AbstractTensorHandlePtr sum_dims;
  {
    vector<int32_t> vals(num_dims_out);
    int64_t vals_shape[] = {num_dims_out};
    Range(&vals, 0, num_dims_out);
    AbstractTensorHandle* sum_dims_raw = nullptr;
    TF_RETURN_IF_ERROR(TestTensorHandleWithDims<int32_t, TF_INT32>(
        ctx, vals.data(), vals_shape, 1, &sum_dims_raw));
    sum_dims.reset(sum_dims_raw);
  }

  // Reduce sum the output on all dimensions.

  TFE_TensorHandle* var_handle0 = TestVariable(ctx, 1.0, dev0_name);
  EXPECT_NE(var_handle0, nullptr);
  TFE_TensorHandle* var_handle1 = TestVariable(ctx, 2.0, dev1_name);
  EXPECT_NE(var_handle1, nullptr);

  TFE_TensorHandle* value_handle = nullptr;
  ReadVariable(ctx, var_handle1, &value_handle);
  CheckTFE_TensorHandleHasFloats(value_handle, {2});
  TFE_DeleteTensorHandle(value_handle);

    const char* device_name = TFE_OpGetDevice(matmul, status);
    ASSERT_TRUE(strstr(device_name, "GPU:0") != nullptr);

    TFE_OpSetDevice(matmul, "CPU:0", status);
    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
    device_name = TFE_OpGetDevice(matmul, status);
    ASSERT_TRUE(strstr(device_name, "CPU:0") != nullptr);
  }

  TFE_DeleteOp(matmul);
  TFE_DeleteTensorHandle(m);
  TFE_DeleteContext(ctx);

  TF_Output input{arg, 0};
  TF_Output output{id, 0};
  TF_Function* fn =
      TF_GraphToFunction(function_graph, "ident", 0, 1, &id, 1, &input, 1,
                         &output, nullptr, nullptr, "test", status);
  ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
  TF_DeleteGraph(function_graph);
  TFE_ContextOptions* opts = TFE_NewContextOptions();

  AbstractTensorHandlePtr x;
  {
    AbstractTensorHandle* x_raw = nullptr;
    absl::Status s =
        TestScalarTensorHandle<float, TF_FLOAT>(ctx_.get(), 2.0f, &x_raw);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    x.reset(x_raw);
  }

  AbstractTensorHandlePtr y;
  {
    AbstractTensorHandle* y_raw = nullptr;
    absl::Status s =

  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  auto* arg0 = TF_AddFunctionParameter(graph_ctx, TF_FLOAT, {-1, nullptr}, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
  auto* arg1 = TF_AddFunctionParameter(graph_ctx, TF_FLOAT, {-1, nullptr}, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  // Create a first "Add" computing `arg0 + arg1`.
  TF_AbstractTensor* add_output1;
  {

 private:
  const AbstractContextKind kind_;
};

namespace internal {
struct AbstractContextDeleter {
  void operator()(AbstractContext* p) const {
    if (p != nullptr) {
      p->Release();
    }
  }
};
}  // namespace internal

using AbstractContextPtr =
    std::unique_ptr<AbstractContext, internal::AbstractContextDeleter>;

}  // namespace tensorflow

// gradients.
//
// `ZerosLike` is not expected to be called and returns a nullptr. The creation
// of default zeros grads is handled by the `DefaultGradientFunction` registered
// for each op.
// TODO(srbs): We need to define `ZerosLike` here to keep the compiler happy.
// Figure out a way to avoid this.
// TODO(srbs): Should ZerosLike check-fail instead of returning nullptr?
class TapeTensor {
 public:

  if (remote) {
    TFE_OpSetDevice(func, task1_name, status);
    ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  }

  TFE_TensorHandle* retvals[1] = {nullptr};
  int num_retvals = 1;
  TFE_Execute(func, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  ASSERT_EQ(1, num_retvals);
  TFE_DeleteOp(func);

// Returns the operation in the graph with `oper_name`. Returns nullptr if
// no operation found.
TF_CAPI_EXPORT extern TF_Operation* TF_GraphOperationByName(
    TF_Graph* graph, const char* oper_name);

// Iterate through the operations of a graph.  To use:
// size_t pos = 0;
// TF_Operation* oper;
// while ((oper = TF_GraphNextOperation(graph, &pos)) != nullptr) {
//   DoSomethingWithOperation(oper);
// }