For certain I/O functionality I need to transfer tensors that potentially reside in GPU memory to host memory. The following code works
torch::tensor a; auto a_cpu = a.cpu(); auto a_accessor = a_cpu.accessor<float, 1>();
auto a_accessor = a.cpu().accessor<float, 1>();
does not compile with the following compiler error
libtorch/1.13.1/include/ATen/core/TensorBase.h:562:23: note: candidate function [with T = double, N = 1] has been explicitly deleted TensorAccessor<T,N> accessor() && = delete;
Is it possible to do the transfer to host memory inline?