openocr-mlt-cuda 0.4.0__tar.gz

openocr_mlt_cuda-0.4.0/MANIFEST.in

@@ -0,0 +1,7 @@
+include README.md THIRD_PARTY_NOTICES.md
+recursive-include examples *.py *.png
+recursive-include tools *.py
+recursive-include constraints *.txt *.lock
+prune tests
+recursive-include docs *.md
+recursive-include docs *.json

openocr_mlt_cuda-0.4.0/PKG-INFO

@@ -0,0 +1,120 @@
+Metadata-Version: 2.4
+Name: openocr-mlt-cuda
+Version: 0.4.0
+Summary: Offline multilingual OCR with ONNX Prefill/Decode recognition and PP-OCRv6 medium detection
+Maintainer: ericleyu
+Project-URL: Models, https://modelscope.cn/models/Strawberry015/openocr-mlt-models
+Project-URL: Model mirror, https://huggingface.co/ericleyu/openocr-mlt-models
+Project-URL: Upstream recognition, https://github.com/YesianRohn/ScriptMoE
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+Requires-Dist: numpy<2.3,>=2.2
+Requires-Dist: Pillow<12.4,>=12.3
+Requires-Dist: opencv-python-headless<5.1,>=5.0
+Requires-Dist: onnxruntime-gpu==1.20.0
+Requires-Dist: pyclipper==1.4.0
+Requires-Dist: nvidia-cuda-runtime-cu12==12.4.127
+Requires-Dist: nvidia-cublas-cu12==12.4.5.8
+Requires-Dist: nvidia-cudnn-cu12==9.1.0.70
+Requires-Dist: nvidia-cuda-nvrtc-cu12==12.4.127
+Requires-Dist: nvidia-cufft-cu12==11.2.1.3
+Requires-Dist: nvidia-curand-cu12==10.3.5.147
+Provides-Extra: torch
+Requires-Dist: torch<2.6,>=2.5; extra == "torch"
+Requires-Dist: torchvision<0.21,>=0.20; extra == "torch"
+Requires-Dist: safetensors<0.9,>=0.8; extra == "torch"
+
+# openocr-mlt
+
+Multilingual OCR with ScriptMoE ONNX Prefill/Decode recognition and PP-OCRv6 medium text detection. Run whole-page OCR, detection, or cropped text recognition locally.
+
+## Installation
+
+Validated on Linux x86_64 with Python 3.10.12. Create one virtual environment for the installation you want to use.
+
+**CPU**
+
+~~~bash
+python3.10 -m venv .venv-ocr-cpu
+source .venv-ocr-cpu/bin/activate
+python -m pip install openocr-mlt
+~~~
+
+**NVIDIA CUDA**
+
+~~~bash
+python3.10 -m venv .venv-ocr-cuda
+source .venv-ocr-cuda/bin/activate
+python -m pip install openocr-mlt-cuda
+~~~
+
+Both packages provide the same Python import and CLI. Install only one package per environment. The CUDA package requires a compatible NVIDIA driver and installs CUDA 12 and cuDNN 9 user-space libraries. ONNX inference runs without PyTorch.
+
+### Upgrading from 0.3.0
+
+- A 0.3.0 CPU installation can be upgraded with `python -m pip install --upgrade openocr-mlt`.
+- For a 0.3.0 CUDA installation, create a new environment and install `openocr-mlt-cuda` as shown above. The previous `openocr-mlt[cuda]` command applies to 0.3.0. A shared or explicitly configured model cache can be reused.
+
+## Python API
+
+~~~python
+from openocr_mlt import OpenOCRMLT
+
+with OpenOCRMLT(task="ocr") as ocr:
+    page = ocr(image_path="page.jpg")
+    print(page.text)
+
+with OpenOCRMLT(task="det", device="cpu") as detector:
+    boxes = detector(image_path="page.jpg")
+    print([box.points for box in boxes])
+
+with OpenOCRMLT(task="rec", device="cpu") as recognizer:
+    lines = recognizer(image_path="crop.jpg")
