// the documentation for TFE_RegisterCustomDevice for more information).
//
// Tensors may be copied on or off the device explicitly using
// TPUReplicatedInput and TPUReplicatedOutput respectively. For example, with
// two component devices, running `x = TPUReplicatedInput(inputs=[a, b])` on the
// parallel device creates a parallel tensor `x` with `a` on the first of

  // TODO(allenl): We should remove "TPU" from these op names at the very least,
  // or consider other ways of packing/unpacking parallel tensors.
  if (operation_name == std::string("TPUReplicatedInput")) {
    // Special-cased operation for packing per-device tensors into one parallel
    // tensor.
    if (inputs.size() != parallel_device.num_underlying_devices()) {
      std::string message(absl::StrCat(

    TFE_Context* context,
    const std::array<TFE_TensorHandle*, num_replicas>& components,
    const char* device, TF_Status* status) {
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> op(
      TFE_NewOp(context, "TPUReplicatedInput", status), TFE_DeleteOp);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpSetAttrInt(op.get(), "N", num_replicas);
  for (int i = 0; i < num_replicas; ++i) {

    ASSERT_TRUE(TF_GetCode(status.get()) == TF_INVALID_ARGUMENT)
        << TF_Message(status.get());
  }

  {
    // Try to pass a ParallelTensor to TPUReplicatedInput
    std::array<TFE_TensorHandle*, 1> correct_components{value_one.get()};
    TensorHandlePtr combined_value = CreatePerDeviceValues(
        context.get(), correct_components, device_name, status.get());