We currently use at::Tensor directly and don't worry about the distinction between Variable and Tensor in c10d. This came up when working on adding support for reduction of sparse tensors, where Variable instances that come in through pybind are mixed with temporary Tensor instances created on the c10d side.

One solution here is to use torch::autograd::Variable (later to be fully merged into at::Tensor) throughout. This means adding a dependency on torch/csrc/autograd in c10d and changing all temporaries to be created with factory functions in the torch namespace.

A temporary workaround is to unbox and rebox Variable instances at the Python/C++ boundary.