The __init__ function of LMMSEInterpolator runs very slow when GPU is enabled. Specifically, the _build_pilot_mask() function takes a very long time to finish.

The issue is that in _build_pilot_mask(), the variables mask and pilots are not converted to NumPy arrays, which force the following loops to run on the GPU. In my case, the _build_pilot_mask() function takes about 1 minutes instead of 0.2 second.

So the two assign statements from

 mask = np.array(pilot_pattern.mask)
 pilots = np.array(pilot_pattern.pilots)