+    print([(line.text, line.score) for line in lines])
+~~~
+
+Pass an existing image path or use the image argument with a PIL image or a uint8 NumPy array. Arrays use BGR by default. Pass color_order="RGB" for RGB arrays. The default device is `"auto"`. The CPU installation uses CPU. The CUDA installation uses CUDA when it can initialize the selected models and uses CPU when CUDA is unavailable. Use `device="cpu"` or `device="cuda:0"` to request a specific device. Explicit model directories and the original methods remain supported.
+
+## Model preparation
+
+The first task call downloads the FP32 model bundle if it is absent from the cache. The bundle is hosted on [ModelScope](https://modelscope.cn/models/Strawberry015/openocr-mlt-models/files). Hugging Face and its mirror are fallback sources. The downloader verifies the archive and all 40 model files before installing the cache. The default cache is `$XDG_CACHE_HOME/openocr-mlt` or `~/.cache/openocr-mlt`. The next task can reuse it offline.
+
+Import a local archive in an offline environment with the following command.
+
+~~~bash
+python -m openocr_mlt fetch-models \
+  --source ./scriptmoe-medium-fp32-20261101-v1.tar.gz --offline
+~~~
+
+Set `offline=True` on the Python task to require local assets. The fetch-models command prints the verified recognition and detector directories.
+
+## Command line
+
+~~~bash
+python -m openocr_mlt fetch-models --cache-dir ./models
+python -m openocr_mlt ocr ./page.jpg \
+  --model-dir ./models/scriptmoe-medium-fp32-20261101-v1/recognition \
+  --det-model-dir ./models/scriptmoe-medium-fp32-20261101-v1/ppocrv6-medium-det-bce-v1 \
+  --device cpu
+~~~
+
+The CLI also supports detect and recognize. Run python -m openocr_mlt --help for the available options.
+
+## Model Evaluation
+
+The FP32 graphs were evaluated with openocr-mlt 0.3.0.dev14. The model bundle is unchanged.
+
+| Task | Data | Result |
+|---|---|---|
+| CUDA recognition | TextMuSS, 10,899 crops | Official normalized weighted accuracy 85.0812%; ten-group macro accuracy 81.0928% |
+| Whole-page OCR, CPU and CUDA | Fixed CC-OCR subset, 200 pages across ten languages, development/held-out split 100/100 | Held-out text F1: macro 72.5849%, micro 79.4545% |
+
+Those model evaluations used Linux x86_64, Python 3.10.12, ONNX Runtime 1.20.0, an Intel Xeon Gold 6226R CPU, and an NVIDIA RTX 3090 GPU. One CPU/CUDA recognition whitespace difference was observed on TextMuSS. Additional platforms, hardware and quantized execution remain future work.
+
+## Sources
+
+- [ScriptMoE](https://github.com/YesianRohn/ScriptMoE): multilingual recognition model and decoding design.
+- [ppocrv6_onnx](https://github.com/AIwork4me/ppocrv6_onnx): detection preprocessing and postprocessing reference.
+- PP-OCRv6 medium: detector distributed through Paddle's official model storage.
+
+Component attribution and licensing information are included in the package's THIRD_PARTY_NOTICES.txt and in the model repository's [THIRD_PARTY_NOTICES.md](https://huggingface.co/ericleyu/openocr-mlt-models/blob/7ef39160a893fc7e605e146c1544ba64c1a57f43/THIRD_PARTY_NOTICES.md).

