2026-10-27: PP-DocLayoutV3_fp32_batchable.onnx, PP-OCRv6_medium_det.onnx, PP-OCRv6_small_det.onnx, PP-OCRv6_tiny_det.onnx, PP-OCRv6_tiny_rec_ctc_fp16.onnx, PP-OCRv6_medium_rec_ctc_fp16.onnx, PP-OCRv6_small_rec_ctc_fp16.onnx, PP-FormulaNet_plus-M_encoder.onnx, PP-FormulaNet_plus-M_prep.onnx, PP-FormulaNet_plus-M_decoder_step.onnx, PP-FormulaNet_plus-M_tokenizer.json, PP-FormulaNet_plus-S_encoder.onnx, PP-FormulaNet_plus-S_prep.onnx, PP-FormulaNet_plus-S_decoder_step.onnx, PP-FormulaNet_plus-S_tokenizer.json, SLANet_plus_encoder.onnx, SLANet_plus_decoder.bin
833ebc4 verified | license: apache-2.0 | |
| tags: | |
| - onnx | |
| - document-parsing | |
| - layout-analysis | |
| - ocr | |
| - table-structure-recognition | |
| - formula-recognition | |
| # docparser-models | |
| ONNX exports of PaddlePaddle document models, one flat file per model. | |
| `MANIFEST.txt` lists the sha256 of every file. All files are Apache-2.0. | |
| | File | Model | Source | | |
| | --- | --- | --- | | |
| | `PP-DocLayoutV3_fp32_batchable.onnx` | layout detection | [`PaddlePaddle/PP-DocLayoutV3_onnx`](https://huggingface.co/PaddlePaddle/PP-DocLayoutV3_onnx), output reshaped to `[B,300,7]` | | |
| | `PP-OCRv6_{tiny,small,medium}_det.onnx` | text detection | [`PaddlePaddle/PP-OCRv6_*_det_onnx`](https://huggingface.co/PaddlePaddle), unchanged | | |
| | `PP-OCRv6_{tiny,small,medium}_rec_ctc_fp16.onnx` | text recognition | [`PaddlePaddle/PP-OCRv6_*_rec_onnx`](https://huggingface.co/PaddlePaddle), plus `ctc_idx` / `ctc_prob` outputs (argmax and max over the class axis), FP16 weights with FP32 inputs and outputs | | |
| | `SLANet_plus_encoder.onnx`, `SLANet_plus_decoder.bin` | table structure | [`PaddlePaddle/SLANet_plus_onnx`](https://huggingface.co/PaddlePaddle/SLANet_plus_onnx): the encoder graph, and the GRU decoder's weights as raw float32 (run by a Python model, `table_decode`) | | |
| | `PP-FormulaNet_plus-{M,S}_{encoder,prep,decoder_step}.onnx`, `_tokenizer.json` | formula recognition | [`PaddlePaddle/PP-FormulaNet_plus-M`](https://huggingface.co/PaddlePaddle/PP-FormulaNet_plus-M) / [`-S`](https://huggingface.co/PaddlePaddle/PP-FormulaNet_plus-S), split into encoder, cross-attention precompute and one greedy decoder step (FP16, new-slice) whose KV axis is dynamic: one TensorRT engine per tier with a profile per KV length (plus-S 256 / 512 / 1024, plus-M 256 / 512 / 1536) | | |
| Pre- and post-processing follow each source model's `inference.yml`. | |