Fix attention caching to make transcription run 30% faster

[vickianand]

For a dict.get(a, f()) call, f() is run even when the dict already has a. This is causing the bug in attention caching logic.
So making this change fixes the caching logic, and gets a sweet (expected) speedup of close to 30%.

For a close to 5 min test audio, on an Nvidia Quadro RTX 8000 gpu:

• before this fix, the large model took 102.8s, and after this fix it takes 68.9s
• before this fix, the medium.en model took 55.1s, and after this fix it takes 39.7s

[vickianand]

All thanks to @ritheshkumar95 for finding this.

[akashmjn]

Can repro on my end as well

17:38 wav file
On Quadro RTX 5000 (with --condition_on_previous_text=True, small.en model)

	Runtime (m:s)
Original	01:42
With kv cache fixed	01:24

Really elegant catch! 🙌 learnt something about Python today 🙂

[jongwook]

Thank you very much! This was an oversight when I was factoring out the caching part for the open source repo.

[nlgtuankiet]

@vickianand @ritheshkumar95
Thank you very much!
I was getting around 0.88-1.1x for large models, now it 1.6x

[rjwilmsi]

Wow, this is great news, thank you to all involved. Title says 30% faster on GPU but the speed up is much more on CPU, at least for me.

Ryzen 5 4500U / 6 core laptop CPU. 6m30s youtube video, English (https://www.youtube.com/watch?v=GFu64hnqzVo):

Model	Before PR	After PR
tiny.en	6m19s	2m9s
base.en	15m39s	4m29s
small.en	60m45s	11m18s