openocr_mlt_cuda-0.4.0/README.md

@@ -0,0 +1,94 @@
+# openocr-mlt
+
+Multilingual OCR with ScriptMoE ONNX Prefill/Decode recognition and PP-OCRv6 medium text detection. Run whole-page OCR, detection, or cropped text recognition locally.
+
+## Installation
+
+Validated on Linux x86_64 with Python 3.10.12. Create one virtual environment for the installation you want to use.
+
+**CPU**
+
+~~~bash
+python3.10 -m venv .venv-ocr-cpu
+source .venv-ocr-cpu/bin/activate
+python -m pip install openocr-mlt
+~~~
+
+**NVIDIA CUDA**
+
+~~~bash
+python3.10 -m venv .venv-ocr-cuda
+source .venv-ocr-cuda/bin/activate
+python -m pip install openocr-mlt-cuda
+~~~
+
+Both packages provide the same Python import and CLI. Install only one package per environment. The CUDA package requires a compatible NVIDIA driver and installs CUDA 12 and cuDNN 9 user-space libraries. ONNX inference runs without PyTorch.
+
+### Upgrading from 0.3.0
+
+- A 0.3.0 CPU installation can be upgraded with `python -m pip install --upgrade openocr-mlt`.
+- For a 0.3.0 CUDA installation, create a new environment and install `openocr-mlt-cuda` as shown above. The previous `openocr-mlt[cuda]` command applies to 0.3.0. A shared or explicitly configured model cache can be reused.
+
+## Python API
+
+~~~python
+from openocr_mlt import OpenOCRMLT
+
+with OpenOCRMLT(task="ocr") as ocr:
+    page = ocr(image_path="page.jpg")
+    print(page.text)
+
+with OpenOCRMLT(task="det", device="cpu") as detector:
+    boxes = detector(image_path="page.jpg")
+    print([box.points for box in boxes])
+
+with OpenOCRMLT(task="rec", device="cpu") as recognizer:
+    lines = recognizer(image_path="crop.jpg")
+    print([(line.text, line.score) for line in lines])
+~~~
+
+Pass an existing image path or use the image argument with a PIL image or a uint8 NumPy array. Arrays use BGR by default. Pass color_order="RGB" for RGB arrays. The default device is `"auto"`. The CPU installation uses CPU. The CUDA installation uses CUDA when it can initialize the selected models and uses CPU when CUDA is unavailable. Use `device="cpu"` or `device="cuda:0"` to request a specific device. Explicit model directories and the original methods remain supported.
+
+## Model preparation
+
+The first task call downloads the FP32 model bundle if it is absent from the cache. The bundle is hosted on [ModelScope](https://modelscope.cn/models/Strawberry015/openocr-mlt-models/files). Hugging Face and its mirror are fallback sources. The downloader verifies the archive and all 40 model files before installing the cache. The default cache is `$XDG_CACHE_HOME/openocr-mlt` or `~/.cache/openocr-mlt`. The next task can reuse it offline.
+
+Import a local archive in an offline environment with the following command.
+
+~~~bash
+python -m openocr_mlt fetch-models \
+  --source ./scriptmoe-medium-fp32-20261101-v1.tar.gz --offline
+~~~
+
+Set `offline=True` on the Python task to require local assets. The fetch-models command prints the verified recognition and detector directories.
+
+## Command line
+
+~~~bash
+python -m openocr_mlt fetch-models --cache-dir ./models
+python -m openocr_mlt ocr ./page.jpg \
+  --model-dir ./models/scriptmoe-medium-fp32-20261101-v1/recognition \
+  --det-model-dir ./models/scriptmoe-medium-fp32-20261101-v1/ppocrv6-medium-det-bce-v1 \
+  --device cpu
+~~~
+
+The CLI also supports detect and recognize. Run python -m openocr_mlt --help for the available options.
+
+## Model Evaluation
+
+The FP32 graphs were evaluated with openocr-mlt 0.3.0.dev14. The model bundle is unchanged.
+
+| Task | Data | Result |
+|---|---|---|
+| CUDA recognition | TextMuSS, 10,899 crops | Official normalized weighted accuracy 85.0812%; ten-group macro accuracy 81.0928% |
+| Whole-page OCR, CPU and CUDA | Fixed CC-OCR subset, 200 pages across ten languages, development/held-out split 100/100 | Held-out text F1: macro 72.5849%, micro 79.4545% |
+
+Those model evaluations used Linux x86_64, Python 3.10.12, ONNX Runtime 1.20.0, an Intel Xeon Gold 6226R CPU, and an NVIDIA RTX 3090 GPU. One CPU/CUDA recognition whitespace difference was observed on TextMuSS. Additional platforms, hardware and quantized execution remain future work.
+
+## Sources
+
+- [ScriptMoE](https://github.com/YesianRohn/ScriptMoE): multilingual recognition model and decoding design.
+- [ppocrv6_onnx](https://github.com/AIwork4me/ppocrv6_onnx): detection preprocessing and postprocessing reference.
+- PP-OCRv6 medium: detector distributed through Paddle's official model storage.
+
+Component attribution and licensing information are included in the package's THIRD_PARTY_NOTICES.txt and in the model repository's [THIRD_PARTY_NOTICES.md](https://huggingface.co/ericleyu/openocr-mlt-models/blob/7ef39160a893fc7e605e146c1544ba64c1a57f43/THIRD_PARTY_NOTICES.md).

