Release 74.84M Nepali Hindi English research OCR with source and native Paddle weights

.gitattributes

@@ -1,35 +1,5 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.pdiparams filter=lfs diff=lfs merge=lfs -text
+*.so filter=lfs diff=lfs merge=lfs -text
+paddle/candidate/english-assets/sentencepiece.bpe.model filter=lfs diff=lfs merge=lfs -text
+source-model/english-assets/sentencepiece.bpe.model filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,17 @@
+Ampixa Non-Commercial Research Release Terms
+
+The Ampixa-authored portions of this research checkpoint and its accompanying
+release utilities are made available for non-commercial research and education
+only, to the extent of rights held by Ampixa. Commercial use, commercial hosting,
+and incorporation in commercial products are not licensed by this release.
+
+Retain this notice, the model card, training-data provenance and all applicable
+third-party copyright and license notices when redistributing permitted portions.
+Do not imply an unrestricted license for the checkpoint or its training data.
+This release grants no rights to redistribute the underlying training datasets.
+Third-party components remain subject to their original terms; this notice does
+not replace or broaden those rights. See THIRD_PARTY_NOTICES.md and TRAINING_DATA.md.
+
+THE RELEASE IS PROVIDED AS IS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND.
+The authors make no guarantee of accuracy, fitness for purpose, unseen-writer
+performance, or clearance of third-party rights for other uses.

README.md

@@ -0,0 +1,102 @@
+---
+language:
+- ne
+- hi
+- en
+license: other
+license_name: ampixa-non-commercial-research
+license_link: LICENSE
+pipeline_tag: image-to-text
+base_model: microsoft/trocr-small-handwritten
+tags:
+- ocr
+- handwriting-recognition
+- devanagari
+- nepali
+- hindi
+- pytorch
+- paddleocr
+- research
+---
+
+# Ampixa Deva OCR 75M — research checkpoint
+
+A **74,842,752-parameter** recognizer for cropped Nepali/Hindi words or lines and English handwriting. This release includes source Safetensors weights and a native Paddle/PaddleX export of the same checkpoint. Each format contains one shared encoder; the two formats are alternatives, not models to load together.
+
+This is the established character-decoder checkpoint after an 8,000-update continuation (16,000 cumulative updates), rather than the newer experimental BPE model. It is a research release, not a claim of state-of-the-art OCR or achievement of 4% Nepali handwriting CER.
+
+## Measured results
+
+| Fixed evaluation set | Crops | CER |
+|---|---:|---:|
+| Nepali handwriting words | 300 | 12.9921% |
+| Hindi handwriting words | 2,000 | 8.0803% |
+| English handwriting lines, original English branch | 976 | 2.3150% |
+| English handwriting lines, character branch | 976 | 6.0940% |
+| Real print crops | 2,000 | 11.1904% |
+| Synthetic print crops | 800 | 2.6169% |
+| English print crops, character branch | 1,744 | 13.9357% |
+
+CER is micro-averaged NFC Unicode codepoint edit distance. Every evaluated crop is retained, including errors, empty outputs and generation caps. See `metrics.json` for exact totals and generation settings. All 8,796 native outputs from the two explicit branches matched source tokens, text, caps and errors on these fixed populations. Native evaluation performed zero Torch tensor operations.
+
+These are existing development populations, not a blind benchmark. Nepali is word-level; it does not establish performance on unseen writers, full handwritten notes or pages. An earlier checkpoint of this architecture reached 12.53% Nepali word CER; it is not the checkpoint in this release. On a separate 336-paragraph English comparison, the retained English recognizer achieved 8.16% CER versus 3.01% for Qwen3-VL-4B. This release does not beat that stronger paragraph reference.
+
+## Download and source inference
+
+```bash
+hf download ampixa/deva-ocr-75m-research --local-dir deva-ocr-75m
+cd deva-ocr-75m
+pip install -r requirements-source.txt
+python predict.py /path/to/cropped-line.png --language ne
+python predict.py /path/to/english-line.png --language en
+```
+
+The included CLI uses serial CPU inference and does not download another model. Choose `ne`, `hi` or `en` explicitly. Nepali and Hindi use the character decoder with the adapted encoder; English uses the unchanged original English model. All branches resize each crop to 384×384 RGB. Character decoding uses 10 beams and up to 257 new tokens; English uses 10 beams and up to 256. JSON output includes text, raw token IDs and whether generation reached its cap.
+
+This is a custom architecture. Use the included loader/CLI; `AutoModel.from_pretrained` and the generic Transformers OCR pipeline are not configured for this repository.
+
+## Native Paddle
+
+Install the native dependencies in a separate environment:
+
+```bash
+pip install -r requirements-paddle.txt
+python predict.py /path/to/cropped-line.png --language ne --backend paddle
+```
+
+The prebuilt operator library was validated on Linux x86_64 with FMA, PaddlePaddle 3.3.1 and PaddleX 3.7.1, using two CPU threads. Native GPU execution and other platforms are unvalidated. Use the source backend on platforms incompatible with the prebuilt library. The custom operator source and its third-party notices are under `paddle/code/`.
+
+For the actual PaddleX factory:
+
+```python
+from pathlib import Path
+import sys
+from paddlex.inference.models import create_predictor
+
+root = Path("deva-ocr-75m").resolve()
+sys.path.insert(0, str(root / "paddle/code"))
+from paddlex_shared_character import register, MODEL_NAME
+register()
+predictor = create_predictor(
+    MODEL_NAME, model_dir=str(root / "paddle"), device="cpu",
+    engine="flexible", research_only_allow_unqualified=True,
+)
+for result in predictor.predict("cropped-line.png", branch="character"):
+    print(result["rec_text"])
+```
+
+Use `branch="english"` for English. The research opt-in remains intentional: fixed-crop numerical parity does not establish full product acceptance. Scores are uncalibrated beam scores. Detection, reading order, word boxes, full-page segmentation and automatic language routing are not included.
+
+## Architecture and provenance
+
+The original Microsoft TrOCR-small English model supplies the shared DeiT encoder and English decoder. The character branch applies stored rank-71 encoder corrections and a separate literal-Unicode character decoder without storing a second encoder. Source weights use Safetensors, with checksummed configuration and tokenizer assets.
+
+`source-model/` holds source weights and their original manifest. `code/` contains the source loader and codec. `paddle/` contains the native encoder, two decoder graphs, runtime and operator library. `release-manifest.json` binds this release to its checkpoint and prior native package; `checksums.sha256` inventories published files. Packaging excludes training images and conversion-control prediction packets. Existing graph parameter files and runtime source bytes are retained.
+
+## Use and data terms
+
+Use is restricted to non-commercial research and education. Training includes research-restricted IAM and IIIT handwriting sources, CC0 Nepali word crops, CC BY material, synthetic renders and internally used newspaper-derived crops. This repository does not grant rights to the underlying datasets or permission for commercial deployment. See `LICENSE`, `TRAINING_DATA.md` and `THIRD_PARTY_NOTICES.md`. No Indra or Qwen weights are included; those systems were comparators, not teachers for this checkpoint.
+
+## Release packaging checks
+
+The download-layout CLI was checked on fixed Nepali, Hindi and English inputs in both backends (six exact reference matches). The actual PaddleX factory was rerun on 24 path inputs and 24 BGR inputs, with 24 JSON results checked and zero Torch tensor operations. These are packaging controls; the larger fixed-corpus accuracy results are reported above.

THIRD_PARTY_NOTICES.md

@@ -0,0 +1,8 @@
+# Third-party notices
+
+- Microsoft TrOCR-small-handwritten: the unchanged original English weights and architecture originate from [microsoft/trocr-small-handwritten](https://huggingface.co/microsoft/trocr-small-handwritten), whose publisher model card declares MIT. Microsoft TrOCR source is in [microsoft/unilm](https://github.com/microsoft/unilm/tree/master/trocr); its original model card is retained at `source-model/english-assets/README.md`.
+- Hugging Face Transformers, PyTorch, PaddlePaddle, PaddleX, Safetensors, SentencePiece, tokenizers, Pillow and NumPy are external runtime dependencies, governed by their respective licenses.
+- The native CPU operator code incorporates or adapts third-party numerical implementations. Their exact notices are retained under `paddle/code/`: `trocr_paddle_norm_probe.LICENSE`, `trocr_paddle_sleef_probe.LICENSE`, and `trocr_selective_onednn_pass.LICENSE`. Retain those files with native redistributions.
+- The project-authored loader, codec and beam-search runtime retain their source comments and are supplied under the research release terms in `LICENSE`. This does not relicense third-party code or model components.
+
+The source and native formats encode the same 74,842,752-value recognizer. Neither training datasets nor third-party comparator weights are included.

TRAINING_DATA.md

@@ -0,0 +1,12 @@
+# Training provenance and scope
+
+This checkpoint descends from earlier OCR experiments; its final continuation used a fixed mixture of Nepali words, IIIT Hindi words, English handwriting lines, GNHK lines, newspaper-derived print, NVIDIA English print, IIIT Hindi Unicode words and conjunct renders. Earlier ancestors also used additional word/character and synthetic replay. This is a research training lineage, not a clean-room unrestricted checkpoint.
+
+- **Nepali handwriting words:** [Sweekar Dahal's collection](https://github.com/dahalsweekar/Nepali-Handwritten-Dataset-for-Recognition); CC0 uploader metadata. The final mixture uses 3,187 training words. Evaluation crops remain excluded from training.
+- **Hindi handwriting:** [IIIT-INDIC-HW-WORDS](https://cvit.iiit.ac.in/research/projects/cvit-projects/iiit-indic-hw-words), IIIT-HW-Dev and [IIIT-Indic-HW-UC](https://cvit.iiit.ac.in/usodi/ucciohd.php). These were used in their research/education context. No unrestricted commercial grant is asserted.
+- **English handwriting:** [IAM](https://fki.tic.heia-fr.ch/databases/iam-handwriting-database), used under its non-commercial research context, and [GNHK](https://www.gnhk-dataset.org/), with upstream attribution and dataset terms. Public mirrors do not broaden the original terms.
+- **English synthetic print replay:** [NVIDIA OCR-Synthetic-Multilingual-v1](https://huggingface.co/datasets/nvidia/OCR-Synthetic-Multilingual-v1), attributed CC BY 4.0 material in the experiment's original source inventory. The model card does not grant ownership of these source images.
+- **Nepali print:** Gorkhapatra-derived chunks were used internally with converted PDF-text silver labels. They are not newly human-verified labels or an unrestricted public dataset.
+- **Synthetic text/conjunct renders:** Unicode text rendered with OFL-font families; font licenses remain separate. No font binaries are redistributed here.
+
+Training images, image crops, label files, raw evaluation predictions and training secrets are excluded from this repository. Training on research data and publication of these research weights do not establish permission for commercial use. Obtain appropriate rights from the relevant owners before considering a separate commercial release.

checksums.sha256

@@ -0,0 +1,57 @@
+6914e9fe1b84a0abfe32e9a1148a9b29db07a78dc646b82416b50ee0f5f20acb  .gitattributes
+37908899ed8fb456bd644323faae79bdd5b54ebdf90f07c7cf795d5ccfdd2866  LICENSE
+fc0aee088bfeb7f7449f98426a6a8185e1b81e767da6da45b13016bf687fdbd6  README.md
+3beebf852bf869560dd1adabeb93be212c61a55ce002fbe2bdea2b5f8167cb68  THIRD_PARTY_NOTICES.md
+d986e14af668016c081480cb05e316987bc8c519764f9f004c33db66e855a5d5  TRAINING_DATA.md
+0498d8b7d61f3c8e4a4fdcd927a507464cfaf8156bf874a211484a5474697719  code/shared_character_recognizer.py
+7ede21e57fd4985153570820873abf690cc48de2e677b38828753fbedfdbd5e2  code/trocr_character_codec.py
+965c8de2c0cf3d005ac70cb01a205f5e599fea85d888fa88f0c0164599594db7  metrics.json
+364b288087110ed4a47b6e4c0cfa1367dfcff6249c57d082b3becd472e338e58  paddle/bundle.json
+0510d21cd0bb7a249dafac167fd2e371ee63bd6f83a016a6c4f2e7c6bb843ede  paddle/candidate/bundle.json
+7bb65c86bc29b555f8da3af81b7023e370bdebe6ced8c84d104d55165c6ffb0a  paddle/candidate/character-assets/character_codec.json
+36a945a7cc645688b9ef64dabae16979cf5f7c1c448569cc306694edc0598b9b  paddle/candidate/character-assets/preprocessor_config.json
+7e0f183a2697956065aea4e3ec9ca718967be1b7afe446cff4dcdbc8a64fa52b  paddle/candidate/english-assets/README.md
+eec0c686d2e2ea6d5887f87e1c6a34ef5bf1ba5b2ac98b8675e131da646fa1d2  paddle/candidate/english-assets/preprocessor_config.json
+6f5e2fefcf793761a76a6bfb8ad35489f9c203b25557673284b6d032f41043f4  paddle/candidate/english-assets/sentencepiece.bpe.model
+e098c98c6834fbcd3f3e75215e3e8b8868f0fe5bf07aef93d28e4f7196db293c  paddle/candidate/english-assets/tokenizer_config.json
+6a7b1d60aa337ba04514c9910e68f96d1bafdff383f3779e83f22a587707a3e3  paddle/code/paddlex_shared_character.py
+394b45df79c5c4e84c2053b62b8d4080ef6a875034b6b16c661ac63257f56a30  paddle/code/shared_character_native.py
+7ede21e57fd4985153570820873abf690cc48de2e677b38828753fbedfdbd5e2  paddle/code/trocr_character_codec.py
+e4f1d87113404dfc16fdddd78a1cdd6477a2f886d76cef99594fc537fc531028  paddle/code/trocr_numpy_beam.py
+19d6a3818fa4927890b59222656f155faae8f8817ebc9f40f779fdb6ab5a65d2  paddle/code/trocr_paddle_norm_probe.LICENSE
+5f35cc826b3a0001293ed62e8ea95e5784f89f5e25300bf437d34be1d35e8a75  paddle/code/trocr_paddle_sleef_probe.LICENSE
+ff6311c56c1b22c9209c96cc8afc16da91c24ee55a15e14ebc46fe1960dbd856  paddle/code/trocr_selective_onednn_pass.LICENSE
+ff304ef7e20f22cc1bf3c3199c6614bc3e870db640e569d465ca87c6704051b6  paddle/code/trocr_selective_onednn_pass.cc
+63bb541df8f7a1f8635c946f2407aa0f449e63cfe480cbbe6c0b8d6fb2f5f4c2  paddle/code/trocr_source_cpu_ops.cc
+555d79e651df3b22fef37f7cc873c843678c135bbe39e69488f4ef3b2acb3e5d  paddle/decoders/character/decoder-step.json
+e35ba8d8743abdcf8b6cdb48a572a669cbf3e504d38d2a48827b68657e32aeff  paddle/decoders/character/decoder-step.pdiparams
+eaea1b3b5001549a6d0a4c4b1bd810371a82bdae2092452ac2f73214762ede95  paddle/decoders/english/decoder-step.json
+ec5ec71ad13c5305a1a95c8c2846c4de526571e58ce8df00fa5dbac529940a16  paddle/decoders/english/decoder-step.pdiparams
+24a40b59ddedd55f423035cb40def1e251633948a67802707d14cf3151a67ba8  paddle/decoders/report.json
+eddaf1699760460ba7f2bd7fbdee9b7de9166b581e5145d9dc3c8da177bd72b7  paddle/encoder/encoder.json
+b2a632ecd78bd35441f0575b5073ebdfe2dd8edebff08c892b386205655700b0  paddle/encoder/encoder.pdiparams
+6e19e7fe532e7d24d504f9e323e08d5c77e1ca07ceaeb1ebcc0f616369834dbc  paddle/encoder/report.json
+42c4cf89a86fc3de900949348079f7b0c90d4e58ca2d28a105e326785a3eacd5  paddle/encoder/run.json
+fce69529c6a3752247be1192a495de7d1c395f8493039ad0d17546a42f936c60  paddle/encoder-library.so
+4748d485e0280227d6b395faf8c0b956217389abb4a3c267ff6cf5d427116863  paddle/inference.yml
+ec18fa5c8940fee3624d1c05e0ea4be12f1f3105a184eb03f18e7b1991a8ee7c  predict.py
+0f20efb6a6b119ea4de9752b6120a19df0274c1c52621e6349e466b64b7a714b  release-manifest.json
+7b26b16767a3a722eaed9b6499b3872927773a0817298bf8013a590c89e0e0bc  requirements-paddle.txt
+2d2115dd37ef7dec5f883532694a7508120fe1fe20933d66bd843ff961460edf  requirements-source.txt
+0510d21cd0bb7a249dafac167fd2e371ee63bd6f83a016a6c4f2e7c6bb843ede  source-model/bundle.json
+7bb65c86bc29b555f8da3af81b7023e370bdebe6ced8c84d104d55165c6ffb0a  source-model/character-assets/character_codec.json
+decbe7681cc6000bd5bf83226b295e7a051e8280bc12255961e011d2caadbfce  source-model/character-assets/config.json
+9577d3da76b60ee97a70d1516ac144eba329a8a70dd0e96c6f721c910434dcff  source-model/character-assets/generation_config.json
+36a945a7cc645688b9ef64dabae16979cf5f7c1c448569cc306694edc0598b9b  source-model/character-assets/preprocessor_config.json
+02b8abbca9b3f1e5fb8ea2057b0aea84e1f488dd5dd2ca52a75850170bc3b08b  source-model/character-generation.json
+7e0f183a2697956065aea4e3ec9ca718967be1b7afe446cff4dcdbc8a64fa52b  source-model/english-assets/README.md
+2b6e06edec319984eca99c932d940bfa1f308db3f74d6279440b906c373bf727  source-model/english-assets/config.json
+41149cdcffec4d657f32dfcddd9b208037f01286c9e07945c724908c58ed0193  source-model/english-assets/generation_config.json
+eec0c686d2e2ea6d5887f87e1c6a34ef5bf1ba5b2ac98b8675e131da646fa1d2  source-model/english-assets/preprocessor_config.json
+6f5e2fefcf793761a76a6bfb8ad35489f9c203b25557673284b6d032f41043f4  source-model/english-assets/sentencepiece.bpe.model
+58dc0a3d15e3bad4861bcf2dee770ea327344d185182d4c5d2bf253285e53a83  source-model/english-assets/special_tokens_map.json
+e098c98c6834fbcd3f3e75215e3e8b8868f0fe5bf07aef93d28e4f7196db293c  source-model/english-assets/tokenizer_config.json
+cddf3683b66672c43fee2a6d966829eebf2a00c61d376d0eb30db972693210be  source-model/english-generation.json
+1c598da84fcbe26af738d5ef1c0ea04f80f965e7690b1089aeaa8c7977cd4359  source-model/model-configs.json
+9b58d252978d3a5d33adcdaea95e113162daf5146e20a9e1b8aa47e21e48ed93  source-model/weights.safetensors
+bf281f9f2fbabbec2885fc0556e59cc528bd2eed5f7033c20dd718bb480d9249  validation.json

