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) {

    // on to one TFE_Op per device and just call TFE_ResetOp to avoid parsing
    // device names repeatedly.
    std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> const_op(
        TFE_NewOp(context, "Const", status), TFE_DeleteOp);
    if (TF_GetCode(status) != TF_OK) return nullptr;
    TFE_OpSetDevice(const_op.get(), underlying_devices_[device_index].c_str(),
                    status);

// (3) Implementation detail: Avoid use of NodeBuilder/NodeDefBuilder since
//     the additional sanity checks there seem unnecessary;
typedef struct TFE_Op TFE_Op;

TF_CAPI_EXPORT extern TFE_Op* TFE_NewOp(TFE_Context* ctx,
                                        const char* op_or_function_name,
                                        TF_Status* status);
TF_CAPI_EXPORT extern void TFE_DeleteOp(TFE_Op* op);