Was Whisper trained with FP16?

[iblueai]

Thank you for the authors' great work.
I have a small question. I've been examining the Whisper Large v3 model, and it seems that the weights exceed the accuracy range of FP16.
Additionally, when roughly checking the scale of the final logits, it appears to be of a scale that cannot be computed with FP16 weights and losses.
While the paper briefly mentions training with FP16 and the source code does not load the model as FP16, there seems to be ambiguity.
Some aspects suggest loading the model as FP16, which might lead to unintended operations.
It would be helpful to clearly understand whether Whisper was trained with Mixed Precision or purely with FP16.
This clarification could assist future research.
Thank you.