code/shared_character_recognizer.py

@@ -0,0 +1,115 @@
+"""Research recognizer sharing one English encoder across two explicit branches.
+
+The character branch applies saved low-rank corrections through functional_call
+and supplies precomputed encoder outputs to a decoder-only VED container. No
+second learned encoder is registered or saved. Functional parameter overrides
+are serial CPU research operations, not a concurrent/native serving interface.
+"""
+import copy
+import hashlib
+import json
+from pathlib import Path
+
+import torch
+from torch import nn
+from safetensors.torch import load_file
+from transformers import DeiTModel, GenerationConfig, VisionEncoderDecoderConfig, VisionEncoderDecoderModel
+
+
+def sha(path):
+    value = hashlib.sha256()
+    with Path(path).open('rb') as stream:
+        for block in iter(lambda: stream.read(1024 * 1024), b''): value.update(block)
+    return value.hexdigest()
+
+
+def eager(config):
+    config._attn_implementation = 'eager'
+    config.encoder._attn_implementation = 'eager'
+    config.decoder._attn_implementation = 'eager'
+    return config
+
+
+def load_english(folder):
+    config = eager(VisionEncoderDecoderConfig.from_pretrained(folder, local_files_only=True))
+    model, report = VisionEncoderDecoderModel.from_pretrained(folder, config=config,
+        encoder=DeiTModel(config.encoder, add_pooling_layer=False), local_files_only=True,
+        dtype=torch.float32, attn_implementation='eager', output_loading_info=True)
+    assert not any(report.get(k) for k in ['missing_keys', 'unexpected_keys', 'mismatched_keys', 'error_msgs'])
+    assert sum(p.numel() for p in model.parameters()) == 61448832
+    return model.eval().requires_grad_(False)
+
+
+class PrecomputedOnlyEncoder(nn.Module):
+    main_input_name = 'pixel_values'
+
+    def __init__(self, config):
+        super().__init__(); self.config = copy.deepcopy(config)
+
+    def get_output_embeddings(self):
+        return None
+
+    def forward(self, pixel_values=None, **kwargs):
+        raise RuntimeError('Character decoding requires the shared encoder outputs')
+
+
+class Correction(nn.Module):
+    def __init__(self, spec):
+        super().__init__(); self.spec = spec
+        if spec['kind'] == 'factor':
+            self.left = nn.Parameter(torch.empty(spec['matrix_shape'][0], spec['rank']))
+            self.right = nn.Parameter(torch.empty(spec['rank'], spec['matrix_shape'][1]))
+        else:
+            assert spec['kind'] == 'replacement'
+            self.value = nn.Parameter(torch.empty(spec['shape']))
+
+    def apply(self, base):
+        if self.spec['kind'] == 'replacement': return self.value
+        return base + (self.left @ self.right).reshape(self.spec['shape'])
+
+
+class SharedCharacterRecognizer(nn.Module):
+    def __init__(self, original, character_decoder, character_projection, character_config,
+                 english_generation, character_generation, specifications):
+        super().__init__()
+        self.original = original
+        cfg = eager(copy.deepcopy(character_config))
+        self.character = VisionEncoderDecoderModel(config=cfg,
+            encoder=PrecomputedOnlyEncoder(cfg.encoder), decoder=character_decoder)
+        if character_projection is not None: self.character.enc_to_dec_proj = character_projection
+        self.original.generation_config = copy.deepcopy(english_generation)
+        self.character.generation_config = copy.deepcopy(character_generation)
+        self.specifications = specifications
+        self.corrections = nn.ModuleDict({spec['key']: Correction(spec) for spec in specifications})
+
+    def generate_pixels(self, pixels, branch, generation):
+        if self.training: raise ValueError('Explicit eval mode is required')
+        if branch == 'english': return self.original.generate(pixel_values=pixels, **generation)
+        if branch != 'character': raise ValueError('Require an explicit English or character branch')
+        base = self.original.encoder.state_dict()
+        overrides = {spec['name']: self.corrections[spec['key']].apply(base[spec['name']])
+                     for spec in self.specifications}
+        output = torch.func.functional_call(self.original.encoder, overrides, (),
+            dict(pixel_values=pixels, return_dict=True), strict=False)
+        return self.character.generate(encoder_outputs=output, **generation)
+
+    @classmethod
+    def from_bundle(cls, folder):
+        folder = Path(folder); manifest = json.loads((folder / 'bundle.json').read_text())
+        for name, expected in manifest['files_sha256'].items(): assert sha(folder / name) == expected
+        configs = json.loads((folder / 'model-configs.json').read_text())
+        english = eager(VisionEncoderDecoderConfig.from_dict(configs['english']))
+        character = eager(VisionEncoderDecoderConfig.from_dict(configs['character']))
+        original = VisionEncoderDecoderModel(config=english,
+            encoder=DeiTModel(english.encoder, add_pooling_layer=False))
+        private = VisionEncoderDecoderModel(config=character,
+            encoder=PrecomputedOnlyEncoder(character.encoder))
+        model = cls(original, private.decoder, getattr(private, 'enc_to_dec_proj', None), character,
+            GenerationConfig.from_pretrained(folder, config_file_name='english-generation.json', local_files_only=True),
+            GenerationConfig.from_pretrained(folder, config_file_name='character-generation.json', local_files_only=True),
+            manifest['corrections'])
+        result = model.load_state_dict(load_file(str(folder / 'weights.safetensors')), strict=True)
+        assert not result.missing_keys and not result.unexpected_keys
+        assert sum(p.numel() for p in model.parameters()) == manifest['parameters'] == 74842752
+        assert len(model.character.encoder.state_dict()) == 0
+        return model.eval().requires_grad_(False)

code/trocr_character_codec.py

@@ -0,0 +1,46 @@
+"""Literal NFC character codec for the experimental compact TrOCR decoder.
+
+Tokenizer BOS0 is serialization only; training targets exclude it. Generation
+starts with EOS2 and suppresses BOS/PAD/UNK. Unknown outputs remain visible as
+replacement characters when decoded, never silently removed.
+"""
+import unicodedata
+
+
+class CharacterCodec:
+    def __init__(self, characters):
+        if len(characters) != len(set(characters)) or any(len(c) != 1 for c in characters):
+            raise ValueError('Require ordered, unique Unicode codepoints')
+        self.characters = list(characters)
+        self.native = set(characters)
+        self.character_ids = {c: i + 4 for i, c in enumerate(characters)}
+        self.id_characters = {i: c for c, i in self.character_ids.items()}
+        self.classes = len(characters) + 4
+
+    def encode(self, text):
+        if not text or text != unicodedata.normalize('NFC', text).strip():
+            raise ValueError('Require nonempty, stripped NFC text')
+        if any(c not in self.native for c in text):
+            raise ValueError('Text outside the fixed OCR alphabet')
+        return [0, *[self.character_ids[c] for c in text], 2], 'characters'
+
+    def decode(self, ids):
+        pieces = []
+        for index in ids:
+            if index in [0, 1, 2]:
+                continue
+            if index == 3:
+                pieces.append('\ufffd')
+            elif index in self.id_characters:
+                pieces.append(self.id_characters[index])
+            else:
+                raise ValueError('Unassigned character output ID')
+        return unicodedata.normalize('NFC', ''.join(pieces))
+
+    def manifest(self):
+        return {'version': 1, 'kind': 'literal_NFC_characters', 'classes': self.classes,
+                'special_ids': {'BOS_serialization': 0, 'PAD': 1, 'EOS_and_decoder_start': 2, 'UNK': 3},
+                'character_ids': self.character_ids, 'generation_suppressed_ids': [0, 1, 3],
+                'training_target': 'Drop serialization BOS0 only; retain every text codepoint and EOS2',
+                'unknown_output': 'Decode3 as U+FFFD, never silently remove it',
+                'scope': 'Experimental codec, not recognition quality or Paddle qualification.'}

metrics.json

@@ -0,0 +1,117 @@
+{
+  "physical_neural_values_per_format": 74842752,
+  "metrics": {
+    "character": {
+      "images": 7820,
+      "sets": {
+        "hw_nep_dev": {
+          "images": 300,
+          "chars": 1524,
+          "edits": 198,
+          "exact": 196,
+          "failed": 0,
+          "empty": 0,
+          "capped": 0,
+          "CER": 0.12992125984251968
+        },
+        "hw_hindi_dev": {
+          "images": 2000,
+          "chars": 12054,
+          "edits": 974,
+          "exact": 1440,
+          "failed": 0,
+          "empty": 0,
+          "capped": 0,
+          "CER": 0.08080305292848847
+        },
+        "hw_en_dev": {
+          "images": 976,
+          "chars": 42074,
+          "edits": 2564,
+          "exact": 219,
+          "failed": 0,
+          "empty": 0,
+          "capped": 0,
+          "CER": 0.06094024813423967
+        },
+        "real_dev": {
+          "images": 2000,
+          "chars": 43430,
+          "edits": 4860,
+          "exact": 1618,
+          "failed": 0,
+          "empty": 0,
+          "capped": 0,
+          "CER": 0.11190421367718167
+        },
+        "synth_dev": {
+          "images": 800,
+          "chars": 10050,
+          "edits": 263,
+          "exact": 675,
+          "failed": 0,
+          "empty": 0,
+          "capped": 0,
+          "CER": 0.026169154228855723
+        },
+        "en_dev": {
+          "images": 1744,
+          "chars": 46033,
+          "edits": 6415,
+          "exact": 908,
+          "failed": 0,
+          "empty": 0,
+          "capped": 0,
+          "CER": 0.13935654856298743
+        }
+      },
+      "generation": {
+        "do_sample": false,
+        "num_beams": 10,
+        "max_new_tokens": 257,
+        "min_length": 0,
+        "length_penalty": 1.0,
+        "early_stopping": false,
+        "use_cache": true,
+        "suppress_tokens": [
+          0,
+          1,
+          3
+        ]
+      },
+      "native_source_exact": true,
+      "verification_sha256": "7df9ebd369fd4e323039639f1ce9602f95193e0d6a42fc20cba37ece7f052e1b"
+    },
+    "english": {
+      "images": 976,
+      "sets": {
+        "hw_en_dev": {
+          "images": 976,
+          "chars": 42074,
+          "edits": 974,
+          "exact": 605,
+          "failed": 0,
+          "empty": 0,
+          "capped": 0,
+          "CER": 0.023149688643818035
+        }
+      },
+      "generation": {
+        "do_sample": false,
+        "num_beams": 10,
+        "max_new_tokens": 256,
+        "min_length": 0,
+        "length_penalty": 1.0,
+        "early_stopping": false,
+        "use_cache": true
+      },
+      "native_source_exact": true,
+      "verification_sha256": "daa40c7506e76966acaa17313a015c4b7badb904a15aebd9ab5a26371b6098ab"
+    }
+  },
+  "normalization": "NFC codepoint CER; all samples, failures, empties and caps retained",
+  "explicit_language_branch": true,
+  "automatic_routing": false,
+  "blind_unseen_writer_page_evaluation": false,
+  "handwriting_4percent_target_achieved": false
+}

paddle/bundle.json

@@ -0,0 +1,44 @@
+{
+  "scope": "Package only native graph parameters and required assets for PaddleX research.\n\nNo source Torch weights or training controls are included. Relocation and\nPaddleX path/BGR/result behavior require a separate actual-factory audit.\n",
+  "config_sha256": "4748d485e0280227d6b395faf8c0b956217389abb4a3c267ff6cf5d427116863",
+  "assets_sha256": {
+    "candidate/bundle.json": "0510d21cd0bb7a249dafac167fd2e371ee63bd6f83a016a6c4f2e7c6bb843ede",
+    "candidate/character-assets/character_codec.json": "7bb65c86bc29b555f8da3af81b7023e370bdebe6ced8c84d104d55165c6ffb0a",
+    "candidate/character-assets/preprocessor_config.json": "36a945a7cc645688b9ef64dabae16979cf5f7c1c448569cc306694edc0598b9b",
+    "candidate/english-assets/sentencepiece.bpe.model": "6f5e2fefcf793761a76a6bfb8ad35489f9c203b25557673284b6d032f41043f4",
+    "candidate/english-assets/tokenizer_config.json": "e098c98c6834fbcd3f3e75215e3e8b8868f0fe5bf07aef93d28e4f7196db293c",
+    "candidate/english-assets/preprocessor_config.json": "eec0c686d2e2ea6d5887f87e1c6a34ef5bf1ba5b2ac98b8675e131da646fa1d2",
+    "candidate/english-assets/README.md": "7e0f183a2697956065aea4e3ec9ca718967be1b7afe446cff4dcdbc8a64fa52b",
+    "encoder/encoder.json": "eddaf1699760460ba7f2bd7fbdee9b7de9166b581e5145d9dc3c8da177bd72b7",
+    "encoder/encoder.pdiparams": "b2a632ecd78bd35441f0575b5073ebdfe2dd8edebff08c892b386205655700b0",
+    "encoder/report.json": "6e19e7fe532e7d24d504f9e323e08d5c77e1ca07ceaeb1ebcc0f616369834dbc",
+    "encoder/run.json": "42c4cf89a86fc3de900949348079f7b0c90d4e58ca2d28a105e326785a3eacd5",
+    "decoders/report.json": "24a40b59ddedd55f423035cb40def1e251633948a67802707d14cf3151a67ba8",
+    "decoders/english/decoder-step.json": "eaea1b3b5001549a6d0a4c4b1bd810371a82bdae2092452ac2f73214762ede95",
+    "decoders/english/decoder-step.pdiparams": "ec5ec71ad13c5305a1a95c8c2846c4de526571e58ce8df00fa5dbac529940a16",
+    "decoders/character/decoder-step.json": "555d79e651df3b22fef37f7cc873c843678c135bbe39e69488f4ef3b2acb3e5d",
+    "decoders/character/decoder-step.pdiparams": "e35ba8d8743abdcf8b6cdb48a572a669cbf3e504d38d2a48827b68657e32aeff",
+    "encoder-library.so": "fce69529c6a3752247be1192a495de7d1c395f8493039ad0d17546a42f936c60",
+    "code/paddlex_shared_character.py": "6a7b1d60aa337ba04514c9910e68f96d1bafdff383f3779e83f22a587707a3e3",
+    "code/shared_character_native.py": "394b45df79c5c4e84c2053b62b8d4080ef6a875034b6b16c661ac63257f56a30",
+    "code/trocr_character_codec.py": "7ede21e57fd4985153570820873abf690cc48de2e677b38828753fbedfdbd5e2",
+    "code/trocr_numpy_beam.py": "e4f1d87113404dfc16fdddd78a1cdd6477a2f886d76cef99594fc537fc531028",
+    "code/trocr_source_cpu_ops.cc": "63bb541df8f7a1f8635c946f2407aa0f449e63cfe480cbbe6c0b8d6fb2f5f4c2",
+    "code/trocr_selective_onednn_pass.cc": "ff304ef7e20f22cc1bf3c3199c6614bc3e870db640e569d465ca87c6704051b6",
+    "code/trocr_paddle_norm_probe.LICENSE": "19d6a3818fa4927890b59222656f155faae8f8817ebc9f40f779fdb6ab5a65d2",
+    "code/trocr_paddle_sleef_probe.LICENSE": "5f35cc826b3a0001293ed62e8ea95e5784f89f5e25300bf437d34be1d35e8a75",
+    "code/trocr_selective_onednn_pass.LICENSE": "ff6311c56c1b22c9209c96cc8afc16da91c24ee55a15e14ebc46fe1960dbd856"
+  },
+  "parameters": 74842752,
+  "source_Torch_weights_included": false,
+  "second_encoder_stored": false,
+  "native_generation_report_sha256": "63682d42f3edd72df566613a546aef35d86cc4a9580873a823132a3736e86722",
+  "runtime_change_since_direct_controls": "Library relocation argument, RGB preprocessing helper and graph-path diagnostics; actual factory audit is mandatory.",
+  "source_sha256": "d2d2338a9d8efafde6d1bb8995addba0c7482976ffb2dae3481de41433acc2b4",
+  "factory_contract_qualified": false,
+  "native_Paddle_qualified": false,
+  "PaddleX_qualified": false,
+  "model_promoted": false,
+  "public_release_parent_bundle_sha256": "7f4091be58fc6398363d7a192fab73af6832a543a193efada11b91f188e7d120",
+  "public_packaging_change": "Exclude conversion-control predictions; graph parameters, runtime sources, required identity reports and text assets unchanged."
+}

paddle/candidate/bundle.json

@@ -0,0 +1,2075 @@
+{
+  "format": "shared-character-private-continuation-source-v1",
+  "parameters": 74842752,
+  "corrections": [
+    {
+      "key": "t000",
+      "name": "embeddings.cls_token",
+      "shape": [
+        1,
+        1,
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t001",
+      "name": "embeddings.distillation_token",
+      "shape": [
+        1,
+        1,
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t002",
+      "name": "embeddings.patch_embeddings.projection.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t003",
+      "name": "embeddings.patch_embeddings.projection.weight",
+      "shape": [
+        384,
+        3,
+        16,
+        16
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        768
+      ]
+    },
+    {
+      "key": "t004",
+      "name": "embeddings.position_embeddings",
+      "shape": [
+        1,
+        578,
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t005",
+      "name": "encoder.layer.0.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t006",
+      "name": "encoder.layer.0.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t007",
+      "name": "encoder.layer.0.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t008",
+      "name": "encoder.layer.0.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t009",
+      "name": "encoder.layer.0.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t010",
+      "name": "encoder.layer.0.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t011",
+      "name": "encoder.layer.0.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t012",
+      "name": "encoder.layer.0.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t013",
+      "name": "encoder.layer.0.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t014",
+      "name": "encoder.layer.0.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t015",
+      "name": "encoder.layer.0.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t016",
+      "name": "encoder.layer.0.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t017",
+      "name": "encoder.layer.0.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t018",
+      "name": "encoder.layer.0.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t019",
+      "name": "encoder.layer.0.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t020",
+      "name": "encoder.layer.0.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t021",
+      "name": "encoder.layer.1.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t022",
+      "name": "encoder.layer.1.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t023",
+      "name": "encoder.layer.1.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t024",
+      "name": "encoder.layer.1.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t025",
+      "name": "encoder.layer.1.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t026",
+      "name": "encoder.layer.1.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t027",
+      "name": "encoder.layer.1.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t028",
+      "name": "encoder.layer.1.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t029",
+      "name": "encoder.layer.1.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t030",
+      "name": "encoder.layer.1.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t031",
+      "name": "encoder.layer.1.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t032",
+      "name": "encoder.layer.1.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t033",
+      "name": "encoder.layer.1.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t034",
+      "name": "encoder.layer.1.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t035",
+      "name": "encoder.layer.1.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t036",
+      "name": "encoder.layer.1.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t037",
+      "name": "encoder.layer.10.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t038",
+      "name": "encoder.layer.10.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t039",
+      "name": "encoder.layer.10.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t040",
+      "name": "encoder.layer.10.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t041",
+      "name": "encoder.layer.10.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t042",
+      "name": "encoder.layer.10.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t043",
+      "name": "encoder.layer.10.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t044",
+      "name": "encoder.layer.10.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t045",
+      "name": "encoder.layer.10.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t046",
+      "name": "encoder.layer.10.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t047",
+      "name": "encoder.layer.10.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t048",
+      "name": "encoder.layer.10.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t049",
+      "name": "encoder.layer.10.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t050",
+      "name": "encoder.layer.10.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t051",
+      "name": "encoder.layer.10.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t052",
+      "name": "encoder.layer.10.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t053",
+      "name": "encoder.layer.11.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t054",
+      "name": "encoder.layer.11.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t055",
+      "name": "encoder.layer.11.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t056",
+      "name": "encoder.layer.11.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t057",
+      "name": "encoder.layer.11.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t058",
+      "name": "encoder.layer.11.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t059",
+      "name": "encoder.layer.11.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t060",
+      "name": "encoder.layer.11.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t061",
+      "name": "encoder.layer.11.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t062",
+      "name": "encoder.layer.11.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t063",
+      "name": "encoder.layer.11.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t064",
+      "name": "encoder.layer.11.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t065",
+      "name": "encoder.layer.11.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t066",
+      "name": "encoder.layer.11.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t067",
+      "name": "encoder.layer.11.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t068",
+      "name": "encoder.layer.11.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t069",
+      "name": "encoder.layer.2.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t070",
+      "name": "encoder.layer.2.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t071",
+      "name": "encoder.layer.2.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t072",
+      "name": "encoder.layer.2.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t073",
+      "name": "encoder.layer.2.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t074",
+      "name": "encoder.layer.2.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t075",
+      "name": "encoder.layer.2.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t076",
+      "name": "encoder.layer.2.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t077",
+      "name": "encoder.layer.2.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t078",
+      "name": "encoder.layer.2.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t079",
+      "name": "encoder.layer.2.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t080",
+      "name": "encoder.layer.2.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t081",
+      "name": "encoder.layer.2.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t082",
+      "name": "encoder.layer.2.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t083",
+      "name": "encoder.layer.2.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t084",
+      "name": "encoder.layer.2.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t085",
+      "name": "encoder.layer.3.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t086",
+      "name": "encoder.layer.3.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t087",
+      "name": "encoder.layer.3.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t088",
+      "name": "encoder.layer.3.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t089",
+      "name": "encoder.layer.3.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t090",
+      "name": "encoder.layer.3.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t091",
+      "name": "encoder.layer.3.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t092",
+      "name": "encoder.layer.3.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t093",
+      "name": "encoder.layer.3.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t094",
+      "name": "encoder.layer.3.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t095",
+      "name": "encoder.layer.3.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t096",
+      "name": "encoder.layer.3.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t097",
+      "name": "encoder.layer.3.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t098",
+      "name": "encoder.layer.3.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t099",
+      "name": "encoder.layer.3.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t100",
+      "name": "encoder.layer.3.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t101",
+      "name": "encoder.layer.4.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t102",
+      "name": "encoder.layer.4.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t103",
+      "name": "encoder.layer.4.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t104",
+      "name": "encoder.layer.4.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t105",
+      "name": "encoder.layer.4.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t106",
+      "name": "encoder.layer.4.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t107",
+      "name": "encoder.layer.4.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t108",
+      "name": "encoder.layer.4.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t109",
+      "name": "encoder.layer.4.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t110",
+      "name": "encoder.layer.4.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t111",
+      "name": "encoder.layer.4.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t112",
+      "name": "encoder.layer.4.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t113",
+      "name": "encoder.layer.4.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t114",
+      "name": "encoder.layer.4.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t115",
+      "name": "encoder.layer.4.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t116",
+      "name": "encoder.layer.4.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t117",
+      "name": "encoder.layer.5.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t118",
+      "name": "encoder.layer.5.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t119",
+      "name": "encoder.layer.5.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t120",
+      "name": "encoder.layer.5.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t121",
+      "name": "encoder.layer.5.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t122",
+      "name": "encoder.layer.5.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t123",
+      "name": "encoder.layer.5.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t124",
+      "name": "encoder.layer.5.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t125",
+      "name": "encoder.layer.5.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t126",
+      "name": "encoder.layer.5.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t127",
+      "name": "encoder.layer.5.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t128",
+      "name": "encoder.layer.5.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t129",
+      "name": "encoder.layer.5.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t130",
+      "name": "encoder.layer.5.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t131",
+      "name": "encoder.layer.5.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t132",
+      "name": "encoder.layer.5.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t133",
+      "name": "encoder.layer.6.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t134",
+      "name": "encoder.layer.6.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t135",
+      "name": "encoder.layer.6.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t136",
+      "name": "encoder.layer.6.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t137",
+      "name": "encoder.layer.6.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t138",
+      "name": "encoder.layer.6.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t139",
+      "name": "encoder.layer.6.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t140",
+      "name": "encoder.layer.6.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t141",
+      "name": "encoder.layer.6.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t142",
+      "name": "encoder.layer.6.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t143",
+      "name": "encoder.layer.6.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t144",
+      "name": "encoder.layer.6.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t145",
+      "name": "encoder.layer.6.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t146",
+      "name": "encoder.layer.6.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t147",
+      "name": "encoder.layer.6.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t148",
+      "name": "encoder.layer.6.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t149",
+      "name": "encoder.layer.7.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t150",
+      "name": "encoder.layer.7.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t151",
+      "name": "encoder.layer.7.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t152",
+      "name": "encoder.layer.7.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t153",
+      "name": "encoder.layer.7.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t154",
+      "name": "encoder.layer.7.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t155",
+      "name": "encoder.layer.7.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t156",
+      "name": "encoder.layer.7.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t157",
+      "name": "encoder.layer.7.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t158",
+      "name": "encoder.layer.7.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t159",
+      "name": "encoder.layer.7.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t160",
+      "name": "encoder.layer.7.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t161",
+      "name": "encoder.layer.7.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t162",
+      "name": "encoder.layer.7.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t163",
+      "name": "encoder.layer.7.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t164",
+      "name": "encoder.layer.7.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t165",
+      "name": "encoder.layer.8.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t166",
+      "name": "encoder.layer.8.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t167",
+      "name": "encoder.layer.8.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t168",
+      "name": "encoder.layer.8.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t169",
+      "name": "encoder.layer.8.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t170",
+      "name": "encoder.layer.8.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t171",
+      "name": "encoder.layer.8.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t172",
+      "name": "encoder.layer.8.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t173",
+      "name": "encoder.layer.8.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t174",
+      "name": "encoder.layer.8.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t175",
+      "name": "encoder.layer.8.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t176",
+      "name": "encoder.layer.8.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t177",
+      "name": "encoder.layer.8.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t178",
+      "name": "encoder.layer.8.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t179",
+      "name": "encoder.layer.8.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t180",
+      "name": "encoder.layer.8.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t181",
+      "name": "encoder.layer.9.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t182",
+      "name": "encoder.layer.9.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t183",
+      "name": "encoder.layer.9.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t184",
+      "name": "encoder.layer.9.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t185",
+      "name": "encoder.layer.9.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t186",
+      "name": "encoder.layer.9.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t187",
+      "name": "encoder.layer.9.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t188",
+      "name": "encoder.layer.9.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t189",
+      "name": "encoder.layer.9.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t190",
+      "name": "encoder.layer.9.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t191",
+      "name": "encoder.layer.9.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t192",
+      "name": "encoder.layer.9.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t193",
+      "name": "encoder.layer.9.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t194",
+      "name": "encoder.layer.9.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t195",
+      "name": "encoder.layer.9.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t196",
+      "name": "encoder.layer.9.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t197",
+      "name": "layernorm.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t198",
+      "name": "layernorm.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    }
+  ],
+  "files_sha256": {
+    "weights.safetensors": "9b58d252978d3a5d33adcdaea95e113162daf5146e20a9e1b8aa47e21e48ed93",
+    "english-assets/README.md": "7e0f183a2697956065aea4e3ec9ca718967be1b7afe446cff4dcdbc8a64fa52b",
+    "english-assets/config.json": "2b6e06edec319984eca99c932d940bfa1f308db3f74d6279440b906c373bf727",
+    "english-assets/generation_config.json": "41149cdcffec4d657f32dfcddd9b208037f01286c9e07945c724908c58ed0193",
+    "english-assets/preprocessor_config.json": "eec0c686d2e2ea6d5887f87e1c6a34ef5bf1ba5b2ac98b8675e131da646fa1d2",
+    "english-assets/sentencepiece.bpe.model": "6f5e2fefcf793761a76a6bfb8ad35489f9c203b25557673284b6d032f41043f4",
+    "english-assets/special_tokens_map.json": "58dc0a3d15e3bad4861bcf2dee770ea327344d185182d4c5d2bf253285e53a83",
+    "english-assets/tokenizer_config.json": "e098c98c6834fbcd3f3e75215e3e8b8868f0fe5bf07aef93d28e4f7196db293c",
+    "character-assets/config.json": "decbe7681cc6000bd5bf83226b295e7a051e8280bc12255961e011d2caadbfce",
+    "character-assets/generation_config.json": "9577d3da76b60ee97a70d1516ac144eba329a8a70dd0e96c6f721c910434dcff",
+    "character-assets/character_codec.json": "7bb65c86bc29b555f8da3af81b7023e370bdebe6ced8c84d104d55165c6ffb0a",
+    "character-assets/preprocessor_config.json": "36a945a7cc645688b9ef64dabae16979cf5f7c1c448569cc306694edc0598b9b",
+    "model-configs.json": "1c598da84fcbe26af738d5ef1c0ea04f80f965e7690b1089aeaa8c7977cd4359",
+    "english-generation.json": "cddf3683b66672c43fee2a6d966829eebf2a00c61d376d0eb30db972693210be",
+    "character-generation.json": "02b8abbca9b3f1e5fb8ea2057b0aea84e1f488dd5dd2ca52a75850170bc3b08b"
+  },
+  "uniform_rank": 71,
+  "routing_parameters_reserved": 100000,
+  "automatic_dispatch": false,
+  "scope": "Serial CPU source candidate only; no accuracy or native Paddle qualification",
+  "original_model_sha256": "1b83102cbc1520dee1c3937ac334da83da18cf4683d46ebd1bd4e93ebe584dd7",
+  "parent_bundle_sha256": "b9281e7da68d2b9c331697a9fb15da3a2366fb937fce00742d658987d924318e",
+  "training_run_sha256": "713dfd450e19d12523585d9973aad7a7236097e64206e035800bbdcfeacb21e3",
+  "training_step": 8000,
+  "training_chain": "5e79546a138135b739b6485f3af4d81d412baa44119e849ec307ac8d87dff4ec",
+  "initial_character_model_sha256": "fcc1346b8c6aa36ae04b8628deb5cd309aa09d4574f0ed9ee0ce4d124091174a",
+  "root_bundle_sha256": "a52a68c8950c7eada39215d52b7b1eaf893d23577297951019bce382689baf48",
+  "parent_training_step": 8000,
+  "cumulative_training_step": 16000,
+  "parent_training_run_sha256": "f33da3168d15a743a4e73a59a77a55f5c85e41ee675cf006a00c940b9a23c70b"
+}

