correct docs regarding flash attention

[BBC-Esq]

Flash Attention is still listed in the documentation:

https://opennmt.net/CTranslate2/python/ctranslate2.Generator.html

I'd recommend keeping it since you do not have a history of documentation for different versions...and then just add a bullet point mentioning that it's only available for version 4.3.1 or 4.3 or whatever...

Not sure if it's mentioned elsewhere in the documentation, didn't check everywhere...

[minhthuc2502]

Thank you for your feedback. I will update the doc and version handler ASAP.

[winstxnhdw]

@BBC-Esq have you seen any benefits to using Flash Attention?

[BBC-Esq]

I re-ran some limited testing, including some very large contexts, and did not see much if any improvement - either in tokens/s or VRAM usage. However, I haven't been able to fully test it yet. My previous tests, for whatever reason, indicated a discernable improvement when running models that have a certain architecture (I believe it was Mistral) but not Llama-based for some reason. I'd have to revisit my prior benchmarks and update them to be certain and will post when I get the time to do that.