It only works with Python 3.9.9/3.8-3.11?

[HubKing]

The front page says:

We used Python 3.9.9 and PyTorch 1.10.1 to train and test our models, but the codebase is expected to be compatible with Python 3.8-3.11 and recent PyTorch versions

Do I have to use the exact version of one of the two? I already have installed Python 3.11 and PyTorch 2.3, so I wonder if I could use that. Arch Linux.

[glangford]

I have used Python 3.11.7 with torch 2.2.2 (Mac) successfully.

[Et0h]

I have Python 3.12.6 64-bit and Torch 2.4.1 (Windows 10) successfully.

[filtercodes]

Python 3.10.9 here with torch==2.2.1

-> RuntimeError: Given groups=1, weight of size [1280, 128, 3], expected input[1, 80, 3000] to have 128 channels, but got 80 channels instead

Mac OS metal though

[glangford]

@filtercodes This is likely because an older version of whisper is being used in combination with the large-v3 model. The input to large-v3 uses 128 Mel frequency bins instead of 80.

From the announcement,

The large-v3 model is available in openai-whisper==20231106 and after

• `large-v3` release #1762

[filtercodes]

@filtercodes This is likely because an older version of whisper is being used in combination with the large-v3 model. The input to large-v3 uses 128 Mel frequency bins instead of 80.

From the announcement,

The large-v3 model is available in openai-whisper==20231106 and after

• large-v3 release #1762

This is the version that shows up -> openai-whisper @ git+https://github.com/openai/whisper.git@25639fc17ddc013d56c594bfbf7644f2185fad84

Command I was running was: pip install --upgrade --no-deps --force-reinstall git+https://github.com/openai/whisper.git and that was supposed to update whisper to the latest version. Not sure what I'm missing here....

[glangford]

Is the code calling log_mel_spectrogram directly? If so you can specify the number of channels with n_mels.

• Error in the "large-v3" model #1778 (comment)

To confirm that the right version of whisper is being called (and the Python code is not an issue), you could try using the command line with large-v3.

[filtercodes]

Is the code calling log_mel_spectrogram directly? If so you can specify the number of channels with n_mels.

• Error in the "large-v3" model #1778 (comment)

To confirm that the right version of whisper is being called (and the Python code is not an issue), you could try using the command line with large-v3.

That was it :) n_mels has to be redefined because it uses 80 as a default even it gets it from model.dims.n_mels.
Also it works only on CPU, not on metal. A simple to("mps") didn't do the trick. Then a (mostly) single threaded inference takes couple of minutes, which doesn't sound very "turbo" at all

[glangford]

Good news. :)

mps acceleration is still a work in progress, for "too much history" see the thread below. For the record, the turbo model on CUDA is insanely quick.

• General MPS op coverage tracking issue pytorch/pytorch#77764