paddle/candidate/character-assets/character_codec.json

@@ -0,0 +1,590 @@
+{
+  "version": 1,
+  "kind": "literal_NFC_characters",
+  "classes": 573,
+  "special_ids": {
+    "BOS_serialization": 0,
+    "PAD": 1,
+    "EOS_and_decoder_start": 2,
+    "UNK": 3
+  },
+  "character_ids": {
+    "!": 4,
+    "\"": 5,
+    "#": 6,
+    "$": 7,
+    "%": 8,
+    "&": 9,
+    "'": 10,
+    "(": 11,
+    ")": 12,
+    "*": 13,
+    "+": 14,
+    ",": 15,
+    "-": 16,
+    ".": 17,
+    "/": 18,
+    "0": 19,
+    "1": 20,
+    "2": 21,
+    "3": 22,
+    "4": 23,
+    "5": 24,
+    "6": 25,
+    "7": 26,
+    "8": 27,
+    "9": 28,
+    ":": 29,
+    ";": 30,
+    "<": 31,
+    "=": 32,
+    ">": 33,
+    "?": 34,
+    "A": 35,
+    "B": 36,
+    "C": 37,
+    "D": 38,
+    "E": 39,
+    "F": 40,
+    "G": 41,
+    "H": 42,
+    "I": 43,
+    "J": 44,
+    "K": 45,
+    "L": 46,
+    "M": 47,
+    "N": 48,
+    "O": 49,
+    "P": 50,
+    "Q": 51,
+    "R": 52,
+    "S": 53,
+    "T": 54,
+    "U": 55,
+    "V": 56,
+    "W": 57,
+    "X": 58,
+    "Y": 59,
+    "Z": 60,
+    "[": 61,
+    "]": 62,
+    "_": 63,
+    "`": 64,
+    "a": 65,
+    "b": 66,
+    "c": 67,
+    "d": 68,
+    "e": 69,
+    "f": 70,
+    "g": 71,
+    "h": 72,
+    "i": 73,
+    "j": 74,
+    "k": 75,
+    "l": 76,
+    "m": 77,
+    "n": 78,
+    "o": 79,
+    "p": 80,
+    "q": 81,
+    "r": 82,
+    "s": 83,
+    "t": 84,
+    "u": 85,
+    "v": 86,
+    "w": 87,
+    "x": 88,
+    "y": 89,
+    "z": 90,
+    "©": 91,
+    "‥": 92,
+    "{": 93,
+    "}": 94,
+    "\\": 95,
+    "|": 96,
+    "@": 97,
+    "^": 98,
+    "~": 99,
+    "÷": 100,
+    "∕": 101,
+    "∙": 102,
+    "⋅": 103,
+    "·": 104,
+    "±": 105,
+    "∓": 106,
+    "∩": 107,
+    "∪": 108,
+    "□": 109,
+    "←": 110,
+    "↔": 111,
+    "⇒": 112,
+    "⇐": 113,
+    "⇔": 114,
+    "∀": 115,
+    "∃": 116,
+    "∄": 117,
+    "∴": 118,
+    "∵": 119,
+    "∝": 120,
+    "∞": 121,
+    "⊥": 122,
+    "∟": 123,
+    "∠": 124,
+    "∡": 125,
+    "∢": 126,
+    "′": 127,
+    "″": 128,
+    "∥": 129,
+    "⊾": 130,
+    "⊿": 131,
+    "∂": 132,
+    "∫": 133,
+    "∬": 134,
+    "∭": 135,
+    "∮": 136,
+    "∯": 137,
+    "∰": 138,
+    "∑": 139,
+    "∏": 140,
+    "√": 141,
+    "∛": 142,
+    "∜": 143,
+    "∱": 144,
+    "∲": 145,
+    "∳": 146,
+    "∶": 147,
+    "∷": 148,
+    "∼": 149,
+    "®": 150,
+    "℉": 151,
+    "Ω": 152,
+    "℧": 153,
+    "Å": 154,
+    "⌀": 155,
+    "ℏ": 156,
+    "⅀": 157,
+    "⍺": 158,
+    "⍵": 159,
+    "¢": 160,
+    "€": 161,
+    "£": 162,
+    "¥": 163,
+    "₿": 164,
+    "Ⅰ": 165,
+    "Ⅱ": 166,
+    "Ⅲ": 167,
+    "Ⅳ": 168,
+    "Ⅴ": 169,
+    "Ⅵ": 170,
+    "Ⅶ": 171,
+    "Ⅷ": 172,
+    "Ⅸ": 173,
+    "Ⅹ": 174,
+    "Ⅺ": 175,
+    "Ⅻ": 176,
+    "ⅰ": 177,
+    "ⅱ": 178,
+    "ⅲ": 179,
+    "ⅳ": 180,
+    "ⅴ": 181,
+    "ⅵ": 182,
+    "ⅶ": 183,
+    "ⅷ": 184,
+    "ⅸ": 185,
+    "ⅹ": 186,
+    "ⅺ": 187,
+    "ⅻ": 188,
+    "➀": 189,
+    "➁": 190,
+    "➂": 191,
+    "➃": 192,
+    "➄": 193,
+    "➅": 194,
+    "➆": 195,
+    "➇": 196,
+    "➈": 197,
+    "➉": 198,
+    "➊": 199,
+    "➋": 200,
+    "➌": 201,
+    "➍": 202,
+    "➎": 203,
+    "➏": 204,
+    "➐": 205,
+    "➑": 206,
+    "➒": 207,
+    "➓": 208,
+    "❶": 209,
+    "❷": 210,
+    "❸": 211,
+    "❹": 212,
+    "❺": 213,
+    "❻": 214,
+    "❼": 215,
+    "❽": 216,
+    "❾": 217,
+    "❿": 218,
+    "①": 219,
+    "②": 220,
+    "③": 221,
+    "④": 222,
+    "⑤": 223,
+    "⑥": 224,
+    "⑦": 225,
+    "⑧": 226,
+    "⑨": 227,
+    "⑩": 228,
+    "●": 229,
+    "▶": 230,
+    "𝑢": 231,
+    "︽": 232,
+    "–": 233,
+    "﹥": 234,
+    "𝜓": 235,
+    "•": 236,
+    "∋": 237,
+    "ƒ": 238,
+    "⬆": 239,
+    "Ạ": 240,
+    "◀": 241,
+    "": 242,
+    "▫": 243,
+    "︾": 244,
+    "À": 245,
+    "Á": 246,
+    "Â": 247,
+    "Ã": 248,
+    "Ä": 249,
+    "Å": 250,
+    "Æ": 251,
+    "Ç": 252,
+    "È": 253,
+    "É": 254,
+    "Ê": 255,
+    "Ë": 256,
+    "Ì": 257,
+    "Í": 258,
+    "Î": 259,
+    "Ï": 260,
+    "Ð": 261,
+    "Ñ": 262,
+    "Ò": 263,
+    "Ó": 264,
+    "Ô": 265,
+    "Õ": 266,
+    "Ö": 267,
+    "Ø": 268,
+    "Ù": 269,
+    "Ú": 270,
+    "Û": 271,
+    "Ü": 272,
+    "Ý": 273,
+    "Þ": 274,
+    "à": 275,
+    "á": 276,
+    "â": 277,
+    "ã": 278,
+    "ä": 279,
+    "å": 280,
+    "æ": 281,
+    "ç": 282,
+    "è": 283,
+    "é": 284,
+    "ê": 285,
+    "ë": 286,
+    "ì": 287,
+    "í": 288,
+    "î": 289,
+    "ï": 290,
+    "ð": 291,
+    "ñ": 292,
+    "ò": 293,
+    "ó": 294,
+    "ô": 295,
+    "õ": 296,
+    "ö": 297,
+    "ø": 298,
+    "ù": 299,
+    "ú": 300,
+    "û": 301,
+    "ü": 302,
+    "ý": 303,
+    "þ": 304,
+    "ÿ": 305,
+    "¡": 306,
+    "¤": 307,
+    "¦": 308,
+    "§": 309,
+    "¨": 310,
+    "ª": 311,
+    "«": 312,
+    "¬": 313,
+    "¯": 314,
+    "°": 315,
+    "²": 316,
+    "³": 317,
+    "´": 318,
+    "µ": 319,
+    "¶": 320,
+    "¸": 321,
+    "¹": 322,
+    "º": 323,
+    "»": 324,
+    "¼": 325,
+    "½": 326,
+    "¾": 327,
+    "¿": 328,
+    "×": 329,
+    "‐": 330,
+    "‑": 331,
+    "‒": 332,
+    "—": 333,
+    "―": 334,
+    "‖": 335,
+    "‗": 336,
+    "‘": 337,
+    "’": 338,
+    "‚": 339,
+    "‛": 340,
+    "“": 341,
+    "”": 342,
+    "„": 343,
+    "‟": 344,
+    "†": 345,
+    "‡": 346,
+    "‣": 347,
+    "․": 348,
+    "…": 349,
+    "‧": 350,
+    "‰": 351,
+    "‴": 352,
+    "‵": 353,
+    "‶": 354,
+    "‷": 355,
+    "‸": 356,
+    "‹": 357,
+    "›": 358,
+    "※": 359,
+    "‼": 360,
+    "‽": 361,
+    "‾": 362,
+    "₤": 363,
+    "₡": 364,
+    "₹": 365,
+    "−": 366,
+    "∖": 367,
+    "∗": 368,
+    "≈": 369,
+    "≠": 370,
+    "≡": 371,
+    "≤": 372,
+    "≥": 373,
+    "⊂": 374,
+    "⊃": 375,
+    "↑": 376,
+    "→": 377,
+    "↓": 378,
+    "↕": 379,
+    "™": 380,
+    "Ω": 381,
+    "℮": 382,
+    "∆": 383,
+    "✓": 384,
+    "✗": 385,
+    "✘": 386,
+    "▪": 387,
+    "◼": 388,
+    "✔": 389,
+    "✕": 390,
+    "☑": 391,
+    "☒": 392,
+    "№": 393,
+    "₽": 394,
+    "₴": 395,
+    "Α": 396,
+    "α": 397,
+    "Β": 398,
+    "β": 399,
+    "Γ": 400,
+    "γ": 401,
+    "Δ": 402,
+    "δ": 403,
+    "Ε": 404,
+    "ε": 405,
+    "Ζ": 406,
+    "ζ": 407,
+    "Η": 408,
+    "η": 409,
+    "Θ": 410,
+    "θ": 411,
+    "Ι": 412,
+    "ι": 413,
+    "Κ": 414,
+    "κ": 415,
+    "Λ": 416,
+    "λ": 417,
+    "Μ": 418,
+    "μ": 419,
+    "Ν": 420,
+    "ν": 421,
+    "Ξ": 422,
+    "ξ": 423,
+    "Ο": 424,
+    "ο": 425,
+    "Π": 426,
+    "π": 427,
+    "Ρ": 428,
+    "ρ": 429,
+    "Σ": 430,
+    "σ": 431,
+    "ς": 432,
+    "Τ": 433,
+    "τ": 434,
+    "Υ": 435,
+    "υ": 436,
+    "Φ": 437,
+    "φ": 438,
+    "Χ": 439,
+    "χ": 440,
+    "Ψ": 441,
+    "ψ": 442,
+    "ω": 443,
+    "ऀ": 444,
+    "ँ": 445,
+    "ं": 446,
+    "ः": 447,
+    "ऄ": 448,
+    "अ": 449,
+    "आ": 450,
+    "इ": 451,
+    "ई": 452,
+    "उ": 453,
+    "ऊ": 454,
+    "ऋ": 455,
+    "ऌ": 456,
+    "ऍ": 457,
+    "ऎ": 458,
+    "ए": 459,
+    "ऐ": 460,
+    "ऑ": 461,
+    "ऒ": 462,
+    "ओ": 463,
+    "औ": 464,
+    "क": 465,
+    "ख": 466,
+    "ग": 467,
+    "घ": 468,
+    "ङ": 469,
+    "च": 470,
+    "छ": 471,
+    "ज": 472,
+    "झ": 473,
+    "ञ": 474,
+    "ट": 475,
+    "ठ": 476,
+    "ड": 477,
+    "ढ": 478,
+    "ण": 479,
+    "त": 480,
+    "थ": 481,
+    "द": 482,
+    "ध": 483,
+    "न": 484,
+    "ऩ": 485,
+    "प": 486,
+    "फ": 487,
+    "ब": 488,
+    "भ": 489,
+    "म": 490,
+    "य": 491,
+    "र": 492,
+    "ऱ": 493,
+    "ल": 494,
+    "ळ": 495,
+    "ऴ": 496,
+    "व": 497,
+    "श": 498,
+    "ष": 499,
+    "स": 500,
+    "ह": 501,
+    "ऺ": 502,
+    "ऻ": 503,
+    "़": 504,
+    "ऽ": 505,
+    "ा": 506,
+    "ि": 507,
+    "ी": 508,
+    "ु": 509,
+    "ू": 510,
+    "ृ": 511,
+    "ॄ": 512,
+    "ॅ": 513,
+    "ॆ": 514,
+    "े": 515,
+    "ै": 516,
+    "ॉ": 517,
+    "ॊ": 518,
+    "ो": 519,
+    "ौ": 520,
+    "्": 521,
+    "ॎ": 522,
+    "ॏ": 523,
+    "ॐ": 524,
+    "॑": 525,
+    "॒": 526,
+    "॓": 527,
+    "॔": 528,
+    "ॕ": 529,
+    "ॖ": 530,
+    "ॗ": 531,
+    "क़": 532,
+    "ख़": 533,
+    "ग़": 534,
+    "ज़": 535,
+    "ड़": 536,
+    "ढ़": 537,
+    "फ़": 538,
+    "य़": 539,
+    "ॠ": 540,
+    "ॡ": 541,
+    "ॢ": 542,
+    "ॣ": 543,
+    "।": 544,
+    "॥": 545,
+    "०": 546,
+    "१": 547,
+    "२": 548,
+    "३": 549,
+    "४": 550,
+    "५": 551,
+    "६": 552,
+    "७": 553,
+    "८": 554,
+    "९": 555,
+    "॰": 556,
+    "ॱ": 557,
+    "ॲ": 558,
+    "ॳ": 559,
+    "ॴ": 560,
+    "ॵ": 561,
+    "ॶ": 562,
+    "ॷ": 563,
+    "ॸ": 564,
+    "ॹ": 565,
+    "ॺ": 566,
+    "ॻ": 567,
+    "ॼ": 568,
+    "ॽ": 569,
+    "ॾ": 570,
+    "ॿ": 571,
+    " ": 572
+  },
+  "generation_suppressed_ids": [
+    0,
+    1,
+    3
+  ],
+  "training_target": "Drop serialization BOS0 only; retain every text codepoint and EOS2",
+  "unknown_output": "Decode3 as U+FFFD, never silently remove it",
+  "scope": "Experimental codec, not recognition quality or Paddle qualification."
+}

