Incompatible with CUDA v12.1 about faster-whisper

[guillaumekln]

As mentioned in the README, the Python module ctranslate2 requires the library cuBLAS 11.x which is installed with CUDA 11.x.

Support for CUDA 12.x is not planned at the moment.

[pandruszkow]

Makes sense. Would you accept a README change that makes this information more prominent? I know it'd have saved me some time if I hadn't missed this info.

[guillaumekln]

Sure, feel free to send a PR to improve the README.

[Geremia]

@guillaumekln Why isn't CUDA 12.x support planned?

[guillaumekln]

If we update the CTranslate2 binaries to use CUDA 12, it will break all existing users so this is a change we would like to avoid for now.

A possible solution is to release 2 versions of the binaries, one for CUDA 11 and another for CUDA 12. However, we don't have the resources to set that up and maintain it.

Given CUDA 11 applications still work on newer graphic cards, CUDA 12 support for the released binaries is not a priority.

[Geremia]

@guillaumekln Are there instructions how I can build it myself for CUDA 12?

[Geremia]

@Geremia https://github.com/OpenNMT/CTranslate2/blob/master/docs/installation.md#install-from-sources

[guillaumekln]

You can also check out the Dockerfile which acts as a complete build example.