We've all been there. You're setting up a custom installation of Jax/Torch/TensorFlow and... Agh! The cuDNN version you installed doesn't match what your framework supports! You hastily install a new cuDNN distribution and now you have missmatch with your CUDA toolkit, and a pretty big mess. cuDNN-versions aims to ease some of this strain. This repository provides a scoop bucket containing every cuDNN build, for every applicable CUDA version, so that you can easily search and install what you're looking for.
This repo can be added like any other scoop bucket by running the following command.
scoop bucket add cudnn-versions https://github.com/shmishtopher/cudnn-versions
You should be able to search for cuDNN distributions with scoop search.
❯ scoop search cuDNNv8.2
Results from local buckets...
Name Version Source Binaries
---- ------- ------ --------
cuDNNv8.2.0-CUDAv10.2-windows10 8.2.0 cudnn-versions
cuDNNv8.2.0-CUDAv11.3-windows 8.2.0 cudnn-versions
cuDNNv8.2.1-CUDAv10.2-windows10 8.2.1 cudnn-versions
cuDNNv8.2.1-CUDAv11.3-windows 8.2.1 cudnn-versions
cuDNNv8.2.2-CUDAv10.2-windows10 8.2.2 cudnn-versions
cuDNNv8.2.2-CUDAv11.4-windows 8.2.2 cudnn-versions
cuDNNv8.2.4-CUDAv10.2-windows10 8.2.4 cudnn-versions
cuDNNv8.2.4-CUDAv11.4-windows 8.2.4 cudnn-versions
Every cuDNN build is index by a target tripple that looks like <CUDNN_VERSION>-<CUDA_VERSION>-<PLATFORM>.
| Field | Description | Example |
|---|---|---|
| CUDNN_VERSION | The version of the cuDNN library | cuDNNv8.4.1 |
| CUDA_VERSION | The compatible CUDA toolkit version | CUDAv11.6 |
| PLATFORM | The platform the distribution is built for | windows10 |
Manifests in this bucket are created automatically by the discover-versions.ps1 script in the root of this repository. Any changes that you want to see should be made to that file.
For assistance, please open up an issue or contact me directly at me@shmish.dev.