RuntimeError: imag is not implemented for tensors with non-complex dtypes

[thet0ast3r]

Hi, im facing an issue with an m3 macbook air,
i installed all requirements for mlx according to the instructions and tried running:

python convert.py --torch-name-or-path large-v3-turbo

however, i only get this error, pretty often:

Exception ignored in PyObject_HasAttrString(); consider using PyObject_HasAttrStringWithError(), PyObject_GetOptionalAttrString() or PyObject_GetAttrString():
Traceback (most recent call last):
  File "/Users/tobi-mba/priv/liteASR/LiteASR/src/mlx/mlx_whisper/convert.py", line 240, in remap
    value = mx.array(value.detach())
RuntimeError: imag is not implemented for tensors with non-complex dtypes.
Exception ignored in PyObject_HasAttrString(); consider using PyObject_HasAttrStringWithError(), PyObject_GetOptionalAttrString() or PyObject_GetAttrString():
Traceback (most recent call last):
  File "/Users/tobi-mba/priv/liteASR/LiteASR/src/mlx/mlx_whisper/convert.py", line 240, in remap
    value = mx.array(value.detach())
RuntimeError: imag is not implemented for tensors with non-complex dtypes.
Exception ignored in PyObject_HasAttrString(); consider using PyObject_HasAttrStringWithError(), PyObject_GetOptionalAttrString() or PyObject_GetAttrString():
Traceback (most recent call last):
  File "/Users/tobi-mba/priv/liteASR/LiteASR/src/mlx/mlx_whisper/convert.py", line 240, in remap
    value = mx.array(value.detach())
RuntimeError: imag is not implemented for tensors with non-complex dtypes.
Exception ignored in PyObject_HasAttrString(); consider using PyObject_HasAttrStringWithError(), PyObject_GetOptionalAttrString() or PyObject_GetAttrString():
Traceback (most recent call last):
  File "/Users/tobi-mba/priv/liteASR/LiteASR/src/mlx/mlx_whisper/convert.py", line 240, in remap

Sadly i can't really come up with a fix, is there someone that can point me at the root cause of this?

python 3.13, macos 15.3.2

[MarktHart]

You probably need to change a few lines in audio.py (or where the log mel spectrogram is done):

This uses imaginary numbers:

stft = torch.stft(audio, N_FFT, HOP_LENGTH, window=window, return_complex=True)
magnitudes = stft[..., :-1].abs() ** 2

This should (I didn't run it) be equivalent without:

stft = torch.stft(audio, N_FFT, HOP_LENGTH, window=window, return_complex=False)
magnitudes = (stft[..., :-1, :] ** 2).sum(dim=-1)

[kamahori]

Is this issue resolved? @thet0ast3r