// 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

  TF_SetStatus(status, TF_OK, "");
  if (retvals[0] != nullptr) {
    TFE_DeleteTensorHandle(retvals[0]);
  }
  retvals[0] = nullptr;
  TFE_Execute(matmul2, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status));
  TF_Tensor* t = TFE_TensorHandleResolve(retvals[0], status);
  EXPECT_EQ(TF_OK, TF_GetCode(status));
  TF_DeleteTensor(t);

  TFE_TensorHandle* retvals[1] = {nullptr};
  int num_retvals = 1;
  TFE_Execute(func, &retvals[0], &num_retvals, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  ASSERT_EQ(1, num_retvals);
  TF_Tensor* t = TFE_TensorHandleResolve(retvals[0], status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteTensorHandle(retvals[0]);
  float sum = 0;

// Execute the operation defined by 'op' and return handles to computed
// 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.
//

  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);
  float product[4] = {0};

    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
    TFE_TensorHandle* retvals[1];
    int num_retvals = 1;
    TFE_Execute(shape_op, &retvals[0], &num_retvals, status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

    device_type = TFE_TensorHandleDeviceType(retvals[0], status.get());
    ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  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);

  auto* retval_task0 =
      TFE_TensorHandleCopyToDevice(retvals[0], ctx, local_device_name, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

                     int* num_retvals, ForwardOperation* forward_op_,
                     Tape* tape, const GradientRegistry& registry) {
  TF_RETURN_IF_ERROR(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]);
  }

                                 ForwardOperation*);

// Make the call to `Tape::RecordOperation`.
absl::Status Execute(AbstractOperation*, AbstractContext*,
                     absl::Span<AbstractTensorHandle*> retvals,
                     int* num_retvals, ForwardOperation*, Tape*,
                     const GradientRegistry&);

}  // namespace internal
}  // namespace gradients
}  // namespace tensorflow

                       ImmediateExecutionTensorHandle** retvals,
                       int* num_retvals) override {
    std::vector<TFE_TensorHandle*> outputs(*num_retvals);
    TF_Status status;
    device_.execute(tensorflow::wrap(op), num_retvals, outputs.data(), &status,
                    info_);
    if (status.status.ok()) {
      for (int i = 0; i < *num_retvals; ++i) {
        retvals[i] = tensorflow::unwrap(outputs[i]);