openocr_mlt_cuda-0.4.0/THIRD_PARTY_NOTICES.md

@@ -0,0 +1,26 @@
+# Third-Party Model Sources and Licensing
+
+## ScriptMoE Recognition
+
+Recognition models are derived from [ScriptMoE](https://github.com/YesianRohn/ScriptMoE) at upstream revision `5dde99867aaee6b4af3e75a3020b225eab17ac53`. ONNX conversion preserves the character dictionary and autoregressive decoding rules and uses KV caching and optimized MoE execution.
+
+| Artifact | SHA256 |
+|---|---|
+| Original recognition weights | 60c74a924404b652c3a9550ec370c74d78850151785341729bdac268c014650d |
+| Character dictionary | 7b4eb2ca5fd6c9b6c6ecf423f69d35d8da48088697303e53624af196ee9ab6c0 |
+
+## PP-OCRv6 Medium Detection
+
+The detector is distributed by Paddle as [PP-OCRv6_medium_det_onnx.tar](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/tmp/PP-OCRv6_medium_det_onnx.tar).
+
+| Property | Value |
+|---|---|
+| File | inference.onnx |
+| Size | 62,026,563 bytes |
+| SHA256 | a0d14865d3a297970899b5c246f0e47c43879bb406c3a1fa4e505aba84b134e4 |
+
+The accompanying Python package adapts detection preprocessing and postprocessing from [ppocrv6_onnx](https://github.com/AIwork4me/ppocrv6_onnx), revision `d638eb0bb5d8222522abcc1bae4394cbeeb636f0`, under its MIT license. The package includes the license text for that code.
+
+## Licensing Status
+
+A unified redistribution license has not been specified for this combined model bundle. The MIT license of the detection adapter applies to that code; model weights have separate licensing requirements.

openocr_mlt_cuda-0.4.0/docs/INSTALL.md

@@ -0,0 +1,3 @@
+# Installation
+
+See the [installation and usage guide](../README.md#installation) for separate CPU and CUDA installations, Python requirements, and runtime selection.

openocr_mlt_cuda-0.4.0/docs/MODELS.md

@@ -0,0 +1,3 @@
+# Model preparation
+
+See the [model preparation guide](../README.md#model-preparation) for first-use download, the verified model cache, and offline import.

openocr_mlt_cuda-0.4.0/docs/VALIDATION.md

@@ -0,0 +1,5 @@
+# Validation
+
+The FP32 model and inference implementation in this release use the same pipeline evaluated with version 0.3.0.dev14. See the [evaluation summary](verification/fp32-v14-summary.json) for the historical TextMuSS and CC-OCR quality and performance results.
+
+The [installation record](verification/acceptance-summary.json) describes the earlier 0.3.0.dev15 trial. It covers separate Linux x86_64 CPU/CUDA installations, model import, cache reuse, whole-page OCR and cropped recognition.