paddle/candidate/character-assets/preprocessor_config.json

@@ -0,0 +1,28 @@
+{
+  "crop_size": {
+    "height": 224,
+    "width": 224
+  },
+  "do_center_crop": false,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "image_processor_type": "DeiTImageProcessor",
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "processor_class": "TrOCRProcessor",
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "height": 384,
+    "width": 384
+  }
+}

paddle/candidate/english-assets/README.md

@@ -0,0 +1,61 @@
+---
+tags:
+- trocr
+- image-to-text
+widget:
+- src: https://fki.tic.heia-fr.ch/static/img/a01-122-02.jpg
+  example_title: Note 1
+- src: https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcSoolxi9yWGAT5SLZShv8vVd0bz47UWRzQC19fDTeE8GmGv_Rn-PCF1pP1rrUx8kOjA4gg&usqp=CAU
+  example_title: Note 2
+- src: https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcRNYtTuSBpZPV_nkBYPMFwVVD9asZOPgHww4epu9EqWgDmXW--sE2o8og40ZfDGo87j5w&usqp=CAU
+  example_title: Note 3
+---
+
+# TrOCR (small-sized model, fine-tuned on IAM) 
+
+TrOCR model fine-tuned on the [IAM dataset](https://fki.tic.heia-fr.ch/databases/iam-handwriting-database). It was introduced in the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Li et al. and first released in [this repository](https://github.com/microsoft/unilm/tree/master/trocr). 
+
+
+## Model description
+
+The TrOCR model is an encoder-decoder model, consisting of an image Transformer as encoder, and a text Transformer as decoder. The image encoder was initialized from the weights of DeiT, while the text decoder was initialized from the weights of UniLM.
+
+Images are presented to the model as a sequence of fixed-size patches (resolution 16x16), which are linearly embedded. One also adds absolute position embeddings before feeding the sequence to the layers of the Transformer encoder. Next, the Transformer text decoder autoregressively generates tokens.
+
+## Intended uses & limitations
+
+You can use the raw model for optical character recognition (OCR) on single text-line images. See the [model hub](https://huggingface.co/models?search=microsoft/trocr) to look for fine-tuned versions on a task that interests you.
+
+### How to use
+
+Here is how to use this model in PyTorch:
+
+```python
+from transformers import TrOCRProcessor, VisionEncoderDecoderModel
+from PIL import Image
+import requests
+
+# load image from the IAM database
+url = 'https://fki.tic.heia-fr.ch/static/img/a01-122-02-00.jpg'
+image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
+
+processor = TrOCRProcessor.from_pretrained('microsoft/trocr-small-handwritten')
+model = VisionEncoderDecoderModel.from_pretrained('microsoft/trocr-small-handwritten')
+pixel_values = processor(images=image, return_tensors="pt").pixel_values
+
+generated_ids = model.generate(pixel_values)
+generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+```
+
+### BibTeX entry and citation info
+
+```bibtex
+@misc{li2021trocr,
+      title={TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models}, 
+      author={Minghao Li and Tengchao Lv and Lei Cui and Yijuan Lu and Dinei Florencio and Cha Zhang and Zhoujun Li and Furu Wei},
+      year={2021},
+      eprint={2109.10282},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+```

paddle/candidate/english-assets/preprocessor_config.json

@@ -0,0 +1,19 @@
+{
+  "crop_size": 224,
+  "do_center_crop": false,
+  "do_normalize": true,
+  "do_resize": true,
+  "image_processor_type": "DeiTImageProcessor",
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "resample": 3,
+  "size": 384
+}

paddle/candidate/english-assets/sentencepiece.bpe.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f5e2fefcf793761a76a6bfb8ad35489f9c203b25557673284b6d032f41043f4
+size 1356293

paddle/candidate/english-assets/tokenizer_config.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "cls_token": "<s>", "pad_token": "<pad>", "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "sp_model_kwargs": {}, "tokenizer_class": "XLMRobertaTokenizer"}

paddle/code/paddlex_shared_character.py

@@ -0,0 +1,54 @@
+"""Process-local PaddleX adapter for the shared two-branch research recognizer."""
+import math
+from pathlib import Path
+
+import numpy as np
+from PIL import Image
+from paddlex.inference.models.predictors.local_model_predictor import LocalModelPredictor
+from paddlex.inference.models.text_recognition.predictor import ImageBatchSampler,TextRecResult
+from paddlex.inference.models.bindings.registry import register_predictor_binding_map
+
+from shared_character_native import SharedCharacterNative,preprocess_rgb,sha
+
+
+MODEL_NAME='TrOCR_shared_character_Paddle_research_rec'
+
+
+class PaddleXSharedCharacter(LocalModelPredictor):
+    def __init__(self,*args,research_only_allow_unqualified=False,**kwargs):
+        super().__init__(*args,**kwargs)
+        if self.model_name!=MODEL_NAME or self.device!='cpu':raise ValueError('Require the shared CPU research model')
+        native=self.config['Native']
+        if native['format']!='shared-character-paddle-research-v1' or native['parameters']!=74842752:
+            raise ValueError('Unexpected shared recognizer profile')
+        for name,digest in native['files_sha256'].items():
+            relative=Path(name)
+            if relative.is_absolute() or '..' in relative.parts or relative.as_posix()!=name:raise ValueError('Invalid bundle asset path')
+            if sha(self.model_dir/relative)!=digest:raise ValueError('Shared bundle asset bytes changed')
+        self.runtime=SharedCharacterNative(self.model_dir/'candidate',self.model_dir/'encoder',self.model_dir/'decoders',
+            allow_unqualified=research_only_allow_unqualified,library_path=self.model_dir/'encoder-library.so')
+
+    def _build_batch_sampler(self):return ImageBatchSampler()
+    def _get_result_class(self):return TextRecResult
+
+    def process(self,batch_data,branch=None,return_word_box=False):
+        if branch not in ['english','character']:raise ValueError('Explicit research branch is required')
+        if return_word_box:raise ValueError('Autoregressive word-box alignment is not implemented')
+        images=[];outputs=[]
+        for item in batch_data.instances:
+            if isinstance(item,(str,Path)):
+                with Image.open(item) as source:image=source.convert('RGB')
+                bgr=np.asarray(image)[:,:,::-1].copy()
+            elif isinstance(item,np.ndarray) and item.dtype==np.uint8 and item.ndim==3 and item.shape[2]==3:
+                bgr=item;image=Image.fromarray(np.ascontiguousarray(bgr[:,:,::-1]))
+            else:raise ValueError('Require an image path or uint8 HWC BGR array')
+            images.append(bgr);outputs.append(self.runtime.predict_pixels(preprocess_rgb(image),branch))
+        return dict(input_path=batch_data.input_paths,page_index=batch_data.page_indexes,input_img=images,
+            rec_text=[r['text'] for r in outputs],rec_score=[min(1.,math.exp(r['score'])) for r in outputs],
+            beam_score=[r['score'] for r in outputs],tokens=[r['tokens'] for r in outputs],
+            decoder_iterations=[r['decoder_iterations'] for r in outputs],
+            token_limit_reached=[r['token_limit_reached'] for r in outputs],recognizer_branch=[branch]*len(outputs),
+            score_kind=['Uncalibrated geometric mean of generated token probabilities']*len(outputs),vis_font=[None]*len(outputs))
+
+
+def register():register_predictor_binding_map(PaddleXSharedCharacter,{'flexible':[MODEL_NAME]})

paddle/code/shared_character_native.py

@@ -0,0 +1,139 @@
+"""Serial CPU Paddle research runtime for the three shared recognizer graphs.
+
+Consumes only native graph parameters, codec/tokenizer assets and the compiled
+encoder operator library. No Torch weights, source logits, source caches or
+forced token histories enter free generation. Explicit branch selection only.
+"""
+import hashlib
+import json
+from pathlib import Path
+
+import numpy as np
+import paddle
+from paddle.utils.cpp_extension import load_op_meta_info_and_register_op
+import sentencepiece as spm
+from PIL import Image
+
+from trocr_character_codec import CharacterCodec
+from trocr_numpy_beam import TrOCRBeamState
+
+
+_LIBRARY_HASH=None
+
+
+def preprocess_rgb(image):
+    image=image.convert('RGB').resize((384,384),resample=Image.Resampling.BICUBIC)
+    pixels=(np.asarray(image).astype(np.float64)*(1./255.)).astype(np.float32)
+    return np.ascontiguousarray(((pixels-np.float32(.5))/np.float32(.5)).transpose(2,0,1)[None])
+
+
+def read(path):return json.loads(Path(path).read_text())
+def sha(path):
+    value=hashlib.sha256()
+    with Path(path).open('rb') as stream:
+        for part in iter(lambda:stream.read(1024*1024),b''):value.update(part)
+    return value.hexdigest()
+
+
+class SharedCharacterNative:
+    def __init__(self,candidate,encoder,decoders,allow_unqualified=False,library_path=None,allow_control=False):
+        if not allow_unqualified:raise ValueError('Source research model; full native/PaddleX acceptance is pending')
+        if paddle.device.get_device()!='cpu':raise ValueError('Only the validated CPU operator profile is supported')
+        candidate,encoder,decoders=map(Path,[candidate,encoder,decoders])
+        manifest=read(candidate/'bundle.json');enc=read(encoder/'report.json');enc_run=read(encoder/'run.json');dec=read(decoders/'report.json')
+        digest=sha(candidate/'bundle.json')
+        assert digest==enc_run['candidate_manifest_sha256']==dec['candidate_manifest_sha256']
+        if digest=='a52a68c8950c7eada39215d52b7b1eaf893d23577297951019bce382689baf48':
+            assert not allow_control
+        else:
+            certificate=enc_run['bundle_identity']
+            assert certificate==dec['bundle_identity'] and certificate['candidate_manifest_sha256']==digest
+            assert certificate['trained_bundle'] and certificate['control_weights']==allow_control
+            assert certificate['weights_sha256']==manifest['files_sha256']['weights.safetensors']
+            if manifest['format']=='shared-character-private-continuation-source-v1':
+                assert manifest['parent_bundle_sha256']=='b9281e7da68d2b9c331697a9fb15da3a2366fb937fce00742d658987d924318e'
+                assert manifest['root_bundle_sha256']=='a52a68c8950c7eada39215d52b7b1eaf893d23577297951019bce382689baf48'
+                assert manifest['parent_training_step']==8000 and manifest['cumulative_training_step']==8000+manifest['training_step']
+            else:
+                assert manifest['format']=='shared-character-private-trained-source-v1'
+                assert manifest['parent_bundle_sha256']=='a52a68c8950c7eada39215d52b7b1eaf893d23577297951019bce382689baf48'
+            assert manifest['training_step'] in ([2] if allow_control else [2000,8000])
+        assert sha(encoder/'report.json')==dec['encoder_report_sha256']
+        assert enc['encoder_numerical_pass'] and enc['saved_neural_population_byte_exact']
+        assert dec['total_graph_neural_values']==manifest['parameters']==74842752
+        for name in ['character-assets/character_codec.json','english-assets/sentencepiece.bpe.model','english-assets/tokenizer_config.json']:
+            assert sha(candidate/name)==manifest['files_sha256'][name]
+        self.codec=CharacterCodec(list(read(candidate/'character-assets/character_codec.json')['character_ids']))
+        assert self.codec.classes==573
+        token_config=read(candidate/'english-assets/tokenizer_config.json')
+        assert token_config['tokenizer_class']=='XLMRobertaTokenizer' and token_config.get('clean_up_tokenization_spaces',False) is False
+        self.pieces=spm.SentencePieceProcessor(model_file=str(candidate/'english-assets/sentencepiece.bpe.model'))
+        assert self.pieces.get_piece_size()==64000
+        library=dec['encoder_library'];resolved_library=str(library_path or library['path'])
+        assert sha(resolved_library)==library['sha256']
+        self.operator_library_path=str(Path(resolved_library).resolve())
+        self.graph_files=[]
+        global _LIBRARY_HASH
+        if _LIBRARY_HASH is None:
+            load_op_meta_info_and_register_op(resolved_library);_LIBRARY_HASH=library['sha256']
+        else:assert _LIBRARY_HASH==library['sha256'],'Use a fresh process for a different operator library'
+
+        def load(folder,stem,hashes,selective):
+            for name,value in hashes.items():assert sha(folder/name)==value
+            self.graph_files.extend([str((folder/(stem+suffix)).resolve()) for suffix in ['.json','.pdiparams']])
+            config=paddle.inference.Config(str(folder/(stem+'.json')),str(folder/(stem+'.pdiparams')))
+            config.disable_gpu();config.set_cpu_math_library_num_threads(2)
+            if selective:
+                config.enable_mkldnn();config.enable_custom_passes(['ocr_conv_gelu_onednn_v1'],True)
+            else:config.disable_mkldnn()
+            config.switch_ir_optim(False);config.disable_glog_info()
+            return paddle.inference.create_predictor(config)
+        self.encoder=load(encoder,'encoder',enc['graph_sha256'],True)
+        self.decoders={branch:load(decoders/branch,'decoder-step',record['graph_sha256'],False)
+                       for branch,record in dec['decoders'].items()}
+        assert set(self.encoder.get_input_names())=={'pixels','character_branch'}
+        assert all(set(p.get_input_names())=={'token','self_keys','self_values','cross_keys','cross_values'} for p in self.decoders.values())
+
+    @staticmethod
+    def run(predictor,feeds):
+        assert set(predictor.get_input_names())==set(feeds)
+        for name,value in feeds.items():predictor.get_input_handle(name).copy_from_cpu(np.ascontiguousarray(value))
+        predictor.run()
+        return [predictor.get_output_handle(n).copy_to_cpu() for n in predictor.get_output_names()]
+
+    def encode(self,pixels,branch):
+        assert branch in self.decoders and pixels.dtype==np.float32 and pixels.shape==(1,3,384,384)
+        outputs=self.run(self.encoder,dict(pixels=pixels,character_branch=np.array([branch=='character'],np.bool_)))
+        assert len(outputs)==2 and all(v.shape==(6,1,8,578,32) for v in outputs)
+        return [np.repeat(v,10,axis=1) for v in outputs]
+
+    def step(self,branch,token,self_keys,self_values,cross_keys,cross_values):
+        outputs=self.run(self.decoders[branch],dict(token=token,self_keys=self_keys,self_values=self_values,
+                                                  cross_keys=cross_keys,cross_values=cross_values))
+        assert len(outputs)==3 and outputs[0].shape==(10,64044 if branch=='english' else 573)
+        assert outputs[1].shape==outputs[2].shape==(6,10,8,self_keys.shape[3]+1,32)
+        assert all(v.dtype==np.float32 and np.isfinite(v).all() for v in outputs)
+        return outputs
+
+    def decode(self,tokens,branch):
+        if branch=='character':return self.codec.decode(tokens)
+        pieces=[]
+        for token in tokens:
+            if token in [0,1,2,3,64001]:continue
+            if not 4<=token<=64000:raise ValueError('Unassigned original output row')
+            pieces.append(self.pieces.id_to_piece(token-1))
+        return ''.join(pieces).replace('\u2581',' ').strip()
+
+    def predict_pixels(self,pixels,branch):
+        cross=self.encode(pixels,branch)
+        state=TrOCRBeamState(beams=10,max_new_tokens=256 if branch=='english' else 257,
+            length_penalty=1.,suppress_tokens=() if branch=='english' else (0,1,3))
+        keys=np.zeros((6,10,8,0,32),np.float32);values=keys.copy();token=np.full((10,1),2,np.int64)
+        while not state.done:
+            logits,next_keys,next_values=self.step(branch,token,keys,values,*cross)
+            parents,selected=state.advance(logits)
+            keys=np.ascontiguousarray(next_keys[:,parents]);values=np.ascontiguousarray(next_values[:,parents]);token=selected[:,None]
+        tokens,score=state.result()
+        return dict(tokens=tokens,text=self.decode(tokens,branch),score=score,
+            token_limit_reached=len(tokens)-1>=(256 if branch=='english' else 257) and tokens[-1]!=2,
+            decoder_iterations=state.current-1)

paddle/code/trocr_character_codec.py

@@ -0,0 +1,46 @@
+"""Literal NFC character codec for the experimental compact TrOCR decoder.
+
+Tokenizer BOS0 is serialization only; training targets exclude it. Generation
+starts with EOS2 and suppresses BOS/PAD/UNK. Unknown outputs remain visible as
+replacement characters when decoded, never silently removed.
+"""
+import unicodedata
+
+
+class CharacterCodec:
+    def __init__(self, characters):
+        if len(characters) != len(set(characters)) or any(len(c) != 1 for c in characters):
+            raise ValueError('Require ordered, unique Unicode codepoints')
+        self.characters = list(characters)
+        self.native = set(characters)
+        self.character_ids = {c: i + 4 for i, c in enumerate(characters)}
+        self.id_characters = {i: c for c, i in self.character_ids.items()}
+        self.classes = len(characters) + 4
+
+    def encode(self, text):
+        if not text or text != unicodedata.normalize('NFC', text).strip():
+            raise ValueError('Require nonempty, stripped NFC text')
+        if any(c not in self.native for c in text):
+            raise ValueError('Text outside the fixed OCR alphabet')
+        return [0, *[self.character_ids[c] for c in text], 2], 'characters'
+
+    def decode(self, ids):
+        pieces = []
+        for index in ids:
+            if index in [0, 1, 2]:
+                continue
+            if index == 3:
+                pieces.append('\ufffd')
+            elif index in self.id_characters:
+                pieces.append(self.id_characters[index])
+            else:
+                raise ValueError('Unassigned character output ID')
+        return unicodedata.normalize('NFC', ''.join(pieces))
+
+    def manifest(self):
+        return {'version': 1, 'kind': 'literal_NFC_characters', 'classes': self.classes,
+                'special_ids': {'BOS_serialization': 0, 'PAD': 1, 'EOS_and_decoder_start': 2, 'UNK': 3},
+                'character_ids': self.character_ids, 'generation_suppressed_ids': [0, 1, 3],
+                'training_target': 'Drop serialization BOS0 only; retain every text codepoint and EOS2',
+                'unknown_output': 'Decode3 as U+FFFD, never silently remove it',
+                'scope': 'Experimental codec, not recognition quality or Paddle qualification.'}

paddle/code/trocr_numpy_beam.py

@@ -0,0 +1,80 @@
+"""Batch-one NumPy beam state for the pinned TrOCR generation policy.
+
+Implements Transformers4.57's finished/running beam rules. Equal-score top-k
+ties use lower indices, explicitly; Torch CPU/CUDA tie order is not portable.
+Source equivalence and complete OCR qualification must be measured separately.
+"""
+import numpy as np
+
+
+def top_indices(scores, count):
+    scores = np.asarray(scores)
+    assert scores.ndim == 1 and 1 <= count <= scores.size
+    assert not np.isnan(scores).any() and not np.isposinf(scores).any() and np.isfinite(scores).sum() >= count
+    threshold = np.partition(scores, scores.size - count)[scores.size - count]
+    better = np.flatnonzero(scores > threshold)
+    tied = np.flatnonzero(scores == threshold)[:count - len(better)]
+    indices = np.concatenate([better, tied])
+    return indices[np.lexsort((indices, -scores[indices]))]
+
+
+def log_softmax(scores):
+    scores = np.asarray(scores, dtype=np.float32)
+    shifted = scores - scores.max(axis=-1, keepdims=True)
+    return shifted - np.log(np.exp(shifted).sum(axis=-1, keepdims=True))
+
+
+class TrOCRBeamState:
+    def __init__(self, beams=10, max_new_tokens=256, length_penalty=1., suppress_tokens=()):
+        if beams < 2 or not 1 <= max_new_tokens <= 512 or length_penalty < 0:
+            raise ValueError('Unsupported beam search settings')
+        self.beams = beams;self.maximum = max_new_tokens + 1;self.penalty = length_penalty
+        self.suppressed = tuple(suppress_tokens)
+        if any(type(t) is not int or t < 0 for t in self.suppressed) or len(set(self.suppressed)) != len(self.suppressed):
+            raise ValueError('Suppressed token IDs must be unique nonnegative integers')
+        self.current = 1
+        self.running = np.full((beams, self.maximum), 1, np.int64);self.running[:, 0] = 2
+        self.running_scores = np.full(beams, -1e9, np.float32);self.running_scores[0] = 0
+        self.finished = self.running.copy();self.finished_scores = np.full(beams, -1e9, np.float32)
+        self.finished_flags = np.zeros(beams, bool);self.finished_lengths = np.zeros(beams, np.int64)
+        self.done = False
+
+    def advance(self, logits):
+        logits = np.asarray(logits, dtype=np.float32)
+        if self.done or logits.ndim != 2 or logits.shape[0] != self.beams or logits.shape[1] < 3 or not np.isfinite(logits).all():
+            raise ValueError('Invalid scores or completed beam search')
+        classes = logits.shape[1]
+        if self.suppressed and max(self.suppressed) >= classes:raise ValueError('Suppressed ID outside vocabulary')
+        logprobs = log_softmax(logits)
+        # HF applies suppression after log-softmax, without renormalization.
+        logprobs[:, self.suppressed] = -np.inf
+        accumulated = logprobs + self.running_scores[:, None]
+        choices = top_indices(accumulated.reshape(-1), 2 * self.beams)
+        scores = accumulated.reshape(-1)[choices]
+        parents, tokens = choices // classes, choices % classes
+        candidates = self.running[parents].copy();candidates[:, self.current] = tokens
+        stopped = (tokens == 2) | (self.current + 1 == self.maximum)
+
+        live_scores = scores + stopped.astype(np.float32) * np.float32(-1e9)
+        live = top_indices(live_scores, self.beams)
+        self.running = candidates[live].copy();self.running_scores = live_scores[live].copy()
+
+        eligible = stopped & (np.arange(2 * self.beams) < self.beams)
+        normalized = scores / np.float32(self.current ** self.penalty)
+        normalized = normalized + (~eligible).astype(np.float32) * np.float32(-1e9)
+        merged_scores = np.concatenate([self.finished_scores, normalized])
+        selected = top_indices(merged_scores, self.beams)
+        self.finished = np.concatenate([self.finished, candidates])[selected].copy()
+        self.finished_scores = merged_scores[selected].copy()
+        self.finished_flags = np.concatenate([self.finished_flags, eligible])[selected].copy()
+        self.finished_lengths = np.concatenate([self.finished_lengths, np.full(2 * self.beams, self.current, np.int64)])[selected].copy()
+
+        self.current += 1
+        possible = self.running_scores[0] / np.float32((self.current - 1) ** self.penalty)
+        worst = self.finished_scores.min() if self.finished_flags.all() else np.float32(-1e9)
+        self.done = bool(possible <= worst or stopped.all())
+        return parents[live].copy(), tokens[live].astype(np.int64, copy=True)
+
+    def result(self):
+        if not self.done or not self.finished_flags[0]:raise ValueError('Beam search has no completed result')
+        return self.finished[0, :1 + self.finished_lengths[0]].tolist(), float(self.finished_scores[0])

