RapidOCRLite

RapidOCRLite (RapidOCRLite) is published on PyPI as RapidOCRLite, separate from the official rapidocr. The import name is still rapidocr: pip install rapidocrlite, then from rapidocr import RapidOCR. Both projects install the same rapidocr package directory, so do not install them in one environment (run pip uninstall rapidocr first). RapidOCRLite is the lightweight native PP-OCRv6 engine with a RapidOCR-compatible API. The ONNX Runtime 3.x line (latest tag v3.9.2) stays in RapidAI/RapidOCR.

from rapidocr import RapidOCR, __version__

engine = RapidOCR(params={
    "Det.model_path": "PP-OCRv6_det_tiny.onnx",
    "Rec.model_path": "PP-OCRv6_rec_tiny.onnx",
    "Rec.rec_keys_path": "ppocrv6_tiny_dict.txt",
    "Global.use_cls": False,
})
result = engine("page.png")
print(__version__)
print(result.txts, result.scores)
print(result.to_json())

If model paths are omitted, RapidOCR() downloads the same files as 3.x: default_models.yaml selects the URL, the file is saved under Global.model_root_dir (the package models/ directory when that is null) using the URL filename, and the SHA256 is checked. download_models() and download_models(config_yaml) follow 3.x and return None. PP-OCRv6 ONNX recognizers also download the dictionary published next to the MNN bundles (ppocrv6_tiny_dict.txt / ppocrv6_dict.txt).

Build from a checkout (CMake and a C++20 compiler required). The repository root is the installable project, so the sdist contains the native sources:

pip install .

pip install ./python also works from a checkout, because that project points CMake at the parent tree. It does not produce a self-contained sdist.

The binding is ctypes over the C ABI (include/rapidocrlite/rapidocrlite.h), so the wheel ships librapidocrlite and does not add a second C++ wrapper.

Vulkan

Wheels are built with -DPPOCR_ENABLE_VULKAN=OFF. They do not contain the compute shaders and do not need libvulkan. A checkout that has the Vulkan-Headers submodule and glslangValidator can build a Vulkan-enabled package with:

CMAKE_ARGS="-DPPOCR_ENABLE_VULKAN=ON -DPPOCR_BUILD_STATIC=OFF" pip install .

The runtime opens libvulkan.so.1 with dlopen. It is not linked into the library. Enable it per engine, in the same style as 3.x use_cuda:

engine = RapidOCR(params={
    "EngineConfig.ppocr_cpp.use_vulkan": True,
    "EngineConfig.ppocr_cpp.device_index": 0,
})

device_index -1 skips CPU Vulkan devices such as lavapipe. Set it to the lavapipe index (0 when it is the only device) together with VK_ICD_FILENAMES pointing at lvp_icd.json. If no compute device exists, the call logs a warning and runs on CPU. backend: vulkan is the same attempt. backend: hybrid matches the C++ hybrid path. See docs/API.md.

Migrating from rapidocr 3.x

RapidOCR(config_path=..., params={"Section.key": value}) and engine(img, use_det=..., use_cls=..., use_rec=..., text_score=..., box_thresh=..., unclip_ratio=...) keep the same names. result.txts, result.scores, result.boxes, result.elapse, to_json(), to_markdown(), and vis() are still there. File paths, URLs, bytes, and PIL images are RGB. A numpy array is treated as BGR, then swapped, which is what 3.x does for OpenCV images.

What does not carry over from 3.x:

A call that only reads txts, scores, and boxes from a PP-OCRv6 model can switch by installing this package and leaving use_cls at its default. Set use_cls=False to skip the warning. Do not pass an older OCR version.

Wheels and publishing

.github/workflows/build-wheels.yml builds the manylinux2014 wheels and the sdist. It runs on pushes and pull requests to main, on tags named v5*, and from the Actions “Run workflow” button. Only a v5* tag push publishes: wheels and sdist go to PyPI, and wheels, sdist, SDK archives and the Java jar go to the GitHub Release of that tag. Pushes to main and PRs only build and test.

LoongArch64 wheels (manylinux_2_38_loongarch64, CPython 3.9–3.13, glibc ≥ 2.38) are attached to the GitHub Release only: PyPI does not accept loongarch64 platform tags. They are verified under QEMU user-mode emulation only (no real-hardware testing, no performance numbers). Install from the release URL, choosing the file for your Python version and release tag:

pip install numpy Pillow PyYAML requests tqdm
pip install https://github.com/RapidAI/RapidOCRLite/releases/download/v5.0.0a13/rapidocrlite-5.0.0a13-cp312-cp312-manylinux_2_38_loongarch64.whl

5.0.0a6 is a pre-release: pip install --pre RapidOCRLite or pip install rapidocrlite==5.0.0a6. pip install rapidocr installs the official 3.x package instead.

How to turn on PyPI Trusted Publishing (or the API-token fallback) and which tag to push are in docs/RELEASING.md.

Ported from main (#755)

• RawMemoryImage(address, length, width, height, pixel_format="BGRA", stride=None, bottom_up=False) reads an uncompressed pixel buffer (GRAY/BGR/RGB/BGRA/BGRX/RGBA/RGBX) by address.
• bytes, bytearray, MemoryImage and PIL inputs honor the EXIF orientation, like paths.
• Model routing lives in default_models.yaml under model_routes. RapidOCR.list_supported_langs("rec", "PP-OCRv6", "tiny") and rapidocr --list-langs --task rec --ocr_version PP-OCRv6 --model_type tiny list the languages.
• Global.model_root_dir accepts str or Path; the recognizer dictionary is stored there too.
• No OpenCV dependency or extra.