openocr_mlt_cuda-0.4.0/docs/verification/acceptance-summary.json

@@ -0,0 +1,48 @@
+{
+  "version": "0.3.0.dev15",
+  "date": "2026-11-01",
+  "passed": true,
+  "public_release": false,
+  "public_model_url_configured": false,
+  "cpu_cuda_clean_installs": [
+    {
+      "kind": "cpu",
+      "python": "3.10.12 (main, Aug 31 2026, 10:18:17) [GCC 11.4.0]",
+      "packages": 13,
+      "version": "0.3.0.dev15",
+      "pip_check": true,
+      "model_files": 40,
+      "cache_reused": true,
+      "page_text": "Security",
+      "crop_text": "Security",
+      "actual_rec_device": "cpu",
+      "actual_det_device": "cpu",
+      "invalid_device_exit": 2
+    },
+    {
+      "kind": "cuda",
+      "python": "3.10.12 (main, Aug 31 2026, 10:18:17) [GCC 11.4.0]",
+      "packages": 19,
+      "version": "0.3.0.dev15",
+      "pip_check": true,
+      "model_files": 40,
+      "cache_reused": true,
+      "page_text": "Security",
+      "crop_text": "Security",
+      "actual_rec_device": "cuda:0",
+      "actual_det_device": "cuda:0",
+      "invalid_device_exit": 2
+    }
+  ],
+  "unit_boundary_tests": 12,
+  "https_test": "isolated transport/redirect tests; no public endpoint configured",
+  "sdist_runtime_files_identical": 31,
+  "hash_lock_dry_run_resolved": {
+    "cpu": 13,
+    "cuda": 19
+  },
+  "inference_modules_and_models_identical_to": "0.3.0.dev14",
+  "new_full_quality_runs": 0,
+  "new_performance_runs": 0,
+  "prior_evidence": "fp32-v14-summary.json"
+}