paddle/code/trocr_paddle_norm_probe.LICENSE

@@ -0,0 +1,96 @@
+The source-order Welford and softmax prototypes in
+trocr_paddle_norm_probe.py and trocr_paddle_attention_probe.py follow the
+PyTorch ATen reduction algorithms at commit
+08187d9e0fba026dc8217405802ab5381dc88d90:
+https://github.com/pytorch/pytorch/blob/08187d9e0fba026dc8217405802ab5381dc88d90/aten/src/ATen/native/cpu/moments_utils.h
+
+https://github.com/pytorch/pytorch/blob/08187d9e0fba026dc8217405802ab5381dc88d90/aten/src/ATen/cpu/vec/functional_base.h
+https://github.com/pytorch/pytorch/blob/08187d9e0fba026dc8217405802ab5381dc88d90/aten/src/ATen/native/cpu/SoftMaxKernel.cpp
+
+The following upstream license notice is retained for those adaptations.
+The prototype is experimental and has not passed the native numerical gate.
+
+From PyTorch:
+
+Copyright (c) 2016-     Facebook, Inc            (Adam Paszke)
+Copyright (c) 2014-     Facebook, Inc            (Soumith Chintala)
+Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)
+Copyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu)
+Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)
+Copyright (c) 2011-2013 NYU                      (Clement Farabet)
+Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston)
+Copyright (c) 2006      Idiap Research Institute (Samy Bengio)
+Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz)
+
+From Caffe2:
+
+Copyright (c) 2016-present, Facebook Inc. All rights reserved.
+
+All contributions by Facebook:
+Copyright (c) 2016 Facebook Inc.
+
+All contributions by Google:
+Copyright (c) 2015 Google Inc.
+All rights reserved.
+
+All contributions by Yangqing Jia:
+Copyright (c) 2015 Yangqing Jia
+All rights reserved.
+
+All contributions by Kakao Brain:
+Copyright 2019-2020 Kakao Brain
+
+All contributions by Cruise LLC:
+Copyright (c) 2022 Cruise LLC.
+All rights reserved.
+
+All contributions by Tri Dao:
+Copyright (c) 2024 Tri Dao.
+All rights reserved.
+
+All contributions by Arm:
+Copyright (c) 2021, 2023-2025 Arm Limited and/or its affiliates
+
+All contributions from Caffe:
+Copyright(c) 2013, 2014, 2015, the respective contributors
+All rights reserved.
+
+All other contributions:
+Copyright(c) 2015, 2016 the respective contributors
+All rights reserved.
+
+Caffe2 uses a copyright model similar to Caffe: each contributor holds
+copyright over their contributions to Caffe2. The project versioning records
+all such contribution and copyright details. If a contributor wants to further
+mark their specific copyright on a particular contribution, they should
+indicate their copyright solely in the commit message of the change when it is
+committed.
+
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+3. Neither the names of Facebook, Deepmind Technologies, NYU, NEC Laboratories America
+   and IDIAP Research Institute nor the names of its contributors may be
+   used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.

paddle/code/trocr_paddle_sleef_probe.LICENSE

@@ -0,0 +1,33 @@
+The experimental xexpf adaptation in trocr_paddle_sleef_probe.py follows SLEEF
+commit 5a1d179df9cf652951b59010a2d2075372d67f68, the submodule pinned by
+PyTorch commit 08187d9e0fba026dc8217405802ab5381dc88d90.
+Source: https://github.com/shibatch/sleef/blob/5a1d179df9cf652951b59010a2d2075372d67f68/src/libm/sleefsimdsp.c
+
+//   Copyright Naoki Shibata and contributors 2010 - 2024.
+// Distributed under the Boost Software License, Version 1.0.
+//    (See accompanying file LICENSE.txt or copy at
+//          http://www.boost.org/LICENSE_1_0.txt)
+
+Boost Software License - Version 1.0 - August 17th, 2003
+
+Permission is hereby granted, free of charge, to any person or organization
+obtaining a copy of the software and accompanying documentation covered by
+this license (the "Software") to use, reproduce, display, distribute,
+execute, and transmit the Software, and to prepare derivative works of the
+Software, and to permit third-parties to whom the Software is furnished to
+do so, all subject to the following:
+
+The copyright notices in the Software and this entire statement, including
+the above license grant, this restriction and the following disclaimer,
+must be included in all copies of the Software, in whole or in part, and
+all derivative works of the Software, unless such copies or derivative
+works are solely in the form of machine-executable object code generated by
+a source language processor.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON-INFRINGEMENT. IN NO EVENT
+SHALL THE COPYRIGHT HOLDERS OR ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE
+FOR ANY DAMAGES OR OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE,
+ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+DEALINGS IN THE SOFTWARE.

paddle/code/trocr_selective_onednn_pass.LICENSE

@@ -0,0 +1,203 @@
+Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

paddle/code/trocr_selective_onednn_pass.cc

@@ -0,0 +1,52 @@
+// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+// Copyright (c) 2026 OCR Tech contributors.
+// SPDX-License-Identifier: Apache-2.0
+// Adapted from Paddle v3.3.1 onednn_placement_pass.cc; see the adjacent LICENSE.
+// Restrict placement to the two source-matched primitives, leaving BLAS intact.
+#include <iostream>
+#include "paddle/fluid/pir/dialect/operator/interface/op_yaml_info.h"
+#include "paddle/fluid/pir/dialect/operator/utils/op_yaml_info_util.h"
+#include "paddle/pir/include/pass/pass.h"
+#include "paddle/pir/include/pass/pass_registry.h"
+#include "paddle/pir/include/pattern_rewrite/pattern_match.h"
+
+namespace {
+class SelectPrimitive final : public pir::RewritePattern {
+ public:
+  SelectPrimitive(const std::string& name, pir::IrContext* context)
+      : pir::RewritePattern(name, 1, context) {}
+  bool MatchAndRewrite(pir::Operation* op,
+                       pir::PatternRewriter& rewriter) const override {
+    const auto name = op->name();
+    if (name != "pd_op.conv2d" && name != "pd_op.gelu") return false;
+    const auto target = "onednn_op" + name.substr(5);
+    const auto info = pir::IrContext::Instance()->GetRegisteredOpInfo(target);
+    PADDLE_ENFORCE_EQ(bool(info), true,
+                     common::errors::NotFound("Missing selective primitive %s", target));
+    auto attrs = op->attributes();
+    const auto* yaml = info.GetInterfaceImpl<paddle::dialect::OpYamlInfoInterface>();
+    const auto runtime = std::get<3>(yaml->get_op_info_(target));
+    for (const auto& entry : runtime.extra_args_default_value) attrs[entry.first] = entry.second;
+    if (attrs.count("is_test")) attrs["is_test"] = rewriter.bool_attr(true);
+    std::vector<pir::Type> types;
+    for (size_t i = 0; i < op->num_results(); ++i) types.push_back(op->result_type(i));
+    auto* replacement = rewriter.Build(op->operands_source(), attrs, types, info);
+    rewriter.ReplaceOp(op, replacement->results());
+    std::cerr << "OCR_SELECTIVE_ONEDNN " << name << " -> " << target << std::endl;
+    return true;
+  }
+};
+
+class SelectivePass final : public pir::PatternRewritePass {
+ public:
+  SelectivePass() : pir::PatternRewritePass("ocr_conv_gelu_onednn_v1", 2) {}
+  pir::RewritePatternSet InitializePatterns(pir::IrContext* context) override {
+    pir::RewritePatternSet patterns(context);
+    for (const auto& name : {"pd_op.conv2d", "pd_op.gelu"}) {
+      patterns.Add(std::make_unique<SelectPrimitive>(name, context));
+    }
+    return patterns;
+  }
+};
+}  // namespace
+REGISTER_IR_PASS(ocr_conv_gelu_onednn_v1, SelectivePass);

paddle/code/trocr_source_cpu_ops.cc

@@ -0,0 +1,141 @@
+// Experimental Paddle CPU operators for the fixed TrOCR encoder profile.
+// Welford/reduction order follows pinned PyTorch; exp follows pinned SLEEF.
+// Retain trocr_paddle_norm_probe.LICENSE and trocr_paddle_sleef_probe.LICENSE.
+#include <array>
+#include <cfenv>
+#include <cmath>
+#include <cstdint>
+#include <vector>
+#include "paddle/extension.h"
+
+namespace {
+constexpr int kWidth = 384;
+constexpr int kAttention = 578;
+struct Moments {
+  int count = 0;
+  std::array<float, 8> mean{};
+  std::array<float, 8> m2{};
+};
+void Merge(Moments& dst, const Moments& src) {
+  const int old = dst.count;
+  const float factor = float(src.count) / float(old + src.count);
+  for (int i = 0; i < 8; ++i) {
+    const float delta = src.mean[i] - dst.mean[i];
+    const float scaled = factor * delta;
+    dst.mean[i] = dst.mean[i] + scaled;
+    dst.m2[i] = std::fma(delta * float(old), scaled, dst.m2[i] + src.m2[i]);
+  }
+  dst.count += src.count;
+}
+void CheckCPU(const paddle::Tensor& value) {
+  PD_CHECK(value.is_cpu(), "Source-order research operators require CPU inputs.");
+  PD_CHECK(value.dtype() == paddle::DataType::FLOAT32, "Require FP32 inputs.");
+  PD_CHECK(std::fegetround() == FE_TONEAREST, "Require round-to-nearest floating point.");
+}
+float NegativeExp(float d) {
+  if (d < -104.f) return 0.f;
+  const int q = int(std::nearbyint(d * 1.4426950408889634073599246810018921374f));
+  float s = std::fma(float(q), -0.693145751953125f, d);
+  s = std::fma(float(q), -1.428606765330187045e-6f, s);
+  float u = 0.000198527617612853646278381f;
+  u = std::fma(u, s, 0.00139304355252534151077271f);
+  u = std::fma(u, s, 0.00833336077630519866943359f);
+  u = std::fma(u, s, 0.0416664853692054748535156f);
+  u = std::fma(u, s, 0.166666671633720397949219f);
+  u = std::fma(u, s, 0.5f);
+  u = 1.f + std::fma(s * s, u, s);
+  const int half = int(std::floor(float(q) * .5f));
+  return std::ldexp(std::ldexp(u, half), q - half);
+}
+}  // namespace
+
+std::vector<paddle::Tensor> SourceNorm384(const paddle::Tensor& x,
+                                         const paddle::Tensor& gamma,
+                                         const paddle::Tensor& beta) {
+  CheckCPU(x); CheckCPU(gamma); CheckCPU(beta);
+  PD_CHECK(!x.shape().empty() && x.shape().back() == kWidth, "Require 384 channels.");
+  PD_CHECK(gamma.numel() == kWidth && beta.numel() == kWidth, "Require 384 affine values.");
+  auto out = paddle::empty(x.shape(), x.dtype(), x.place());
+  const float* input = x.data<float>();
+  const float* weight = gamma.data<float>();
+  const float* bias = beta.data<float>();
+  float* result = out.data<float>();
+  for (int64_t row = 0; row < x.numel() / kWidth; ++row) {
+    const float* values = input + row * kWidth;
+    Moments first, second;
+    for (int chunk = 0; chunk < 3; ++chunk) {
+      Moments local; local.count = 16;
+      for (int j = 0; j < 16; ++j) {
+        const float factor = 1.f / float(j + 1);
+        for (int lane = 0; lane < 8; ++lane) {
+          const float value = values[chunk * 128 + j * 8 + lane];
+          const float delta = value - local.mean[lane];
+          local.mean[lane] = std::fma(factor, delta, local.mean[lane]);
+          local.m2[lane] = std::fma(delta, value - local.mean[lane], local.m2[lane]);
+        }
+      }
+      Merge(first, local);
+      if (chunk == 1) { Merge(second, first); first = Moments{}; }
+    }
+    Merge(first, second);
+    float mean = 0.f, m2 = 0.f;
+    for (int lane = 0; lane < 8; ++lane) {
+      const float factor = 48.f / float((lane + 1) * 48);
+      const float delta = first.mean[lane] - mean;
+      mean = std::fma(factor, delta, mean);
+      const float correction = (delta * delta) * factor;
+      m2 = m2 + std::fma(correction, float(lane * 48), first.m2[lane]);
+    }
+    const float variance = m2 / float(kWidth);
+    const float scale = 1.f / std::sqrt(variance + 1e-12f);
+    for (int i = 0; i < kWidth; ++i) {
+      const float normalized = (values[i] - mean) * scale;
+      result[row * kWidth + i] = std::fma(normalized, weight[i], bias[i]);
+    }
+  }
+  return {out};
+}
+
+std::vector<paddle::Tensor> SourceSoftmax578(const paddle::Tensor& x) {
+  CheckCPU(x);
+  PD_CHECK(!x.shape().empty() && x.shape().back() == kAttention, "Require 578 attention scores.");
+  auto out = paddle::empty(x.shape(), x.dtype(), x.place());
+  const float* input = x.data<float>(); float* output = out.data<float>();
+  for (int64_t row = 0; row < x.numel() / kAttention; ++row) {
+    const float* values = input + row * kAttention;
+    float* result = output + row * kAttention;
+    float maximum = values[0];
+    for (int i = 0; i < kAttention; ++i) {
+      PD_CHECK(std::isfinite(values[i]), "Require finite scores.");
+      if (values[i] > maximum) maximum = values[i];
+    }
+    for (int i = 0; i < kAttention; ++i) result[i] = NegativeExp(values[i] - maximum);
+    std::array<float, 16> sum;
+    for (int i = 0; i < 16; ++i) sum[i] = result[i];
+    for (int block = 1; block < 36; ++block)
+      for (int lane = 0; lane < 16; ++lane) sum[lane] = sum[lane] + result[block * 16 + lane];
+    for (int lane = 0; lane < 2; ++lane) sum[lane] = sum[lane] + result[576 + lane];
+    for (int width : {8, 4, 2, 1})
+      for (int lane = 0; lane < width; ++lane) sum[lane] = sum[lane] + sum[lane + width];
+    const float inverse = 1.f / sum[0];
+    for (int i = 0; i < kAttention; ++i) result[i] = result[i] * inverse;
+  }
+  return {out};
+}
+
+std::vector<std::vector<int64_t>> NormShape(const std::vector<int64_t>& x,
+                                           const std::vector<int64_t>& gamma,
+                                           const std::vector<int64_t>& beta) { return {x}; }
+std::vector<paddle::DataType> NormType(paddle::DataType x, paddle::DataType gamma,
+                                     paddle::DataType beta) { return {x}; }
+std::vector<std::vector<int64_t>> SoftmaxShape(const std::vector<int64_t>& x) { return {x}; }
+std::vector<paddle::DataType> SoftmaxType(paddle::DataType x) { return {x}; }
+
+PD_BUILD_OP(ocr_source_norm384_v1)
+    .Inputs({"X", "Gamma", "Beta"}).Outputs({"Out"})
+    .SetKernelFn(PD_KERNEL(SourceNorm384))
+    .SetInferShapeFn(PD_INFER_SHAPE(NormShape)).SetInferDtypeFn(PD_INFER_DTYPE(NormType));
+PD_BUILD_OP(ocr_source_softmax578_v1)
+    .Inputs({"X"}).Outputs({"Out"})
+    .SetKernelFn(PD_KERNEL(SourceSoftmax578))
+    .SetInferShapeFn(PD_INFER_SHAPE(SoftmaxShape)).SetInferDtypeFn(PD_INFER_DTYPE(SoftmaxType));

paddle/decoders/character/decoder-step.pdiparams

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e35ba8d8743abdcf8b6cdb48a572a669cbf3e504d38d2a48827b68657e32aeff
+size 23830010

paddle/decoders/english/decoder-step.pdiparams

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec5ec71ad13c5305a1a95c8c2846c4de526571e58ce8df00fa5dbac529940a16
+size 153818620

paddle/decoders/report.json

@@ -0,0 +1,63 @@
+{
+  "scope": "Export the two disjoint cached Paddle decoders for the shared encoder.\n\nRound-trip/count every decoder parameter. This completes graph assets only:\ncached numerical trajectories, beam generation and native CER remain untested.\nNo Torch, duplicate encoder, checkpoint modification or deployment.\n",
+  "candidate_manifest_sha256": "0510d21cd0bb7a249dafac167fd2e371ee63bd6f83a016a6c4f2e7c6bb843ede",
+  "bundle_identity": {
+    "candidate_manifest_sha256": "0510d21cd0bb7a249dafac167fd2e371ee63bd6f83a016a6c4f2e7c6bb843ede",
+    "weights_sha256": "9b58d252978d3a5d33adcdaea95e113162daf5146e20a9e1b8aa47e21e48ed93",
+    "trained_bundle": true,
+    "control_weights": false,
+    "source_sha256": {
+      "audit_shared_character_training.py": "3eccbe9b38f6c4aa255ac96d5bbe3229631cf0d1eb9b9a0ee1a6665e5b52b323",
+      "character_handwriting_curriculum.py": "ac6aeac728da7c20e1d5a16248aca81f5522d6614629b07052679d288259e53e",
+      "shared_character_recognizer.py": "0498d8b7d61f3c8e4a4fdcd927a507464cfaf8156bf874a211484a5474697719",
+      "shared_character_training.py": "b0991c9fd0da42c736c4d5b36f9b64f23c9be94303a259867402e8e4cfd42012",
+      "shared_private_protocol.py": "49384fcd48140482e0efc9a190c56878dc8d6d367f3d85aa843f948a269152b1",
+      "train_shared_character_private.py": "d50a1ea8f2f13fefe2f3a5166ddc417aee4c42c2da1ef2ff90c2602d5788fc3a",
+      "training_checkpoint.py": "569a4ca5b0a81b29f678e066e8bfa5bd337fcc2785bbc4b5d0dfd3e8dea7a051",
+      "trocr_character_codec.py": "7ede21e57fd4985153570820873abf690cc48de2e677b38828753fbedfdbd5e2",
+      "trocr_training_loss.py": "cb722cf44ef61f883f0e932763fe5566ecbab1d05a4f69ad68e466334d188f86"
+    },
+    "weight_verification_sha256": "693a3c2542b1a0284fd3967827dbbb77621e0d155ac62f32f00f2eb3a623f2de"
+  },
+  "encoder_report_sha256": "6e19e7fe532e7d24d504f9e323e08d5c77e1ca07ceaeb1ebcc0f616369834dbc",
+  "encoder_parameters": 30432384,
+  "decoders": {
+    "english": {
+      "parameters": 38453760,
+      "classes": 64044,
+      "serialized_neural_population_byte_exact": true,
+      "graph_sha256": {
+        "decoder-step.json": "eaea1b3b5001549a6d0a4c4b1bd810371a82bdae2092452ac2f73214762ede95",
+        "decoder-step.pdiparams": "ec5ec71ad13c5305a1a95c8c2846c4de526571e58ce8df00fa5dbac529940a16"
+      }
+    },
+    "character": {
+      "parameters": 5956608,
+      "classes": 573,
+      "serialized_neural_population_byte_exact": true,
+      "graph_sha256": {
+        "decoder-step.json": "555d79e651df3b22fef37f7cc873c843678c135bbe39e69488f4ef3b2acb3e5d",
+        "decoder-step.pdiparams": "e35ba8d8743abdcf8b6cdb48a572a669cbf3e504d38d2a48827b68657e32aeff"
+      }
+    }
+  },
+  "total_graph_neural_values": 74842752,
+  "total_with_100k_router_reserve": 74942752,
+  "every_source_tensor_owned_once": true,
+  "no_duplicate_encoder": true,
+  "encoder_library": {
+    "path": "/home/ampixa/ocr/ocr-deva-ft/diagnostics/shared-private-continuation-native-r1-20261023/8000/encoder/build/ocr_source_selective_v1/ocr_source_selective_v1.so",
+    "sha256": "fce69529c6a3752247be1192a495de7d1c395f8493039ad0d17546a42f936c60"
+  },
+  "source_sha256": {
+    "export_shared_character_decoders.py": "117dfabb79c52c32778d8e31e6ecb702f07d83db12b23cd14c1e46ac92e96687",
+    "shared_character_paddle.py": "d3947e59ada6971a8ce03a40d94886bd188e147a36872ca255798e4974dc3521",
+    "trocr_paddle_graphs.py": "13422f5004d84084e02dcf5d7fbc17623e9d17f6e7d67be256c5a05b122a880b",
+    "shared_bundle_identity.py": "e696082c61d0f1e3dfc76130e3496d9964a517e2097b7cd15525de9cb267660f"
+  },
+  "decoder_numerical_qualified": false,
+  "generation_qualified": false,
+  "quality_measured": false,
+  "native_Paddle_qualified": false,
+  "model_promoted": false
+}

