AbstractOperation* TapeOperation::GetBackingOperation() { return parent_op_; }
Status TapeOperation::Execute(absl::Span<AbstractTensorHandle*> retvals,
                              int* num_retvals) {
  TF_RETURN_IF_ERROR(parent_op_->Execute(retvals, num_retvals));
  for (int i = 0; i < *num_retvals; i++) {
    // TODO(srbs): Manage refcount of ForwardOperation's inputs/outputs.
    forward_op_.outputs.push_back(retvals[i]);
  }

  TFE_OpSetDevice(matmul, remote_device_name, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TF_Tensor* t = TFE_TensorHandleResolve(retvals[0], status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

                       NodeDef* call_def) {
  Graph* graph = graph_ptr->get();
  const int num_args = input_permutation->size();
  const int num_retvals = output_permutation->size();

  std::vector<Node*> args;
  std::vector<Node*> retvals;
  args.reserve(num_args);
  retvals.reserve(num_retvals);
  for (Node* n : graph->nodes()) {
    if (n->type_string() == "_Arg") {
      // Check if this is a guaranteed constant.

  ASSERT_EQ(errors::OK, s.code()) << s.message();
  int num_retvals = 1;
  std::vector<AbstractTensorHandle*> outputs(1);
  GradientRegistry registry;
  s = RegisterGradients(&registry);
  ASSERT_EQ(errors::OK, s.code()) << s.message();
  auto tape = std::make_unique<Tape>(/*persistent=*/false);
  s = Execute(check_numerics_op.get(), ctx.get(), absl::MakeSpan(outputs),
              &num_retvals, &forward_op, tape.get(), registry);

    TFE_OpSetDevice(op.get(), device_name, status.get());
    if (TF_GetCode(status.get()) != TF_OK) return;

    TFE_TensorHandle* result_handles;
    int num_retvals = 1;
    TFE_Execute(op.get(), &result_handles, &num_retvals, status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_INVALID_ARGUMENT)
        << TF_Message(status.get());
  }
}

  }
  TFE_OpSetDevice(op.get(), device, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;

  TFE_TensorHandle* result_handle;
  int num_retvals = 1;
  TFE_Execute(op.get(), &result_handle, &num_retvals, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  return TensorHandlePtr(result_handle);
}

template <typename value_type>

    // The input handles should never change since they have been mirrored.
    ASSERT_FALSE(remote_arg->HasLocalMirror(nullptr));
  }

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

  // TODO(gjn): Add support for waiting on async local mirrors
  if (!remote && !async && !remote_func_outputs) {

// tensors in `retvals`.
//
// 'retvals' must point to a pre-allocated array of TFE_TensorHandle* and
// '*num_retvals' should be set to the size of this array. It is an error if
// the size of 'retvals' is less than the number of outputs. This call sets
// *num_retvals to the number of outputs.
//
// If async execution is enabled, the call may simply enqueue the execution

  Status AddInput(AbstractTensorHandle* input) override;
  Status AddInputList(absl::Span<AbstractTensorHandle* const> inputs) override;
  Status Execute(absl::Span<AbstractTensorHandle*> retvals,
                 int* num_retvals) override;
  Status SetAttrString(const char* attr_name, const char* data,
                       size_t length) override;
  Status SetAttrInt(const char* attr_name, int64_t value) override;

                 int* num_retvals) override {
    auto* tf_opdesc = op_.release();
    if (tf_opdesc == nullptr) {
      return errors::InvalidArgument("AbstractOp is incomplete.");
    }
    TF_Status* s = TF_NewStatus();
    auto* operation = TF_FinishOperation(tf_opdesc, s);
    TF_RETURN_IF_ERROR(StatusFromTF_Status(s));
    TF_DeleteStatus(s);
    *num_retvals = TF_OperationNumOutputs(operation);