GPU/CUDA Error Handling

[ozancaglayan]

Hello,

I'm looking to provide a more error-robust experience for one of my codebases that is built around faster-whisper. So far, did not get any issues & errors but I wonder what would happen if a device-side assert is triggered, will I get an exception from CTranslate? What would be the exception type?

More importantly, do you have some test cases that could be used to certainly trigger some errors on the device side for which I can test my codebase's recovery possibilities?

Looking for a pynvml based event watching solution for now but was curious what can I get from CTranslate2 & faster-whisper.

Thanks!