TFE_NewOp(context.get(), function_name, status.get()), TFE_DeleteOp);
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  TFE_OpSetDevice(op.get(), device_name, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  TFE_OpAddInput(op.get(), parallel_value.get(), status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  ASSERT_EQ(errors::OK, s.code()) << s.message();
  if (isa<TracingOperation>(check_numerics_op.get())) {
    s = dyn_cast<TracingOperation>(check_numerics_op.get())
            ->SetOpName("check_numerics");
    ASSERT_EQ(errors::OK, s.code()) << s.message();
  }
  s = AddInput(check_numerics_op.get(), t.get(), &forward_op);
  ASSERT_EQ(errors::OK, s.code()) << s.message();

      TFE_NewOp(context.get(), "AssertAndCollective", status.get()),
      TFE_DeleteOp);
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  std::unique_ptr<ParallelTensor> reduced_values =
      parallel_device.ScalarsFromSequence<float>({1.0, 2.0}, context.get(),
                                                 status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  TF_OutputListSetNumOutputs(mm_outputs, 1, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
  TF_ExecuteOperation(fn_op, 1, &input_t, mm_outputs, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  ASSERT_EQ(1, TF_OutputListNumOutputs(mm_outputs));

  TF_Operation* a = Placeholder(graph, s, "A");
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  TF_Operation* b = Placeholder(graph, s, "B");
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  TF_Operation* two = ScalarConst(2, graph, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  TF_Operation* plus2 = Add(a, two, graph, s, "plus2");
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  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;
  ASSERT_EQ(sizeof(sum), TF_TensorByteSize(t));
  memcpy(&sum, TF_TensorData(t), TF_TensorByteSize(t));

                          &numerical_grad_raw);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    numerical_grad.reset(numerical_grad_raw);
  }

  TF_Tensor* numerical_tensor;
  s = GetValue(numerical_grad.get(), &numerical_tensor);
  ASSERT_EQ(errors::OK, s.code()) << s.message();
  auto num_elem_numerical = TF_TensorElementCount(numerical_tensor);
  ASSERT_EQ(num_elem_numerical, num_grad);

  TFE_Op* op = TFE_NewOp(ctx, "ReadVariableOp", status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpSetAttrType(op, "dtype", TF_FLOAT);
  TFE_OpAddInput(op, var, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  int num_retvals = 1;
  TFE_Execute(op, out_value, &num_retvals, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteOp(op);