This repository is a fork/local adaptation of Baidu's Unlimited-OCR model:

Read the original model card and upstream repository first for model details, license, intended usage, limitations, citation, and full inference guidance.

Run a basic local load/inference smoke test:

Export the local model checkout to ONNX:

The default ONNX target exports an image prefill forward graph. It does not export the Python generate() loop or PIL preprocessing from model.infer(). ONNX export uses float16 on CUDA and float32 on CPU when --dtype auto is used. bfloat16 exports are not compatible with ONNX Runtime for this graph's convolution nodes. The image ONNX graph is fixed-length because the model's MoE routing is data-dependent. Increase --image-sequence-length if prompt tokens plus generated tokens exceed the default capacity. The exporter traces with padded attention-mask slots; re-export old ONNX files if inference returns no visible text. ONNX export uses a dense tensor-only MoE route for correctness, so it can be slower than the PyTorch expert-dispatch path.

Text-only export:

Run greedy ONNX Runtime inference:

For a CUDA/float16 ONNX graph, run inference with onnxruntime-gpu so CUDAExecutionProvider is available.