Currently `cuPy` is used as it is much faster. If it is not well-supported (e.g. errors are not informative), revert to pyTORCH tensors or consider having a separate branch supporting pyTORCH.
Currently
cuPyis used as it is much faster.If it is not well-supported (e.g. errors are not informative), revert to pyTORCH tensors or consider having a separate branch supporting pyTORCH.