Accessing the whole binary with Nx.to_flat_list or Nx.backend_transfer is not blocked by ongoing computation, however tensor access involves slicing and therefor does block. So whenever we need to slice, we are going to transfer to binary backend first.
Follow up to #244.
Accessing the whole binary with
Nx.to_flat_list
orNx.backend_transfer
is not blocked by ongoing computation, however tensor access involves slicing and therefor does block. So whenever we need to slice, we are going to transfer to binary backend first.