// Copy host -> device. (Empty tensors don't have backing buffers.)
    // Manually allocate memory so we can allocate as much memory as the device
    // requires (as given by GetByteSizeRequirement). This avoids
    // XlaTransferManager having to reallocate the device buffer later if
    // XlaTransferManager is used.
    VLOG(1) << "Constant output tensor on device";

    TF_RETURN_IF_ERROR(