, not to re-evaluate the upstream model's accuracy or fairness.
• This conversion inherits the upstream model's intended uses, limitations, training data, and risks. See the upstream model card and technical report.

License and attribution

The upstream model is Apache-2.0 licensed. This repository is a format conversion and retains that license and attribution. See LICENSE and NOTICE.

If you use the model, please cite the upstream Heron and Docling work listed in the upstream model card.