transformers

5be46dfc - [Whisper] Fix errors with MPS backend introduced by new code on word-level timestamps computation (#28288)

[Whisper] Fix errors with MPS backend introduced by new code on word-level timestamps computation (#28288) * Update modeling_whisper.py to support MPS backend Fixed some issue with MPS backend. First, the torch.std_mean is not implemented and is not scheduled for implementation, while the single torch.std and torch.mean are. Second, MPS backend does not support float64, so it can not cast from float32 to float64. Inverting the double() when the matrix is in the cpu fixes the issue while should not change the logic. * Found another instruction in modeling_whisper.py not implemented byor MPS After a load test, where I transcribed a 2 hours audio file, I got into a branch that did not fix in the previous commit. Similar fix, where the torch.std_mean is changed into torch.std and torch.mean * Update modeling_whisper.py removed trailing white spaces Removed trailing white spaces * Update modeling_whisper.py to use is_torch_mps_available() Using is_torch_mps_available() instead of capturing the NotImplemented exception * Update modeling_whisper.py sorting the import block Sorting the utils import block * Update src/transformers/models/whisper/modeling_whisper.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update src/transformers/models/whisper/modeling_whisper.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update src/transformers/models/whisper/modeling_whisper.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> --------- Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>