CI : Add CUDA 11.8.0 support #1554
Conversation
|
Just out of curiosity... |
Yes, there are certain users who might face issues updating their drivers. This includes individuals using Kepler GPUs, such as the K80 or K40 models, and those who do not have administrative privileges on their computers. Also, for instance, some users who have installed other software might need specific versions of CUDA drivers. It would be quite troublesome to reinstall drivers many times just to run whisper.cpp. |
Yes. We can also mention that in README. Actually, adding a CI won’t waste much time, as they run in parallel. The total duration of the CI is determined by the longest running one. Currently, the longest ones are armv7 and ppc64le in ubuntu-latest-gcc. In theory, these could be merged with ubuntu-latest, which would speed up the CI process. |
|
Wow, thanks for the excellent explanation, bobqianic! 💯 |
There was a problem hiding this comment.
Thank you for looking into this @bobqianic !
Speeding-up the CI would be useful in case you have ideas how to do it!
* try to fix cublas build in CI * add multiple cuda-toolkit version * Update build.yml * Disable CUDA-toolkit 10.2.89
The reason for the inability to run is due to the incompatibility between the CUDA driver and CUDA toolkit. Previously, the CI for CUDA used version 12.1.0 of the CUDA toolkit, which requires a driver of at least version
>=527.41on Windows. The filewhisper.cppwill dynamically link with the runtime in the CUDA toolkit, and the runtime will in turn dynamically link with the driver. The incompatibility causes it to fail to run. The solution is simple: adding a Windows binary compiled with version 11.X of the CUDA toolkit can solve the problem.Reference: here
Closes #1552