paddle/encoder-library.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fce69529c6a3752247be1192a495de7d1c395f8493039ad0d17546a42f936c60
+size 5347952

paddle/encoder/encoder.pdiparams

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2a632ecd78bd35441f0575b5073ebdfe2dd8edebff08c892b386205655700b0
+size 121743123

paddle/encoder/report.json

@@ -0,0 +1,408 @@
+{
+  "scope": "Export one actual conditional Paddle encoder and check all24 source fixtures.\n\nCount every saved neural value, including conditional factors. No source tensor\ninjection, separate full encoder, decoder, beam generation or quality claim.\nThe unchanged1e-4 bound applies to both branches; failure never changes it.\n",
+  "run_sha256": "42c4cf89a86fc3de900949348079f7b0c90d4e58ca2d28a105e326785a3eacd5",
+  "parameters": 30432384,
+  "whole_candidate_parameters": 74842752,
+  "saved_neural_population_byte_exact": true,
+  "single_shared_encoder": true,
+  "records": [
+    {
+      "id": "nep_hw:images/nep_2122.jpg",
+      "branch": "english",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 1.3473792765289545,
+      "file": "training-00.npz",
+      "sha256": "982a19949ee70841849fed810b856bdba91d010cf8102162d7afa8d9741586e0"
+    },
+    {
+      "id": "nep_hw:images/nep_2122.jpg",
+      "branch": "character",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.7976739667356014,
+      "file": "training-01.npz",
+      "sha256": "977d69d27433eb0e2e08a94c539743f0554c99e677395bf6e4829c0cde4701d0"
+    },
+    {
+      "id": "nep_hw:images/nep_769.jpg",
+      "branch": "english",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6514115799218416,
+      "file": "training-02.npz",
+      "sha256": "245197df27ae2a11e71b003c008dadfe25499f25358b5e322b03e3ba0cf6b0be"
+    },
+    {
+      "id": "nep_hw:images/nep_769.jpg",
+      "branch": "character",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6233616527169943,
+      "file": "training-03.npz",
+      "sha256": "0247fe245a5f3b90c6d2e754025a299273d8d6f08ae170debfc0353270a4f02a"
+    },
+    {
+      "id": "nep_hw:images/nep_3325.jpg",
+      "branch": "english",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6163665913045406,
+      "file": "training-04.npz",
+      "sha256": "aca9384656eecd2a391ba6bd5e34a3ee7a48cfec0c3c63969397f54903318daa"
+    },
+    {
+      "id": "nep_hw:images/nep_3325.jpg",
+      "branch": "character",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6238061562180519,
+      "file": "training-05.npz",
+      "sha256": "2b02130c4d892ff8c2277f3418d93ecd6374b4f85fb0bbaa4fd295c683c2dbfa"
+    },
+    {
+      "id": "nep_hw:images/nep_1370.jpg",
+      "branch": "english",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6203578859567642,
+      "file": "training-06.npz",
+      "sha256": "f7c672af41633ec3b4355cbe29244b7d01ca6ac5a766772248ba8f9d28ab35fc"
+    },
+    {
+      "id": "nep_hw:images/nep_1370.jpg",
+      "branch": "character",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6171446871012449,
+      "file": "training-07.npz",
+      "sha256": "a09266470b57ba8a3070f3db33b50556df611f768a70749a53a9bbb46d8d0d36"
+    },
+    {
+      "id": "iiit_hw:images/train-00001-of-00003_022703.jpg",
+      "branch": "english",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.634559752419591,
+      "file": "training-08.npz",
+      "sha256": "ac94b53c1ff084b2a48225947b1a4d0482c75c2ec2ee5699f312701030e05a34"
+    },
+    {
+      "id": "iiit_hw:images/train-00001-of-00003_022703.jpg",
+      "branch": "character",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6346047893166542,
+      "file": "training-09.npz",
+      "sha256": "d8c1cdd1b7df84f20010e10309e55bd14b38286cdb35414034caeb4dafb2d282"
+    },
+    {
+      "id": "iiit_hw:images/train-00002-of-00003_018536.jpg",
+      "branch": "english",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6370339300483465,
+      "file": "training-10.npz",
+      "sha256": "2d5c6487cc3a36874d248cd831be2a83cac615d904a24c99c407bcbeb16abef0"
+    },
+    {
+      "id": "iiit_hw:images/train-00002-of-00003_018536.jpg",
+      "branch": "character",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6340623814612627,
+      "file": "training-11.npz",
+      "sha256": "cd260e997b84766aa65f74e396adf55a08553fa43fa4289b6bd70d9af2e50f6e"
+    },
+    {
+      "id": "iiit_hw:images/train-00001-of-00003_003887.jpg",
+      "branch": "english",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6398916430771351,
+      "file": "training-12.npz",
+      "sha256": "c43f5416adf537f40b6ed304ee68477d089c58dc55675f559f5f53173f1e388e"
+    },
+    {
+      "id": "iiit_hw:images/train-00001-of-00003_003887.jpg",
+      "branch": "character",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6324335876852274,
+      "file": "training-13.npz",
+      "sha256": "8146a072dc32ea5c11b3b7f266f01de1fdc33b12459f170844d60d7ff2e35850"
+    },
+    {
+      "id": "iiit_hw:images/train-00002-of-00003_005824.jpg",
+      "branch": "english",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6408481299877167,
+      "file": "training-14.npz",
+      "sha256": "af9b1bc5a787894468fe9aa51855249b03ae86b2208ebf3ed74c92ff40973495"
+    },
+    {
+      "id": "iiit_hw:images/train-00002-of-00003_005824.jpg",
+      "branch": "character",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6478282399475574,
+      "file": "training-15.npz",
+      "sha256": "3d76153d3aa898f5d5553f4663d60bd67daf8447e0e296ca0a71be7b898da9bc"
+    },
+    {
+      "id": "en_hw:images/train_003860.jpg",
+      "branch": "english",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6384864822030067,
+      "file": "training-16.npz",
+      "sha256": "305e75cfd3943b5d63056465544c36f2eb3c19b15c19d765f7fb2aa9ea3aae94"
+    },
+    {
+      "id": "en_hw:images/train_003860.jpg",
+      "branch": "character",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6323059760034084,
+      "file": "training-17.npz",
+      "sha256": "fd35160165f846ffee5fe97706fbfc7583524c41fb71b4de72b721cc33ac0626"
+    },
+    {
+      "id": "en_hw:images/train_002056.jpg",
+      "branch": "english",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6381492856889963,
+      "file": "training-18.npz",
+      "sha256": "e4a3e8cabf0b12f93f1ee417803b799598f7b1187b071ac454a8b679ed05712d"
+    },
+    {
+      "id": "en_hw:images/train_002056.jpg",
+      "branch": "character",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6392509918659925,
+      "file": "training-19.npz",
+      "sha256": "a59b65b6edddbb1e222872dcf5cbbf6d6a1efda8dfa51d392e551e2e8f3bd829"
+    },
+    {
+      "id": "en_hw:images/train_002132.jpg",
+      "branch": "english",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6412029340863228,
+      "file": "training-20.npz",
+      "sha256": "cf4faeb40f9af3377682aed08a688efad81832be59c47459065cbfe175bde70e"
+    },
+    {
+      "id": "en_hw:images/train_002132.jpg",
+      "branch": "character",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.7008441369980574,
+      "file": "training-21.npz",
+      "sha256": "a37a7bdea06fd6e5790e2fefa8c9376ac91e7ae73ab6462ff16033a6ba49b0bf"
+    },
+    {
+      "id": "en_hw:images/train_001877.jpg",
+      "branch": "english",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.705221377313137,
+      "file": "training-22.npz",
+      "sha256": "bb4ce80d04227d3cac548e715b220783ce5bbf7323ec4501e500d91d75d01018"
+    },
+    {
+      "id": "en_hw:images/train_001877.jpg",
+      "branch": "character",
+      "max_errors": [
+        0.0,
+        0.0
+      ],
+      "byte_exact": [
+        true,
+        true
+      ],
+      "numerical_pass": true,
+      "inference_seconds": 0.6301882974803448,
+      "file": "training-23.npz",
+      "sha256": "e22b987cd122575465db84b22bfc0faadf8153b82f138f3967dcd5fccc491f6b"
+    }
+  ],
+  "numerical_limit": 0.0001,
+  "encoder_numerical_pass": true,
+  "graph_sha256": {
+    "encoder.json": "eddaf1699760460ba7f2bd7fbdee9b7de9166b581e5145d9dc3c8da177bd72b7",
+    "encoder.pdiparams": "b2a632ecd78bd35441f0575b5073ebdfe2dd8edebff08c892b386205655700b0"
+  },
+  "decoder_qualified": false,
+  "generation_qualified": false,
+  "quality_measured": false,
+  "native_Paddle_qualified": false,
+  "model_promoted": false,
+  "feature_controls": [],
+  "pooled_features_exposed": false,
+  "additional_learned_feature_parameters": 0
+}

paddle/encoder/run.json

@@ -0,0 +1,55 @@
+{
+  "scope": "Export one actual conditional Paddle encoder and check all24 source fixtures.\n\nCount every saved neural value, including conditional factors. No source tensor\ninjection, separate full encoder, decoder, beam generation or quality claim.\nThe unchanged1e-4 bound applies to both branches; failure never changes it.\n",
+  "PID": 722385,
+  "proc_starttime": "1579109973",
+  "bundle_identity": {
+    "candidate_manifest_sha256": "0510d21cd0bb7a249dafac167fd2e371ee63bd6f83a016a6c4f2e7c6bb843ede",
+    "weights_sha256": "9b58d252978d3a5d33adcdaea95e113162daf5146e20a9e1b8aa47e21e48ed93",
+    "trained_bundle": true,
+    "control_weights": false,
+    "source_sha256": {
+      "audit_shared_character_training.py": "3eccbe9b38f6c4aa255ac96d5bbe3229631cf0d1eb9b9a0ee1a6665e5b52b323",
+      "character_handwriting_curriculum.py": "ac6aeac728da7c20e1d5a16248aca81f5522d6614629b07052679d288259e53e",
+      "shared_character_recognizer.py": "0498d8b7d61f3c8e4a4fdcd927a507464cfaf8156bf874a211484a5474697719",
+      "shared_character_training.py": "b0991c9fd0da42c736c4d5b36f9b64f23c9be94303a259867402e8e4cfd42012",
+      "shared_private_protocol.py": "49384fcd48140482e0efc9a190c56878dc8d6d367f3d85aa843f948a269152b1",
+      "train_shared_character_private.py": "d50a1ea8f2f13fefe2f3a5166ddc417aee4c42c2da1ef2ff90c2602d5788fc3a",
+      "training_checkpoint.py": "569a4ca5b0a81b29f678e066e8bfa5bd337fcc2785bbc4b5d0dfd3e8dea7a051",
+      "trocr_character_codec.py": "7ede21e57fd4985153570820873abf690cc48de2e677b38828753fbedfdbd5e2",
+      "trocr_training_loss.py": "cb722cf44ef61f883f0e932763fe5566ecbab1d05a4f69ad68e466334d188f86"
+    },
+    "weight_verification_sha256": "693a3c2542b1a0284fd3967827dbbb77621e0d155ac62f32f00f2eb3a623f2de"
+  },
+  "candidate_manifest_sha256": "0510d21cd0bb7a249dafac167fd2e371ee63bd6f83a016a6c4f2e7c6bb843ede",
+  "controls_sha256": "31f8b5ec579311d64b6ebfbd1f8e63e076c3565d9f12609398c76a096041b30e",
+  "parameters_by_graph": {
+    "encoder": 30432384,
+    "english_decoder": 38453760,
+    "character_decoder": 5956608
+  },
+  "whole_candidate_parameters": 74842752,
+  "numerical_limit": 0.0001,
+  "source_sha256": {
+    "audit_shared_character_paddle_encoder.py": "69f730d2e5fa581c93d8ccc5c02b5ef79843bf0dffe6fb2d27f689c53348bbff",
+    "shared_character_paddle.py": "d3947e59ada6971a8ce03a40d94886bd188e147a36872ca255798e4974dc3521",
+    "trocr_paddle_graphs.py": "13422f5004d84084e02dcf5d7fbc17623e9d17f6e7d67be256c5a05b122a880b",
+    "trocr_paddle_compact_encoder.py": "f2298a4e5cae283d72efd709e237bc71fcffd527dbc08a709bc599ade9aaea85",
+    "trocr_paddle_source_order_encoder.py": "aa937bb6ba67fd86e6214aa3a0be626e88a0c469a121ad4e9b1f6d250d704737",
+    "trocr_paddle_norm_probe.py": "e991a2a36ac6f89f05906c3635037caa2666ad5cbbc48db9396747abe4ba526f",
+    "trocr_paddle_sleef_probe.py": "14dcf8341e6bba8ad98eb1381dd6628d961bd4499ca5b1595f0153a9ab37d97c",
+    "trocr_paddle_attention_probe.py": "c7405bb5602c4cde7f1a3c5c53a7e2fced0e2e2064cdb78911cdcfff1e66e06f",
+    "trocr_source_cpu_ops.cc": "63bb541df8f7a1f8635c946f2407aa0f449e63cfe480cbbe6c0b8d6fb2f5f4c2",
+    "trocr_selective_onednn_pass.cc": "ff304ef7e20f22cc1bf3c3199c6614bc3e870db640e569d465ca87c6704051b6",
+    "shared_bundle_identity.py": "e696082c61d0f1e3dfc76130e3496d9964a517e2097b7cd15525de9cb267660f"
+  },
+  "Paddle": "3.3.1",
+  "CPU_affinity": [
+    0,
+    1
+  ],
+  "custom_passes": [
+    "ocr_conv_gelu_onednn_v1"
+  ],
+  "GPU_supported": false,
+  "expose_features": false
+}

paddle/inference.yml

@@ -0,0 +1,33 @@
+Global:
+  model_name: TrOCR_shared_character_Paddle_research_rec
+Native:
+  format: shared-character-paddle-research-v1
+  parameters: 74842752
+  files_sha256:
+    candidate/bundle.json: 0510d21cd0bb7a249dafac167fd2e371ee63bd6f83a016a6c4f2e7c6bb843ede
+    candidate/character-assets/character_codec.json: 7bb65c86bc29b555f8da3af81b7023e370bdebe6ced8c84d104d55165c6ffb0a
+    candidate/character-assets/preprocessor_config.json: 36a945a7cc645688b9ef64dabae16979cf5f7c1c448569cc306694edc0598b9b
+    candidate/english-assets/sentencepiece.bpe.model: 6f5e2fefcf793761a76a6bfb8ad35489f9c203b25557673284b6d032f41043f4
+    candidate/english-assets/tokenizer_config.json: e098c98c6834fbcd3f3e75215e3e8b8868f0fe5bf07aef93d28e4f7196db293c
+    candidate/english-assets/preprocessor_config.json: eec0c686d2e2ea6d5887f87e1c6a34ef5bf1ba5b2ac98b8675e131da646fa1d2
+    candidate/english-assets/README.md: 7e0f183a2697956065aea4e3ec9ca718967be1b7afe446cff4dcdbc8a64fa52b
+    encoder/encoder.json: eddaf1699760460ba7f2bd7fbdee9b7de9166b581e5145d9dc3c8da177bd72b7
+    encoder/encoder.pdiparams: b2a632ecd78bd35441f0575b5073ebdfe2dd8edebff08c892b386205655700b0
+    encoder/report.json: 6e19e7fe532e7d24d504f9e323e08d5c77e1ca07ceaeb1ebcc0f616369834dbc
+    encoder/run.json: 42c4cf89a86fc3de900949348079f7b0c90d4e58ca2d28a105e326785a3eacd5
+    decoders/report.json: 24a40b59ddedd55f423035cb40def1e251633948a67802707d14cf3151a67ba8
+    decoders/english/decoder-step.json: eaea1b3b5001549a6d0a4c4b1bd810371a82bdae2092452ac2f73214762ede95
+    decoders/english/decoder-step.pdiparams: ec5ec71ad13c5305a1a95c8c2846c4de526571e58ce8df00fa5dbac529940a16
+    decoders/character/decoder-step.json: 555d79e651df3b22fef37f7cc873c843678c135bbe39e69488f4ef3b2acb3e5d
+    decoders/character/decoder-step.pdiparams: e35ba8d8743abdcf8b6cdb48a572a669cbf3e504d38d2a48827b68657e32aeff
+    encoder-library.so: fce69529c6a3752247be1192a495de7d1c395f8493039ad0d17546a42f936c60
+    code/paddlex_shared_character.py: 6a7b1d60aa337ba04514c9910e68f96d1bafdff383f3779e83f22a587707a3e3
+    code/shared_character_native.py: 394b45df79c5c4e84c2053b62b8d4080ef6a875034b6b16c661ac63257f56a30
+    code/trocr_character_codec.py: 7ede21e57fd4985153570820873abf690cc48de2e677b38828753fbedfdbd5e2
+    code/trocr_numpy_beam.py: e4f1d87113404dfc16fdddd78a1cdd6477a2f886d76cef99594fc537fc531028
+    code/trocr_source_cpu_ops.cc: 63bb541df8f7a1f8635c946f2407aa0f449e63cfe480cbbe6c0b8d6fb2f5f4c2
+    code/trocr_selective_onednn_pass.cc: ff304ef7e20f22cc1bf3c3199c6614bc3e870db640e569d465ca87c6704051b6
+    code/trocr_paddle_norm_probe.LICENSE: 19d6a3818fa4927890b59222656f155faae8f8817ebc9f40f779fdb6ab5a65d2
+    code/trocr_paddle_sleef_probe.LICENSE: 5f35cc826b3a0001293ed62e8ea95e5784f89f5e25300bf437d34be1d35e8a75
+    code/trocr_selective_onednn_pass.LICENSE: ff6311c56c1b22c9209c96cc8afc16da91c24ee55a15e14ebc46fe1960dbd856
+  profile: CPU two threads, FMA, Paddle3.3.1; explicit branch; research only

predict.py

@@ -0,0 +1,60 @@
+"""Recognize one cropped word or line with an explicit language branch."""
+import argparse
+import json
+from pathlib import Path
+import sys
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('image', type=Path)
+    parser.add_argument('--language', choices=['ne', 'hi', 'en'], required=True)
+    parser.add_argument('--backend', choices=['torch', 'paddle'], default='torch')
+    parser.add_argument('--model-dir', type=Path, default=Path(__file__).resolve().parent)
+    args = parser.parse_args()
+    root = args.model_dir.resolve()
+    branch = 'english' if args.language == 'en' else 'character'
+    if args.backend == 'torch':
+        import torch
+        from PIL import Image
+        from transformers import TrOCRProcessor, ViTImageProcessor
+        sys.path.insert(0, str(root / 'code'))
+        from shared_character_recognizer import SharedCharacterRecognizer
+        from trocr_character_codec import CharacterCodec
+        torch.set_num_threads(2)
+        torch.use_deterministic_algorithms(True)
+        folder = root / 'source-model'
+        model = SharedCharacterRecognizer.from_bundle(folder)
+        assets = folder / ('english-assets' if branch == 'english' else 'character-assets')
+        processor = (TrOCRProcessor.from_pretrained(assets, local_files_only=True, use_fast=False)
+                     if branch == 'english' else ViTImageProcessor.from_pretrained(assets, local_files_only=True))
+        with Image.open(args.image) as image:
+            pixels = processor(images=image.convert('RGB'), return_tensors='pt').pixel_values
+        generation = dict(do_sample=False, num_beams=10, max_new_tokens=256 if branch == 'english' else 257,
+                          min_length=0, length_penalty=1., early_stopping=False, use_cache=True)
+        if branch == 'character':
+            generation['suppress_tokens'] = [0, 1, 3]
+        with torch.inference_mode():
+            tokens = model.generate_pixels(pixels, branch, generation)[0].tolist()
+        if branch == 'english':
+            text = processor.batch_decode([tokens], skip_special_tokens=True)[0]
+        else:
+            spec = json.loads((assets / 'character_codec.json').read_text())
+            text = CharacterCodec(list(spec['character_ids'])).decode(tokens)
+        result = dict(text=text, tokens=tokens, token_limit_reached=len(tokens)-1 >= generation['max_new_tokens'] and tokens[-1] != 2)
+    else:
+        import paddle
+        from PIL import Image
+        sys.path.insert(0, str(root / 'paddle/code'))
+        from shared_character_native import SharedCharacterNative, preprocess_rgb
+        paddle.set_device('cpu')
+        folder = root / 'paddle'
+        model = SharedCharacterNative(folder / 'candidate', folder / 'encoder', folder / 'decoders',
+                                     allow_unqualified=True, library_path=folder / 'encoder-library.so')
+        with Image.open(args.image) as image:
+            result = model.predict_pixels(preprocess_rgb(image), branch)
+    print(json.dumps(dict(language=args.language, branch=branch, backend=args.backend, **result), ensure_ascii=False))
+
+
+if __name__ == '__main__':
+    main()

