A better version would be (np.where returns tuples, but we don't want tuples):

x,i=numpy.unique(y, return_inverse=True)
f=[numpy.where(i==ind)[0] for ind in range(len(x))]

You can also do it this way, but it is much harder to read IMO:

x=numpy.unique(y)
f=numpy.split(numpy.argsort(y), numpy.nonzero(numpy.diff(numpy.sort(y)))[0]+1)

This version figures out the indexes needed to put the values of y in sorted order (the same order x uses), then splits it into sub-arrays based on value.  The principle is simpler but the implementation looks like clear to me.

Note that these are only guaranteed to work on 1D arrays, I have not tested them on multidimensional arrays