//
        // TODO(allenl): There may be smarter ways to do this copy in some
        // cases, i.e. with a collective broadcast. We'll need to be careful
        // about things that are taken as inputs on the host or on their
        // existing device (for multi-device functions).
        std::unique_ptr<ParallelTensor> parallel_tensor(
            parallel_device.CopyToParallelDevice(

//
// Design decisions made:
// * Backing store for tensor memory has an associated deallocation
//   function.  This deallocation function will point to client code
//   for tensors populated by the client.  So the client can do things
//   like shadowing a numpy array.
// * We do not provide TF_OK since it is not strictly necessary and we
//   are not optimizing for convenience.
// * We make assumption that one session has one graph.  This should be