release-manifest.json

@@ -0,0 +1,15 @@
+{
+  "model_id": "ampixa/deva-ocr-75m-research",
+  "candidate_manifest_sha256": "0510d21cd0bb7a249dafac167fd2e371ee63bd6f83a016a6c4f2e7c6bb843ede",
+  "source_weights_sha256": "9b58d252978d3a5d33adcdaea95e113162daf5146e20a9e1b8aa47e21e48ed93",
+  "physical_neural_values": 74842752,
+  "source_model_format": "shared-character-private-continuation-source-v1",
+  "training_step": 8000,
+  "cumulative_training_step": 16000,
+  "original_native_bundle_sha256": "7f4091be58fc6398363d7a192fab73af6832a543a193efada11b91f188e7d120",
+  "native_graph_parameters_unchanged": true,
+  "complete_preexisting_native_source_comparison_images": 8796,
+  "packaged_conversion_predictions": false,
+  "packaged_training_data": false,
+  "published_use_scope": "non-commercial research and education; see LICENSE and TRAINING_DATA.md"
+}

requirements-paddle.txt

@@ -0,0 +1,7 @@
+paddlepaddle==3.3.1
+paddlex==3.7.1
+sentencepiece==0.2.1
+tokenizers==0.22.2
+Pillow>=10.4
+numpy>=2.0
+PyYAML>=6

requirements-source.txt

@@ -0,0 +1,6 @@
+torch>=2.6
+transformers==4.57.6
+safetensors>=0.6
+sentencepiece==0.2.1
+Pillow>=10.4
+numpy>=1.26

source-model/bundle.json

@@ -0,0 +1,2075 @@
+{
+  "format": "shared-character-private-continuation-source-v1",
+  "parameters": 74842752,
+  "corrections": [
+    {
+      "key": "t000",
+      "name": "embeddings.cls_token",
+      "shape": [
+        1,
+        1,
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t001",
+      "name": "embeddings.distillation_token",
+      "shape": [
+        1,
+        1,
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t002",
+      "name": "embeddings.patch_embeddings.projection.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t003",
+      "name": "embeddings.patch_embeddings.projection.weight",
+      "shape": [
+        384,
+        3,
+        16,
+        16
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        768
+      ]
+    },
+    {
+      "key": "t004",
+      "name": "embeddings.position_embeddings",
+      "shape": [
+        1,
+        578,
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t005",
+      "name": "encoder.layer.0.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t006",
+      "name": "encoder.layer.0.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t007",
+      "name": "encoder.layer.0.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t008",
+      "name": "encoder.layer.0.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t009",
+      "name": "encoder.layer.0.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t010",
+      "name": "encoder.layer.0.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t011",
+      "name": "encoder.layer.0.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t012",
+      "name": "encoder.layer.0.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t013",
+      "name": "encoder.layer.0.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t014",
+      "name": "encoder.layer.0.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t015",
+      "name": "encoder.layer.0.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t016",
+      "name": "encoder.layer.0.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t017",
+      "name": "encoder.layer.0.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t018",
+      "name": "encoder.layer.0.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t019",
+      "name": "encoder.layer.0.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t020",
+      "name": "encoder.layer.0.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t021",
+      "name": "encoder.layer.1.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t022",
+      "name": "encoder.layer.1.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t023",
+      "name": "encoder.layer.1.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t024",
+      "name": "encoder.layer.1.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t025",
+      "name": "encoder.layer.1.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t026",
+      "name": "encoder.layer.1.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t027",
+      "name": "encoder.layer.1.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t028",
+      "name": "encoder.layer.1.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t029",
+      "name": "encoder.layer.1.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t030",
+      "name": "encoder.layer.1.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t031",
+      "name": "encoder.layer.1.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t032",
+      "name": "encoder.layer.1.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t033",
+      "name": "encoder.layer.1.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t034",
+      "name": "encoder.layer.1.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t035",
+      "name": "encoder.layer.1.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t036",
+      "name": "encoder.layer.1.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t037",
+      "name": "encoder.layer.10.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t038",
+      "name": "encoder.layer.10.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t039",
+      "name": "encoder.layer.10.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t040",
+      "name": "encoder.layer.10.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t041",
+      "name": "encoder.layer.10.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t042",
+      "name": "encoder.layer.10.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t043",
+      "name": "encoder.layer.10.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t044",
+      "name": "encoder.layer.10.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t045",
+      "name": "encoder.layer.10.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t046",
+      "name": "encoder.layer.10.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t047",
+      "name": "encoder.layer.10.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t048",
+      "name": "encoder.layer.10.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t049",
+      "name": "encoder.layer.10.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t050",
+      "name": "encoder.layer.10.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t051",
+      "name": "encoder.layer.10.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t052",
+      "name": "encoder.layer.10.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t053",
+      "name": "encoder.layer.11.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t054",
+      "name": "encoder.layer.11.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t055",
+      "name": "encoder.layer.11.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t056",
+      "name": "encoder.layer.11.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t057",
+      "name": "encoder.layer.11.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t058",
+      "name": "encoder.layer.11.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t059",
+      "name": "encoder.layer.11.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t060",
+      "name": "encoder.layer.11.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t061",
+      "name": "encoder.layer.11.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t062",
+      "name": "encoder.layer.11.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t063",
+      "name": "encoder.layer.11.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t064",
+      "name": "encoder.layer.11.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t065",
+      "name": "encoder.layer.11.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t066",
+      "name": "encoder.layer.11.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t067",
+      "name": "encoder.layer.11.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t068",
+      "name": "encoder.layer.11.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t069",
+      "name": "encoder.layer.2.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t070",
+      "name": "encoder.layer.2.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t071",
+      "name": "encoder.layer.2.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t072",
+      "name": "encoder.layer.2.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t073",
+      "name": "encoder.layer.2.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t074",
+      "name": "encoder.layer.2.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t075",
+      "name": "encoder.layer.2.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t076",
+      "name": "encoder.layer.2.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t077",
+      "name": "encoder.layer.2.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t078",
+      "name": "encoder.layer.2.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t079",
+      "name": "encoder.layer.2.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t080",
+      "name": "encoder.layer.2.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t081",
+      "name": "encoder.layer.2.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t082",
+      "name": "encoder.layer.2.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t083",
+      "name": "encoder.layer.2.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t084",
+      "name": "encoder.layer.2.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t085",
+      "name": "encoder.layer.3.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t086",
+      "name": "encoder.layer.3.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t087",
+      "name": "encoder.layer.3.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t088",
+      "name": "encoder.layer.3.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t089",
+      "name": "encoder.layer.3.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t090",
+      "name": "encoder.layer.3.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t091",
+      "name": "encoder.layer.3.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t092",
+      "name": "encoder.layer.3.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t093",
+      "name": "encoder.layer.3.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t094",
+      "name": "encoder.layer.3.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t095",
+      "name": "encoder.layer.3.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t096",
+      "name": "encoder.layer.3.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t097",
+      "name": "encoder.layer.3.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t098",
+      "name": "encoder.layer.3.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t099",
+      "name": "encoder.layer.3.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t100",
+      "name": "encoder.layer.3.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t101",
+      "name": "encoder.layer.4.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t102",
+      "name": "encoder.layer.4.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t103",
+      "name": "encoder.layer.4.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t104",
+      "name": "encoder.layer.4.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t105",
+      "name": "encoder.layer.4.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t106",
+      "name": "encoder.layer.4.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t107",
+      "name": "encoder.layer.4.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t108",
+      "name": "encoder.layer.4.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t109",
+      "name": "encoder.layer.4.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t110",
+      "name": "encoder.layer.4.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t111",
+      "name": "encoder.layer.4.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t112",
+      "name": "encoder.layer.4.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t113",
+      "name": "encoder.layer.4.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t114",
+      "name": "encoder.layer.4.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t115",
+      "name": "encoder.layer.4.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t116",
+      "name": "encoder.layer.4.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t117",
+      "name": "encoder.layer.5.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t118",
+      "name": "encoder.layer.5.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t119",
+      "name": "encoder.layer.5.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t120",
+      "name": "encoder.layer.5.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t121",
+      "name": "encoder.layer.5.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t122",
+      "name": "encoder.layer.5.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t123",
+      "name": "encoder.layer.5.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t124",
+      "name": "encoder.layer.5.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t125",
+      "name": "encoder.layer.5.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t126",
+      "name": "encoder.layer.5.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t127",
+      "name": "encoder.layer.5.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t128",
+      "name": "encoder.layer.5.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t129",
+      "name": "encoder.layer.5.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t130",
+      "name": "encoder.layer.5.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t131",
+      "name": "encoder.layer.5.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t132",
+      "name": "encoder.layer.5.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t133",
+      "name": "encoder.layer.6.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t134",
+      "name": "encoder.layer.6.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t135",
+      "name": "encoder.layer.6.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t136",
+      "name": "encoder.layer.6.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t137",
+      "name": "encoder.layer.6.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t138",
+      "name": "encoder.layer.6.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t139",
+      "name": "encoder.layer.6.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t140",
+      "name": "encoder.layer.6.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t141",
+      "name": "encoder.layer.6.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t142",
+      "name": "encoder.layer.6.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t143",
+      "name": "encoder.layer.6.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t144",
+      "name": "encoder.layer.6.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t145",
+      "name": "encoder.layer.6.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t146",
+      "name": "encoder.layer.6.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t147",
+      "name": "encoder.layer.6.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t148",
+      "name": "encoder.layer.6.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t149",
+      "name": "encoder.layer.7.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t150",
+      "name": "encoder.layer.7.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t151",
+      "name": "encoder.layer.7.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t152",
+      "name": "encoder.layer.7.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t153",
+      "name": "encoder.layer.7.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t154",
+      "name": "encoder.layer.7.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t155",
+      "name": "encoder.layer.7.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t156",
+      "name": "encoder.layer.7.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t157",
+      "name": "encoder.layer.7.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t158",
+      "name": "encoder.layer.7.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t159",
+      "name": "encoder.layer.7.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t160",
+      "name": "encoder.layer.7.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t161",
+      "name": "encoder.layer.7.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t162",
+      "name": "encoder.layer.7.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t163",
+      "name": "encoder.layer.7.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t164",
+      "name": "encoder.layer.7.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t165",
+      "name": "encoder.layer.8.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t166",
+      "name": "encoder.layer.8.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t167",
+      "name": "encoder.layer.8.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t168",
+      "name": "encoder.layer.8.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t169",
+      "name": "encoder.layer.8.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t170",
+      "name": "encoder.layer.8.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t171",
+      "name": "encoder.layer.8.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t172",
+      "name": "encoder.layer.8.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t173",
+      "name": "encoder.layer.8.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t174",
+      "name": "encoder.layer.8.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t175",
+      "name": "encoder.layer.8.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t176",
+      "name": "encoder.layer.8.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t177",
+      "name": "encoder.layer.8.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t178",
+      "name": "encoder.layer.8.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t179",
+      "name": "encoder.layer.8.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t180",
+      "name": "encoder.layer.8.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t181",
+      "name": "encoder.layer.9.attention.attention.key.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t182",
+      "name": "encoder.layer.9.attention.attention.key.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t183",
+      "name": "encoder.layer.9.attention.attention.query.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t184",
+      "name": "encoder.layer.9.attention.attention.query.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t185",
+      "name": "encoder.layer.9.attention.attention.value.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t186",
+      "name": "encoder.layer.9.attention.attention.value.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t187",
+      "name": "encoder.layer.9.attention.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t188",
+      "name": "encoder.layer.9.attention.output.dense.weight",
+      "shape": [
+        384,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        384
+      ]
+    },
+    {
+      "key": "t189",
+      "name": "encoder.layer.9.intermediate.dense.bias",
+      "shape": [
+        1536
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t190",
+      "name": "encoder.layer.9.intermediate.dense.weight",
+      "shape": [
+        1536,
+        384
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        1536,
+        384
+      ]
+    },
+    {
+      "key": "t191",
+      "name": "encoder.layer.9.layernorm_after.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t192",
+      "name": "encoder.layer.9.layernorm_after.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t193",
+      "name": "encoder.layer.9.layernorm_before.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t194",
+      "name": "encoder.layer.9.layernorm_before.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t195",
+      "name": "encoder.layer.9.output.dense.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t196",
+      "name": "encoder.layer.9.output.dense.weight",
+      "shape": [
+        384,
+        1536
+      ],
+      "kind": "factor",
+      "rank": 71,
+      "matrix_shape": [
+        384,
+        1536
+      ]
+    },
+    {
+      "key": "t197",
+      "name": "layernorm.bias",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    },
+    {
+      "key": "t198",
+      "name": "layernorm.weight",
+      "shape": [
+        384
+      ],
+      "kind": "replacement"
+    }
+  ],
+  "files_sha256": {
+    "weights.safetensors": "9b58d252978d3a5d33adcdaea95e113162daf5146e20a9e1b8aa47e21e48ed93",
+    "english-assets/README.md": "7e0f183a2697956065aea4e3ec9ca718967be1b7afe446cff4dcdbc8a64fa52b",
+    "english-assets/config.json": "2b6e06edec319984eca99c932d940bfa1f308db3f74d6279440b906c373bf727",
+    "english-assets/generation_config.json": "41149cdcffec4d657f32dfcddd9b208037f01286c9e07945c724908c58ed0193",
+    "english-assets/preprocessor_config.json": "eec0c686d2e2ea6d5887f87e1c6a34ef5bf1ba5b2ac98b8675e131da646fa1d2",
+    "english-assets/sentencepiece.bpe.model": "6f5e2fefcf793761a76a6bfb8ad35489f9c203b25557673284b6d032f41043f4",
+    "english-assets/special_tokens_map.json": "58dc0a3d15e3bad4861bcf2dee770ea327344d185182d4c5d2bf253285e53a83",
+    "english-assets/tokenizer_config.json": "e098c98c6834fbcd3f3e75215e3e8b8868f0fe5bf07aef93d28e4f7196db293c",
+    "character-assets/config.json": "decbe7681cc6000bd5bf83226b295e7a051e8280bc12255961e011d2caadbfce",
+    "character-assets/generation_config.json": "9577d3da76b60ee97a70d1516ac144eba329a8a70dd0e96c6f721c910434dcff",
+    "character-assets/character_codec.json": "7bb65c86bc29b555f8da3af81b7023e370bdebe6ced8c84d104d55165c6ffb0a",
+    "character-assets/preprocessor_config.json": "36a945a7cc645688b9ef64dabae16979cf5f7c1c448569cc306694edc0598b9b",
+    "model-configs.json": "1c598da84fcbe26af738d5ef1c0ea04f80f965e7690b1089aeaa8c7977cd4359",
+    "english-generation.json": "cddf3683b66672c43fee2a6d966829eebf2a00c61d376d0eb30db972693210be",
+    "character-generation.json": "02b8abbca9b3f1e5fb8ea2057b0aea84e1f488dd5dd2ca52a75850170bc3b08b"
+  },
+  "uniform_rank": 71,
+  "routing_parameters_reserved": 100000,
+  "automatic_dispatch": false,
+  "scope": "Serial CPU source candidate only; no accuracy or native Paddle qualification",
+  "original_model_sha256": "1b83102cbc1520dee1c3937ac334da83da18cf4683d46ebd1bd4e93ebe584dd7",
+  "parent_bundle_sha256": "b9281e7da68d2b9c331697a9fb15da3a2366fb937fce00742d658987d924318e",
+  "training_run_sha256": "713dfd450e19d12523585d9973aad7a7236097e64206e035800bbdcfeacb21e3",
+  "training_step": 8000,
+  "training_chain": "5e79546a138135b739b6485f3af4d81d412baa44119e849ec307ac8d87dff4ec",
+  "initial_character_model_sha256": "fcc1346b8c6aa36ae04b8628deb5cd309aa09d4574f0ed9ee0ce4d124091174a",
+  "root_bundle_sha256": "a52a68c8950c7eada39215d52b7b1eaf893d23577297951019bce382689baf48",
+  "parent_training_step": 8000,
+  "cumulative_training_step": 16000,
+  "parent_training_run_sha256": "f33da3168d15a743a4e73a59a77a55f5c85e41ee675cf006a00c940b9a23c70b"
+}