openocr_mlt_cuda-0.4.0/docs/verification/fp32-v14-summary.json

@@ -0,0 +1,268 @@
+{
+  "version": "0.3.0.dev14",
+  "date": "2026-11-01",
+  "recognition_model_id": "scriptmoe-onnx-fp32-v9",
+  "detector": "PP-OCRv6 medium ONNX",
+  "recognition": {
+    "cuda": {
+      "samples": 10899,
+      "failed": 0,
+      "skipped": 0,
+      "text_differences": 0,
+      "raw_text_differences": 0,
+      "termination_differences": 0,
+      "decoded_steps_differences": 0,
+      "filtered_exact_accuracy": 0.7975961097348381,
+      "max_score_difference": 8.195638656616211e-05,
+      "official_weighted_accuracy": 0.850812001101018,
+      "official_ten_group_macro": 0.8109283048360594,
+      "tested_package_version": "0.3.0.dev14"
+    },
+    "cpu": {
+      "samples": 10899,
+      "failed": 0,
+      "skipped": 0,
+      "text_differences": 1,
+      "raw_text_differences": 1,
+      "termination_differences": 0,
+      "decoded_steps_differences": 1,
+      "filtered_exact_accuracy": 0.7975043581979998,
+      "max_score_difference": 0.01205642976694632,
+      "official_weighted_accuracy": 0.850812001101018,
+      "official_ten_group_macro": 0.8109283048360594,
+      "tested_package_version": "0.3.0.dev13",
+      "applies_to": "0.3.0.dev14",
+      "cpu_code_and_models_identical": true,
+      "same_cpu_reference": "0.3.0.dev8",
+      "same_cpu_text_steps_eos_score_differences": 0,
+      "installed_current_version_regression_samples": 300
+    }
+  },
+  "cc_ocr": {
+    "dev": {
+      "images": 100,
+      "macro_f1": 0.7131863812825354,
+      "micro_f1": 0.8140047618306818
+    },
+    "eval": {
+      "images": 100,
+      "macro_f1": 0.7258491279221538,
+      "micro_f1": 0.7945454540507626
+    },
+    "images": 200,
+    "boxes": 1957,
+    "failed": 0,
+    "skipped": 0,
+    "page_text_differences": 0,
+    "raw_text_differences": 0,
+    "detection_differences": 0,
+    "decoding_contract_differences": 0
+  },
+  "performance": {
+    "cuda_whole_page": {
+      "rounds": 3,
+      "samples_per_backend": 24,
+      "pairs": [
+        {
+          "round": 1,
+          "torch_ms": 1124.4756288360804,
+          "onnx_ms": 921.554900589399,
+          "ratio": 1.2201938572698157
+        },
+        {
+          "round": 2,
+          "torch_ms": 983.9565325528383,
+          "onnx_ms": 913.2588730426505,
+          "ratio": 1.0774125076657055
+        },
+        {
+          "round": 3,
+          "torch_ms": 1038.3566328091547,
+          "onnx_ms": 953.9304067147896,
+          "ratio": 1.0885035485818277
+        }
+      ],
+      "mean_ms": {
+        "torch": 1048.9295980660245,
+        "onnx": 929.5813934489464
+      },
+      "round_mean_stdev_ms": {
+        "torch": 70.85368586606485,
+        "onnx": 21.490971442195402
+      },
+      "ratio": 1.128389192660441,
+      "text_differences": [],
+      "scope": "same medium+same8 CC-OCR eval pages+sameGPU,three alternating pairs; no broad full-CC speed claim"
+    },
+    "cpu_cropped_recognition": {
+      "rounds": 3,
+      "samples_per_backend": 60,
+      "pairs": [
+        {
+          "round": 1,
+          "torch_ms": 133.5505000781268,
+          "onnx_ms": 133.87302686460316,
+          "ratio": 0.9975908008205224
+        },
+        {
+          "round": 2,
+          "torch_ms": 180.9673246461898,
+          "onnx_ms": 178.97417438216507,
+          "ratio": 1.0111365244226171
+        },
+        {
+          "round": 3,
+          "torch_ms": 178.60913323238492,
+          "onnx_ms": 178.6530858837068,
+          "ratio": 0.9997539776539293
+        }
+      ],
+      "mean_ms": {
+        "torch": 164.37565265223384,
+        "onnx": 163.8334290434917
+      },
+      "ratio": 1.003309603002927,
+      "text_differences": [],
+      "scope": "cropped-recognition only, fixed20,three alternating pairs; not CPU whole-page or full-set performance"
+    }
+  },
+  "environment": {
+    "os": "Linux x86_64",
+    "glibc": "2.35",
