How to activate flash attention?

[pooya-mohammadi]

No description provided.

[Jeronymous]

That is a question for openai-whisper (whisper-timestamped being a wrapper around it).
https://github.com/openai/whisper/discussions?discussions_q=flash+

I know flash attention is implemented in transformers.
So maybe what should be done is to be able to use either openai-whisper or transformers backend in whisper-timestamped...
But that sounds like a big project.

[Jeronymous]

I added (HuggingFace) transformers backend in whisper-timestamped.
Unfortunately, this does not unlock the use of flash attention, because when enabling flash attention in transformers, there is no more possibility of accessing attention weights (and this is require to guess timestamps).
Also, I did not notice any speedup by using HF transformers instead of openai-whisper implementation.