Possible issue in whitening procedure?

Hello everyone,

I am currently running spikeinterface 0.101.2 to process some MaxTwo data. I concatenated some recording segments and centered the data, as I assumed it might be relevant to the issue but it doesn't change the outcome. In the process of running SpykingCircus2 I get an error with the attached error trace basically telling me that the whitening procedure of sklearn has a type mismatch. I have checked that my data is properly conditioned (no nan values, non-singular covariance matrix). I have also found out that if I step in and use the debugging console to change the dtype of the data array to float64 (i.e. the requested double type) it works fine. I have also tried my hand at finding the root cause in the sklearn method but didn't have much success. However, I found out that the procedure causing the issue runs fine for what I assume are all but one iteration and then crashes on the last pass, but I was unable to figure out what might cause the difference that makes it crash. Maybe anyone here has an idea what causes this issue?

The error trace:

sorting = run_sorter(sorter.sorter_name, rec.Extractor, folder = folder, verbose = True, remove_existing_folder=overwrite, **params)    
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^    
File "C:\ProgramData\mambaforge\envs\maxwelltesting\Lib\site-packages\spikeinterface\sorters\runsorter.py", line 199, in run_sorter       
  return run_sorter_local(**common_kwargs)
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\ProgramData\mambaforge\envs\maxwelltesting\Lib\site-packages\spikeinterface\sorters\runsorter.py", line 261, in run_sorter_local 
  SorterClass.run_from_folder(folder, raise_error, verbose)
File "C:\ProgramData\mambaforge\envs\maxwelltesting\Lib\site-packages\spikeinterface\sorters\basesorter.py", line 301, in run_from_folder
  raise SpikeSortingError(
spikeinterface.sorters.utils.misc.SpikeSortingError: Spike sorting error trace:
Traceback (most recent call last):
File "C:\ProgramData\mambaforge\envs\maxwelltesting\Lib\site-packages\spikeinterface\sorters\basesorter.py", line 261, in run_from_folder
  SorterClass._run_from_folder(sorter_output_folder, sorter_params, verbose)
File "C:\ProgramData\mambaforge\envs\maxwelltesting\Lib\site-packages\spikeinterface\sorters\internal\spyking_circus2.py", line 152, in _run_from_folder
  recording_w = whiten(recording_f, mode="local", radius_um=radius_um, dtype="float32", regularize=True)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\ProgramData\mambaforge\envs\maxwelltesting\Lib\site-packages\spikeinterface\preprocessing\whiten.py", line 84, in __init__
  W, M = compute_whitening_matrix(
         ^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\ProgramData\mambaforge\envs\maxwelltesting\Lib\site-packages\spikeinterface\preprocessing\whiten.py", line 209, in compute_whitening_matrix
  estimator.fit(data)
File "C:\ProgramData\mambaforge\envs\maxwelltesting\Lib\site-packages\sklearn\base.py", line 1473, in wrapper
  return fit_method(estimator, *args, **kwargs)
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\ProgramData\mambaforge\envs\maxwelltesting\Lib\site-packages\sklearn\covariance\_graph_lasso.py", line 1109, in fit
  self.covariance_, self.precision_, self.costs_, self.n_iter_ = _graphical_lasso(
                                                                 ^^^^^^^^^^^^^^^^^
File "C:\ProgramData\mambaforge\envs\maxwelltesting\Lib\site-packages\sklearn\covariance\_graph_lasso.py", line 139, in _graphical_lasso
  coefs, _, _, _ = cd_fast.enet_coordinate_descent_gram(
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "_cd_fast.pyx", line 569, in sklearn.linear_model._cd_fast.enet_coordinate_descent_gram
ValueError: Buffer dtype mismatch, expected 'const double' but got 'float'

Start at spikeinterface/sorters/internal/spyking_circus2.py, where whitening is called with dtype="float32", then inspect spikeinterface/preprocessing/whiten.py around compute_whitening_matrix and estimator.fit(data). Reproduce the MaxTwo path and trace the final iteration's dtype; done means the SpykingCircus2 whitening path no longer raises the sklearn buffer mismatch for this input.