Why does Halide allocate memory on CPU when wrapping a PyTorch CUDA tensor?

[twesterhout]

In Halide::PyTorch::wrap when tensor is on CUDA a new tensor is allocated on the CPU:

Halide/src/runtime/HalidePyTorchHelpers.h

Line 95 in da7c66e

buffer = Buffer<scalar_t>(dims);

Is there a way to avoid it?