Repository navigation
opencv-python-cuda 4.14.0-dev0
Pre-release
Pre-release
Preview build of OpenCV 4.14.0 with CUDA support, including hardware video encode/decode (NVDEC/NVENC).
Installation
From the package index (the <5 keeps you on the OpenCV 4.x line):
pip install "opencv-python-cuda<5" --extra-index-url https://breakthrough.github.io/opencv-python-cuda/packages/
Or download the wheel below and install it directly:
pip install opencv_python_cuda-4.14.0+2de54de-cp37-abi3-win_amd64.whl
Make sure to uninstall any existing OpenCV packages first. This package can be uninstalled via pip uninstall opencv-python-cuda.
Build/Version Information
- Build Run
- OpenCV 4.14.0 with all features excluding non-free
- Nvidia CUDA Toolkit v12.9.2
- Nvidia CuDNN 9.28.0.42
- Nvidia Video Codec SDK 13.0.37 (hardware video decode/encode)
- Device compatibility: Maxwell through Blackwell (SASS: 5.0, 5.2, 6.0, 6.1, 7.0, 7.5, 8.0, 8.6, 8.9, 9.0, 10.0; PTX: 10.0)
System Requirements
- Windows 7 64-bit or newer
- Python 3.7 or newer
- Nvidia GPU that is Maxwell-class or newer (GeForce 900-series)
- Nvidia display driver must be installed:
cv2links againstnvcuvid.dll/nvEncodeAPI64.dll, which ship with the driver, soimport cv2fails on machines without it
Changes from 4.13.0-dev1
- Updated OpenCV from 4.13.0 to 4.14.0 (merged upstream opencv-python 4.x, including packaging updates)
- Fixed the DNN CUDA backend. Using
cv2.dnnwithDNN_BACKEND_CUDA(includingcv2.dnn.getAvailableTargets(cv2.dnn.DNN_BACKEND_CUDA)) terminated the Python process withCannot load symbol cudnnGetVersion, because cuDNN 9 could not find its own sub-libraries in thecv2package directory. Earlier releases that bundle cuDNN 9 (including 4.13.0-dev1) are affected; as a workaround for those, add the installedcv2package directory toPATHbefore importingcv2. - Updated Nvidia CUDA Toolkit from 12.9.1 to 12.9.2
- Updated Nvidia CuDNN from 9.18.1.3 to 9.28.0.42
- Wheels can now be installed by name from the package index (see Installation)
- CI now checks that cuDNN's sub-libraries load from the package (
tools/check_cudnn_loading.py)