inference/README.md

AX DocLayout v1.2 CUDA inference source

ONNX Runtime CUDA + TF32 inference for AXDocLayout-M v1.2, using unchanged FP32 weights and 1024 letterbox geometry. Requires supported Linux x86-64/NVIDIA CUDA 12.8 setup and Python 3.12 or 3.13.

python -m pip install ".[onnx]"
python examples/predict_page.py page.png --backend onnx

The ONNX extra installs the pinned GPU runtime and CUDA/cuDNN dependencies. The backend requires CUDA and verifies TF32/compute placement. A GPU failure raises an error. Use the immutable source and weight revisions in the model repository root README. For an explicit CPU comparison, install .[torch] and select --backend pytorch. See LICENSE and NOTICE (AGPL-3.0-only).