Flash-Attention Ruins Timestamps on 1.75

[AeneasZhu]

I found that the timestamps of transcript unmatched the speaker many times when turned -fa on after updating to 1.75. Here are the transcripts:

fa-on.txt

fa-off.txt

You can try the extension to add sub on Youtube: https://github.com/yashagarwal1411/SubtitlesForYoutube and check the difference. (Rename the file ".srt")

Here is the original video: https://www.youtube.com/watch?v=_TTI2ZQZpXc.

After the opening music is off, the timestamps is in a mess (though the transcript quality is good) for almost 1 min. But when I turned off flash-attention by removing -fa, the timestamps could match the speaker.

By the way, I used large-v2 to transcribe it.