Good news for folks who like living on the edge — cuDNN backend packages now have preview releases available on PyPI. This means you can try upcoming features and fixes before they hit the stable channel.
Installing a Preview
pip install nvidia-cudnn-cu13 --pre --extra-index-url https://pypi.nvidia.com
The --pre flag tells pip to include pre-release versions, and the extra index URL points to NVIDIA’s PyPI repository where preview builds are published.
Stable Releases
Nothing changes for stable installs. If you want the latest tested, production-ready build:
pip install nvidia-cudnn-cu13
No extra flags, no extra index — same as always.
Why Previews?
Preview releases give you early access to:
- New kernel support for upcoming GPU architectures
- Performance improvements before they’re fully validated across all configurations
- Bug fixes that haven’t gone through the full release qualification cycle yet
This is especially useful if you’re developing against a new hardware target or want to validate that an upcoming release doesn’t regress your workload.
A Note on Stability
Preview builds go through CI and basic validation, but they haven’t completed the full release qualification matrix. Use them for development and testing — not for production training runs where reproducibility matters. If you hit an issue, file it on GitHub and we’ll prioritize it for the stable release.
CUDA 12 Users
Same pattern, just swap the package name:
# Preview
pip install nvidia-cudnn-cu12 --pre --extra-index-url https://pypi.nvidia.com
# Stable
pip install nvidia-cudnn-cu12