source-model/character-assets/character_codec.json

@@ -0,0 +1,590 @@
+{
+  "version": 1,
+  "kind": "literal_NFC_characters",
+  "classes": 573,
+  "special_ids": {
+    "BOS_serialization": 0,
+    "PAD": 1,
+    "EOS_and_decoder_start": 2,
+    "UNK": 3
+  },
+  "character_ids": {
+    "!": 4,
+    "\"": 5,
+    "#": 6,
+    "$": 7,
+    "%": 8,
+    "&": 9,
+    "'": 10,
+    "(": 11,
+    ")": 12,
+    "*": 13,
+    "+": 14,
+    ",": 15,
+    "-": 16,
+    ".": 17,
+    "/": 18,
+    "0": 19,
+    "1": 20,
+    "2": 21,
+    "3": 22,
+    "4": 23,
+    "5": 24,
+    "6": 25,
+    "7": 26,
+    "8": 27,
+    "9": 28,
+    ":": 29,
+    ";": 30,
+    "<": 31,
+    "=": 32,
+    ">": 33,
+    "?": 34,
+    "A": 35,
+    "B": 36,
+    "C": 37,
+    "D": 38,
+    "E": 39,
+    "F": 40,
+    "G": 41,
+    "H": 42,
+    "I": 43,
+    "J": 44,
+    "K": 45,
+    "L": 46,
+    "M": 47,
+    "N": 48,
+    "O": 49,
+    "P": 50,
+    "Q": 51,
+    "R": 52,
+    "S": 53,
+    "T": 54,
+    "U": 55,
+    "V": 56,
+    "W": 57,
+    "X": 58,
+    "Y": 59,
+    "Z": 60,
+    "[": 61,
+    "]": 62,
+    "_": 63,
+    "`": 64,
+    "a": 65,
+    "b": 66,
+    "c": 67,
+    "d": 68,
+    "e": 69,
+    "f": 70,
+    "g": 71,
+    "h": 72,
+    "i": 73,
+    "j": 74,
+    "k": 75,
+    "l": 76,
+    "m": 77,
+    "n": 78,
+    "o": 79,
+    "p": 80,
+    "q": 81,
+    "r": 82,
+    "s": 83,
+    "t": 84,
+    "u": 85,
+    "v": 86,
+    "w": 87,
+    "x": 88,
+    "y": 89,
+    "z": 90,
+    "©": 91,
+    "‥": 92,
+    "{": 93,
+    "}": 94,
+    "\\": 95,
+    "|": 96,
+    "@": 97,
+    "^": 98,
+    "~": 99,
+    "÷": 100,
+    "∕": 101,
+    "∙": 102,
+    "⋅": 103,
+    "·": 104,
+    "±": 105,
+    "∓": 106,
+    "∩": 107,
+    "∪": 108,
+    "□": 109,
+    "←": 110,
+    "↔": 111,
+    "⇒": 112,
+    "⇐": 113,
+    "⇔": 114,
+    "∀": 115,
+    "∃": 116,
+    "∄": 117,
+    "∴": 118,
+    "∵": 119,
+    "∝": 120,
+    "∞": 121,
+    "⊥": 122,
+    "∟": 123,
+    "∠": 124,
+    "∡": 125,
+    "∢": 126,
+    "′": 127,
+    "″": 128,
+    "∥": 129,
+    "⊾": 130,
+    "⊿": 131,
+    "∂": 132,
+    "∫": 133,
+    "∬": 134,
+    "∭": 135,
+    "∮": 136,
+    "∯": 137,
+    "∰": 138,
+    "∑": 139,
+    "∏": 140,
+    "√": 141,
+    "∛": 142,
+    "∜": 143,
+    "∱": 144,
+    "∲": 145,
+    "∳": 146,
+    "∶": 147,
+    "∷": 148,
+    "∼": 149,
+    "®": 150,
+    "℉": 151,
+    "Ω": 152,
+    "℧": 153,
+    "Å": 154,
+    "⌀": 155,
+    "ℏ": 156,
+    "⅀": 157,
+    "⍺": 158,
+    "⍵": 159,
+    "¢": 160,
+    "€": 161,
+    "£": 162,
+    "¥": 163,
+    "₿": 164,
+    "Ⅰ": 165,
+    "Ⅱ": 166,
+    "Ⅲ": 167,
+    "Ⅳ": 168,
+    "Ⅴ": 169,
+    "Ⅵ": 170,
+    "Ⅶ": 171,
+    "Ⅷ": 172,
+    "Ⅸ": 173,
+    "Ⅹ": 174,
+    "Ⅺ": 175,
+    "Ⅻ": 176,
+    "ⅰ": 177,
+    "ⅱ": 178,
+    "ⅲ": 179,
+    "ⅳ": 180,
+    "ⅴ": 181,
+    "ⅵ": 182,
+    "ⅶ": 183,
+    "ⅷ": 184,
+    "ⅸ": 185,
+    "ⅹ": 186,
+    "ⅺ": 187,
+    "ⅻ": 188,
+    "➀": 189,
+    "➁": 190,
+    "➂": 191,
+    "➃": 192,
+    "➄": 193,
+    "➅": 194,
+    "➆": 195,
+    "➇": 196,
+    "➈": 197,
+    "➉": 198,
+    "➊": 199,
+    "➋": 200,
+    "➌": 201,
+    "➍": 202,
+    "➎": 203,
+    "➏": 204,
+    "➐": 205,
+    "➑": 206,
+    "➒": 207,
+    "➓": 208,
+    "❶": 209,
+    "❷": 210,
+    "❸": 211,
+    "❹": 212,
+    "❺": 213,
+    "❻": 214,
+    "❼": 215,
+    "❽": 216,
+    "❾": 217,
+    "❿": 218,
+    "①": 219,
+    "②": 220,
+    "③": 221,
+    "④": 222,
+    "⑤": 223,
+    "⑥": 224,
+    "⑦": 225,
+    "⑧": 226,
+    "⑨": 227,
+    "⑩": 228,
+    "●": 229,
+    "▶": 230,
+    "𝑢": 231,
+    "︽": 232,
+    "–": 233,
+    "﹥": 234,
+    "𝜓": 235,
+    "•": 236,
+    "∋": 237,
+    "ƒ": 238,
+    "⬆": 239,
+    "Ạ": 240,
+    "◀": 241,
+    "": 242,
+    "▫": 243,
+    "︾": 244,
+    "À": 245,
+    "Á": 246,
+    "Â": 247,
+    "Ã": 248,
+    "Ä": 249,
+    "Å": 250,
+    "Æ": 251,
+    "Ç": 252,
+    "È": 253,
+    "É": 254,
+    "Ê": 255,
+    "Ë": 256,
+    "Ì": 257,
+    "Í": 258,
+    "Î": 259,
+    "Ï": 260,
+    "Ð": 261,
+    "Ñ": 262,
+    "Ò": 263,
+    "Ó": 264,
+    "Ô": 265,
+    "Õ": 266,
+    "Ö": 267,
+    "Ø": 268,
+    "Ù": 269,
+    "Ú": 270,
+    "Û": 271,
+    "Ü": 272,
+    "Ý": 273,
+    "Þ": 274,
+    "à": 275,
+    "á": 276,
+    "â": 277,
+    "ã": 278,
+    "ä": 279,
+    "å": 280,
+    "æ": 281,
+    "ç": 282,
+    "è": 283,
+    "é": 284,
+    "ê": 285,
+    "ë": 286,
+    "ì": 287,
+    "í": 288,
+    "î": 289,
+    "ï": 290,
+    "ð": 291,
+    "ñ": 292,
+    "ò": 293,
+    "ó": 294,
+    "ô": 295,
+    "õ": 296,
+    "ö": 297,
+    "ø": 298,
+    "ù": 299,
+    "ú": 300,
+    "û": 301,
+    "ü": 302,
+    "ý": 303,
+    "þ": 304,
+    "ÿ": 305,
+    "¡": 306,
+    "¤": 307,
+    "¦": 308,
+    "§": 309,
+    "¨": 310,
+    "ª": 311,
+    "«": 312,
+    "¬": 313,
+    "¯": 314,
+    "°": 315,
+    "²": 316,
+    "³": 317,
+    "´": 318,
+    "µ": 319,
+    "¶": 320,
+    "¸": 321,
+    "¹": 322,
+    "º": 323,
+    "»": 324,
+    "¼": 325,
+    "½": 326,
+    "¾": 327,
+    "¿": 328,
+    "×": 329,
+    "‐": 330,
+    "‑": 331,
+    "‒": 332,
+    "—": 333,
+    "―": 334,
+    "‖": 335,
+    "‗": 336,
+    "‘": 337,
+    "’": 338,
+    "‚": 339,
+    "‛": 340,
+    "“": 341,
+    "”": 342,
+    "„": 343,
+    "‟": 344,
+    "†": 345,
+    "‡": 346,
+    "‣": 347,
+    "․": 348,
+    "…": 349,
+    "‧": 350,
+    "‰": 351,
+    "‴": 352,
+    "‵": 353,
+    "‶": 354,
+    "‷": 355,
+    "‸": 356,
+    "‹": 357,
+    "›": 358,
+    "※": 359,
+    "‼": 360,
+    "‽": 361,
+    "‾": 362,
+    "₤": 363,
+    "₡": 364,
+    "₹": 365,
+    "−": 366,
+    "∖": 367,
+    "∗": 368,
+    "≈": 369,
+    "≠": 370,
+    "≡": 371,
+    "≤": 372,
+    "≥": 373,
+    "⊂": 374,
+    "⊃": 375,
+    "↑": 376,
+    "→": 377,
+    "↓": 378,
+    "↕": 379,
+    "™": 380,
+    "Ω": 381,
+    "℮": 382,
+    "∆": 383,
+    "✓": 384,
+    "✗": 385,
+    "✘": 386,
+    "▪": 387,
+    "◼": 388,
+    "✔": 389,
+    "✕": 390,
+    "☑": 391,
+    "☒": 392,
+    "№": 393,
+    "₽": 394,
+    "₴": 395,
+    "Α": 396,
+    "α": 397,
+    "Β": 398,
+    "β": 399,
+    "Γ": 400,
+    "γ": 401,
+    "Δ": 402,
+    "δ": 403,
+    "Ε": 404,
+    "ε": 405,
+    "Ζ": 406,
+    "ζ": 407,
+    "Η": 408,
+    "η": 409,
+    "Θ": 410,
+    "θ": 411,
+    "Ι": 412,
+    "ι": 413,
+    "Κ": 414,
+    "κ": 415,
+    "Λ": 416,
+    "λ": 417,
+    "Μ": 418,
+    "μ": 419,
+    "Ν": 420,
+    "ν": 421,
+    "Ξ": 422,
+    "ξ": 423,
+    "Ο": 424,
+    "ο": 425,
+    "Π": 426,
+    "π": 427,
+    "Ρ": 428,
+    "ρ": 429,
+    "Σ": 430,
+    "σ": 431,
+    "ς": 432,
+    "Τ": 433,
+    "τ": 434,
+    "Υ": 435,
+    "υ": 436,
+    "Φ": 437,
+    "φ": 438,
+    "Χ": 439,
+    "χ": 440,
+    "Ψ": 441,
+    "ψ": 442,
+    "ω": 443,
+    "ऀ": 444,
+    "ँ": 445,
+    "ं": 446,
+    "ः": 447,
+    "ऄ": 448,
+    "अ": 449,
+    "आ": 450,
+    "इ": 451,
+    "ई": 452,
+    "उ": 453,
+    "ऊ": 454,
+    "ऋ": 455,
+    "ऌ": 456,
+    "ऍ": 457,
+    "ऎ": 458,
+    "ए": 459,
+    "ऐ": 460,
+    "ऑ": 461,
+    "ऒ": 462,
+    "ओ": 463,
+    "औ": 464,
+    "क": 465,
+    "ख": 466,
+    "ग": 467,
+    "घ": 468,
+    "ङ": 469,
+    "च": 470,
+    "छ": 471,
+    "ज": 472,
+    "झ": 473,
+    "ञ": 474,
+    "ट": 475,
+    "ठ": 476,
+    "ड": 477,
+    "ढ": 478,
+    "ण": 479,
+    "त": 480,
+    "थ": 481,
+    "द": 482,
+    "ध": 483,
+    "न": 484,
+    "ऩ": 485,
+    "प": 486,
+    "फ": 487,
+    "ब": 488,
+    "भ": 489,
+    "म": 490,
+    "य": 491,
+    "र": 492,
+    "ऱ": 493,
+    "ल": 494,
+    "ळ": 495,
+    "ऴ": 496,
+    "व": 497,
+    "श": 498,
+    "ष": 499,
+    "स": 500,
+    "ह": 501,
+    "ऺ": 502,
+    "ऻ": 503,
+    "़": 504,
+    "ऽ": 505,
+    "ा": 506,
+    "ि": 507,
+    "ी": 508,
+    "ु": 509,
+    "ू": 510,
+    "ृ": 511,
+    "ॄ": 512,
+    "ॅ": 513,
+    "ॆ": 514,
+    "े": 515,
+    "ै": 516,
+    "ॉ": 517,
+    "ॊ": 518,
+    "ो": 519,
+    "ौ": 520,
+    "्": 521,
+    "ॎ": 522,
+    "ॏ": 523,
+    "ॐ": 524,
+    "॑": 525,
+    "॒": 526,
+    "॓": 527,
+    "॔": 528,
+    "ॕ": 529,
+    "ॖ": 530,
+    "ॗ": 531,
+    "क़": 532,
+    "ख़": 533,
+    "ग़": 534,
+    "ज़": 535,
+    "ड़": 536,
+    "ढ़": 537,
+    "फ़": 538,
+    "य़": 539,
+    "ॠ": 540,
+    "ॡ": 541,
+    "ॢ": 542,
+    "ॣ": 543,
+    "।": 544,
+    "॥": 545,
+    "०": 546,
+    "१": 547,
+    "२": 548,
+    "३": 549,
+    "४": 550,
+    "५": 551,
+    "६": 552,
+    "७": 553,
+    "८": 554,
+    "९": 555,
+    "॰": 556,
+    "ॱ": 557,
+    "ॲ": 558,
+    "ॳ": 559,
+    "ॴ": 560,
+    "ॵ": 561,
+    "ॶ": 562,
+    "ॷ": 563,
+    "ॸ": 564,
+    "ॹ": 565,
+    "ॺ": 566,
+    "ॻ": 567,
+    "ॼ": 568,
+    "ॽ": 569,
+    "ॾ": 570,
+    "ॿ": 571,
+    " ": 572
+  },
+  "generation_suppressed_ids": [
+    0,
+    1,
+    3
+  ],
+  "training_target": "Drop serialization BOS0 only; retain every text codepoint and EOS2",
+  "unknown_output": "Decode3 as U+FFFD, never silently remove it",
+  "scope": "Experimental codec, not recognition quality or Paddle qualification."
+}

source-model/character-assets/config.json

@@ -0,0 +1,56 @@
+{
+  "architectures": [
+    "VisionEncoderDecoderModel"
+  ],
+  "decoder": {
+    "activation_dropout": 0.0,
+    "activation_function": "relu",
+    "add_cross_attention": true,
+    "attention_dropout": 0.0,
+    "classifier_dropout": 0.0,
+    "cross_attention_hidden_size": 384,
+    "d_model": 256,
+    "decoder_attention_heads": 8,
+    "decoder_ffn_dim": 1024,
+    "decoder_layerdrop": 0.0,
+    "decoder_layers": 6,
+    "dropout": 0.1,
+    "dtype": "float32",
+    "init_std": 0.02,
+    "is_decoder": true,
+    "layernorm_embedding": true,
+    "max_position_embeddings": 512,
+    "model_type": "trocr",
+    "scale_embedding": true,
+    "tie_word_embeddings": false,
+    "use_cache": false,
+    "use_learned_position_embeddings": true,
+    "vocab_size": 573
+  },
+  "dtype": "float32",
+  "encoder": {
+    "attention_probs_dropout_prob": 0.0,
+    "dtype": "float32",
+    "encoder_stride": 16,
+    "hidden_act": "gelu",
+    "hidden_dropout_prob": 0.0,
+    "hidden_size": 384,
+    "image_size": 384,
+    "initializer_range": 0.02,
+    "intermediate_size": 1536,
+    "layer_norm_eps": 1e-12,
+    "model_type": "deit",
+    "num_attention_heads": 6,
+    "num_channels": 3,
+    "num_hidden_layers": 12,
+    "patch_size": 16,
+    "pooler_act": "tanh",
+    "pooler_output_size": 384,
+    "qkv_bias": true
+  },
+  "eos_token_id": 2,
+  "is_encoder_decoder": true,
+  "model_type": "vision-encoder-decoder",
+  "tie_word_embeddings": false,
+  "transformers_version": "4.57.6"
+}

source-model/character-assets/generation_config.json

@@ -0,0 +1,15 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "decoder_start_token_id": 2,
+  "eos_token_id": 2,
+  "max_new_tokens": 257,
+  "pad_token_id": 1,
+  "suppress_tokens": [
+    0,
+    1,
+    3
+  ],
+  "transformers_version": "4.57.6",
+  "use_cache": false
+}

source-model/character-assets/preprocessor_config.json

@@ -0,0 +1,28 @@
+{
+  "crop_size": {
+    "height": 224,
+    "width": 224
+  },
+  "do_center_crop": false,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "image_processor_type": "DeiTImageProcessor",
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "processor_class": "TrOCRProcessor",
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "height": 384,
+    "width": 384
+  }
+}

source-model/character-generation.json

@@ -0,0 +1,73 @@
+{
+  "max_length": 20,
+  "max_new_tokens": 257,
+  "min_length": 0,
+  "min_new_tokens": null,
+  "early_stopping": false,
+  "max_time": null,
+  "stop_strings": null,
+  "do_sample": false,
+  "num_beams": 1,
+  "use_cache": false,
+  "cache_implementation": null,
+  "cache_config": null,
+  "return_legacy_cache": null,
+  "prefill_chunk_size": null,
+  "temperature": 1.0,
+  "top_k": 50,
+  "top_p": 1.0,
+  "min_p": null,
+  "typical_p": 1.0,
+  "epsilon_cutoff": 0.0,
+  "eta_cutoff": 0.0,
+  "repetition_penalty": 1.0,
+  "encoder_repetition_penalty": 1.0,
+  "length_penalty": 1.0,
+  "no_repeat_ngram_size": 0,
+  "bad_words_ids": null,
+  "renormalize_logits": false,
+  "forced_bos_token_id": null,
+  "forced_eos_token_id": null,
+  "remove_invalid_values": false,
+  "exponential_decay_length_penalty": null,
+  "suppress_tokens": [
+    0,
+    1,
+    3
+  ],
+  "begin_suppress_tokens": null,
+  "sequence_bias": null,
+  "token_healing": false,
+  "guidance_scale": null,
+  "watermarking_config": null,
+  "num_return_sequences": 1,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "output_scores": false,
+  "output_logits": null,
+  "return_dict_in_generate": false,
+  "pad_token_id": 1,
+  "bos_token_id": 0,
+  "eos_token_id": 2,
+  "encoder_no_repeat_ngram_size": 0,
+  "decoder_start_token_id": 2,
+  "is_assistant": false,
+  "num_assistant_tokens": 20,
+  "num_assistant_tokens_schedule": "constant",
+  "assistant_confidence_threshold": 0.4,
+  "prompt_lookup_num_tokens": null,
+  "max_matching_ngram_size": null,
+  "assistant_early_exit": null,
+  "assistant_lookbehind": 10,
+  "target_lookbehind": 10,
+  "disable_compile": false,
+  "low_memory": null,
+  "penalty_alpha": null,
+  "dola_layers": null,
+  "diversity_penalty": 0.0,
+  "num_beam_groups": 1,
+  "constraints": null,
+  "force_words_ids": null,
+  "_from_model_config": true,
+  "transformers_version": "4.57.6"
+}

source-model/english-assets/README.md

@@ -0,0 +1,61 @@
+---
+tags:
+- trocr
+- image-to-text
+widget:
+- src: https://fki.tic.heia-fr.ch/static/img/a01-122-02.jpg
+  example_title: Note 1
+- src: https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcSoolxi9yWGAT5SLZShv8vVd0bz47UWRzQC19fDTeE8GmGv_Rn-PCF1pP1rrUx8kOjA4gg&usqp=CAU
+  example_title: Note 2
+- src: https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcRNYtTuSBpZPV_nkBYPMFwVVD9asZOPgHww4epu9EqWgDmXW--sE2o8og40ZfDGo87j5w&usqp=CAU
+  example_title: Note 3
+---
+
+# TrOCR (small-sized model, fine-tuned on IAM) 
+
+TrOCR model fine-tuned on the [IAM dataset](https://fki.tic.heia-fr.ch/databases/iam-handwriting-database). It was introduced in the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Li et al. and first released in [this repository](https://github.com/microsoft/unilm/tree/master/trocr). 
+
+
+## Model description
+
+The TrOCR model is an encoder-decoder model, consisting of an image Transformer as encoder, and a text Transformer as decoder. The image encoder was initialized from the weights of DeiT, while the text decoder was initialized from the weights of UniLM.
+
+Images are presented to the model as a sequence of fixed-size patches (resolution 16x16), which are linearly embedded. One also adds absolute position embeddings before feeding the sequence to the layers of the Transformer encoder. Next, the Transformer text decoder autoregressively generates tokens.
+
+## Intended uses & limitations
+
+You can use the raw model for optical character recognition (OCR) on single text-line images. See the [model hub](https://huggingface.co/models?search=microsoft/trocr) to look for fine-tuned versions on a task that interests you.
+
+### How to use
+
+Here is how to use this model in PyTorch:
+
+```python
+from transformers import TrOCRProcessor, VisionEncoderDecoderModel
+from PIL import Image
+import requests
+
+# load image from the IAM database
+url = 'https://fki.tic.heia-fr.ch/static/img/a01-122-02-00.jpg'
+image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
+
+processor = TrOCRProcessor.from_pretrained('microsoft/trocr-small-handwritten')
+model = VisionEncoderDecoderModel.from_pretrained('microsoft/trocr-small-handwritten')
+pixel_values = processor(images=image, return_tensors="pt").pixel_values
+
+generated_ids = model.generate(pixel_values)
+generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+```
+
+### BibTeX entry and citation info
+
+```bibtex
+@misc{li2021trocr,
+      title={TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models}, 
+      author={Minghao Li and Tengchao Lv and Lei Cui and Yijuan Lu and Dinei Florencio and Cha Zhang and Zhoujun Li and Furu Wei},
+      year={2021},
+      eprint={2109.10282},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+```

source-model/english-assets/config.json

@@ -0,0 +1,159 @@
+{
+  "architectures": [
+    "VisionEncoderDecoderModel"
+  ],
+  "decoder": {
+    "_name_or_path": "",
+    "activation_dropout": 0.0,
+    "activation_function": "relu",
+    "add_cross_attention": true,
+    "architectures": null,
+    "attention_dropout": 0.0,
+    "bad_words_ids": null,
+    "bos_token_id": 0,
+    "chunk_size_feed_forward": 0,
+    "classifier_dropout": 0.0,
+    "cross_attention_hidden_size": 384,
+    "d_model": 256,
+    "decoder_attention_heads": 8,
+    "decoder_ffn_dim": 1024,
+    "decoder_layerdrop": 0.0,
+    "decoder_layers": 6,
+    "decoder_start_token_id": 2,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "dropout": 0.1,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 2,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "init_std": 0.02,
+    "is_decoder": true,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layernorm_embedding": true,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "max_position_embeddings": 512,
+    "min_length": 0,
+    "model_type": "trocr",
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": 1,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "scale_embedding": true,
+    "sep_token_id": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": false,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": null,
+    "torchscript": false,
+    "transformers_version": "4.14.1",
+    "use_bfloat16": false,
+    "use_cache": false,
+    "use_learned_position_embeddings": true,
+    "vocab_size": 64044
+  },
+  "encoder": {
+    "_name_or_path": "",
+    "add_cross_attention": false,
+    "architectures": null,
+    "attention_probs_dropout_prob": 0.0,
+    "bad_words_ids": null,
+    "bos_token_id": null,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "hidden_act": "gelu",
+    "hidden_dropout_prob": 0.0,
+    "hidden_size": 384,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "image_size": 384,
+    "initializer_range": 0.02,
+    "intermediate_size": 1536,
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layer_norm_eps": 1e-12,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "deit",
+    "no_repeat_ngram_size": 0,
+    "num_attention_heads": 6,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_channels": 3,
+    "num_hidden_layers": 12,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "patch_size": 16,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "qkv_bias": true,
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sep_token_id": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": null,
+    "torchscript": false,
+    "transformers_version": "4.14.1",
+    "use_bfloat16": false
+  },
+  "eos_token_id": 2,
+  "is_encoder_decoder": true,
+  "model_type": "vision-encoder-decoder",
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "transformers_version": null
+}

source-model/english-assets/generation_config.json

@@ -0,0 +1,9 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "decoder_start_token_id": 2,
+  "eos_token_id": 2,
+  "pad_token_id": 1,
+  "transformers_version": "4.27.0.dev0",
+  "use_cache": false
+}