+    "python": "3.10.12",
+    "onnxruntime": "1.20.0",
+    "numpy": "2.2.6",
+    "opencv_python_headless": "5.0.0.93",
+    "cpu": "Xeon Gold 6226R",
+    "gpu": "RTX 3090",
+    "driver": "575.57.08",
+    "cuda_runtime": "12.4.127",
+    "cudnn": "9.1.0.70"
+  },
+  "installation": {
+    "independent_cpu_cuda": true,
+    "torch": false,
+    "outside_source_directory": true,
+    "hash_locked_offline_install": true,
+    "sdist_runtime_files_equal": 29,
+    "model_prepared_ordinary_files": 40,
+    "persistent_real_ocr_examples": true,
+    "complete_offline_dependencies_in_zip": false
+  },
+  "wheel": {
+    "name": "openocr_mlt-0.3.0.dev14-py3-none-any.whl",
+    "bytes": 88248,
+    "sha256": "f5bcfbc81750b87b35ff0948459866f3c054cf6f8b2b1f29af99a2b32ebf0fea"
+  },
+  "sdist": {
+    "name": "openocr_mlt-0.3.0.dev14.tar.gz",
+    "bytes": 91909,
+    "sha256": "34aa38f1732e7d6d81712c80d7e9886ac718293b23bba3cb2500036876fad960"
+  },
+  "cpu_whole_page": {
+    "version": "0.3.0.dev14",
+    "scope": "CPU detection and CPU recognition, fixed nonrandom CC-OCR200,10 languages20 each; not full CC-OCR or detection Hmean",
+    "quality": {
+      "pages": 200,
+      "boxes": 1957,
+      "failed": 0,
+      "skipped": 0,
+      "crop_errors": 0,
+      "same_cpu_comparison": {
+        "passed": true,
+        "pages": 200,
+        "boxes": 1957,
+        "different_pages": 0,
+        "differences_by_page": {
+          "page_text": 0,
+          "detections": 0,
+          "raw_text": 0,
+          "filtered_text": 0,
+          "termination": 0,
+          "decoded_steps": 0,
+          "status": 0
+        },
+        "max_rec_score_difference": 1.7285346984863281e-06
+      },
+      "max_detection_score_difference": 0.0,
+      "metrics": {
+        "torch": {
+          "dev": {
+            "images": 100,
+            "macro_f1": 0.7131863812825354,
+            "micro_f1": 0.8140047618306818
+          },
+          "eval": {
+            "images": 100,
+            "macro_f1": 0.7258491279221538,
+            "micro_f1": 0.7945454540507626
+          }
+        },
+        "onnx": {
+          "dev": {
+            "images": 100,
+            "macro_f1": 0.7131863812825354,
+            "micro_f1": 0.8140047618306818
+          },
+          "eval": {
+            "images": 100,
+            "macro_f1": 0.7258491279221538,
+            "micro_f1": 0.7945454540507626
+          }
+        }
+      },
+      "cross_device_observation": {
+        "reference": "frozen CUDA, same fixed200",
+        "different_page_texts": 0
+      }
+    },
+    "performance": {
+      "rounds": 3,
+      "samples_per_backend": 24,
+      "pairs": [
+        {
+          "round": 1,
+          "torch_ms": 15015.680243377574,
+          "onnx_ms": 14303.62608062569,
+          "ratio": 1.0497813742290398
+        },
+        {
+          "round": 2,
+          "torch_ms": 15510.577811743133,
+          "onnx_ms": 15171.641253167763,
+          "ratio": 1.0223401379534072
+        },
+        {
+          "round": 3,
+          "torch_ms": 15150.97376552876,
+          "onnx_ms": 14863.175587030128,
+          "ratio": 1.0193631688472933
+        }
+      ],
+      "mean_ms": {
+        "torch": 15225.74394021649,
+        "onnx": 14779.48097360786
+      },
+      "ratio": 1.0301947657976307,
+      "all_output_comparisons_passed": true
+    },
+    "runtime_versions": {
+      "torch": {
+        "numpy": "2.2.6",
+        "Pillow": "12.3.0",
+        "opencv-python-headless": "5.0.0.93",
+        "onnxruntime": "1.20.0",
+        "torch": "2.5.1+cpu"
+      },
+      "onnx": {
+        "numpy": "2.2.6",
+        "Pillow": "12.3.0",
+        "opencv-python-headless": "5.0.0.93",
+        "onnxruntime": "1.20.0"
+      }
+    },
+    "frozen_wheel_sha256": "f5bcfbc81750b87b35ff0948459866f3c054cf6f8b2b1f29af99a2b32ebf0fea"
+  },
+  "documentation_updated": "2026-11-01",
+  "prior_frozen_evidence_date": "2026-10-31"
+}

