I have succesfully created my own custom datatype (and some custom operators on this datatype) following the example as in: https://tvm.apache.org/docs/how_to/extend_tvm/bring_your_own_datatypes.html This works great on CPU, using the LLVM back-end.
But when i change the target to cuda (to run on Nvidia GPU), the example in the above link fails. Is it correct that custom datatypes are not supported for the cuda-backend? If so, how can this be added easily?