The python interface currently assumes a copy to host data even if internally data is stored on the GPU. Rather than directly linking against PyTorch, it looks like it may be better to use DLPack which defines a protocol that is compatible with numpy, PyTorch, Jax, etc.
See: https://dmlc.github.io/dlpack/latest/index.html
The python interface currently assumes a copy to host data even if internally data is stored on the GPU. Rather than directly linking against PyTorch, it looks like it may be better to use
DLPackwhich defines a protocol that is compatible with numpy, PyTorch, Jax, etc.See: https://dmlc.github.io/dlpack/latest/index.html