openocr_mlt_cuda-0.4.0/examples/ocr_page.py

@@ -0,0 +1,9 @@
+"""Run complete page OCR with CPU or CUDA ONNX."""
+import argparse,json
+from openocr_mlt import OpenOCRMLT
+p=argparse.ArgumentParser()
+p.add_argument("--model-dir",required=True);p.add_argument("--det-model-dir",required=True)
+p.add_argument("--image",required=True);p.add_argument("--device",default="cpu")
+a=p.parse_args()
+with OpenOCRMLT(model_dir=a.model_dir,det_model_dir=a.det_model_dir,device=a.device) as ocr:
+ print(json.dumps(ocr.ocr(a.image).to_dict(),ensure_ascii=False))

openocr_mlt_cuda-0.4.0/examples/prepared_ocr.py

@@ -0,0 +1,13 @@
+"""Prepare the pinned models, then run one page."""
+import argparse,json
+from openocr_mlt import OpenOCRMLT,fetch_models
+p=argparse.ArgumentParser()
+p.add_argument("--bundle",help="Local bundle path or distributor-provided HTTPS URL")
+p.add_argument("--cache-dir",default="./models-cache")
+p.add_argument("--offline",action="store_true")
+p.add_argument("--image",required=True)
+p.add_argument("--device",default="cpu")
+a=p.parse_args()
+paths=fetch_models(source=a.bundle,cache_dir=a.cache_dir,offline=a.offline)
+with OpenOCRMLT(model_dir=paths["recognition_dir"],det_model_dir=paths["detector_dir"],device=a.device) as engine:
+ print(json.dumps(engine.ocr(a.image).to_dict(),ensure_ascii=False))

openocr_mlt_cuda-0.4.0/examples/recognize.py

@@ -0,0 +1,8 @@
+"""Run the installed package from any working directory."""
+import argparse,json
+from openocr_mlt import OpenOCRMLT
+p=argparse.ArgumentParser()
+p.add_argument("--model-dir",required=True);p.add_argument("--image",required=True);p.add_argument("--device",default="cpu")
+a=p.parse_args()
+with OpenOCRMLT(model_dir=a.model_dir,device=a.device) as ocr:
+ print(json.dumps(ocr.recognize(a.image)[0].to_dict(),ensure_ascii=False))

openocr_mlt_cuda-0.4.0/pyproject.toml

@@ -0,0 +1,29 @@
+[build-system]
+requires=["setuptools>=77"]
+build-backend="setuptools.build_meta"
+
+[project]
+name="openocr-mlt-cuda"
+version="0.4.0"
+maintainers=[{name="ericleyu"}]
+description="Offline multilingual OCR with ONNX Prefill/Decode recognition and PP-OCRv6 medium detection"
+readme="README.md"
+requires-python=">=3.10"
+dependencies=["numpy>=2.2,<2.3","Pillow>=12.3,<12.4","opencv-python-headless>=5.0,<5.1","onnxruntime-gpu==1.20.0","pyclipper==1.4.0","nvidia-cuda-runtime-cu12==12.4.127","nvidia-cublas-cu12==12.4.5.8","nvidia-cudnn-cu12==9.1.0.70","nvidia-cuda-nvrtc-cu12==12.4.127","nvidia-cufft-cu12==11.2.1.3","nvidia-curand-cu12==10.3.5.147"]
+
+[project.optional-dependencies]
+torch=["torch>=2.5,<2.6","torchvision>=0.20,<0.21","safetensors>=0.8,<0.9"]
+
+[project.urls]
+Models="https://modelscope.cn/models/Strawberry015/openocr-mlt-models"
+"Model mirror"="https://huggingface.co/ericleyu/openocr-mlt-models"
+"Upstream recognition"="https://github.com/YesianRohn/ScriptMoE"
+
+[project.scripts]
+openocr-mlt="openocr_mlt.cli:main"
+
+[tool.setuptools.packages.find]
+where=["src"]
+
+[tool.setuptools.package-data]
+openocr_mlt=["assets/*.txt","assets/*.json"]

openocr_mlt_cuda-0.4.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

openocr_mlt_cuda-0.4.0/src/openocr_mlt/__init__.py

@@ -0,0 +1,11 @@
+"""Offline, serial, FP32 ScriptMoE recognition."""
+from .api import OpenOCRMLT
+from .config import RecognitionConfig
+from .types import RecognitionResult, DetectionResult, OCRLine, OCRResult
+from .detection import TextDetector, DetectionConfig
+
+__version__ = "0.4.0"
+__all__ = ["OpenOCRMLT", "RecognitionConfig", "RecognitionResult", "DetectionResult", "OCRLine", "OCRResult", "TextDetector", "DetectionConfig"]
+
+from .model_fetch import fetch_models
+__all__.append("fetch_models")