EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TF_DeleteStatus(status);

  string api_def_overwrites = R"(op: <
  graph_op_name: "TestCApi"
  summary: "New summary"
>
)";
  status = TF_NewStatus();
  TF_ApiDefMapPut(api_def_map, api_def_overwrites.c_str(),
                  api_def_overwrites.size(), status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TF_DeleteStatus(status);

                      "BiasAdd");
}

class CppGradients
    : public ::testing::TestWithParam<std::tuple<const char*, bool, bool>> {
 protected:
  void SetUp() override {
    TF_StatusPtr status(TF_NewStatus());
    TF_SetTracingImplementation(std::get<0>(GetParam()), status.get());
    status_ = StatusFromTF_Status(status.get());
    ASSERT_EQ(errors::OK, status_.code()) << status_.message();

// while `op` is alive.
TF_CAPI_EXPORT extern const TFE_OpAttrs* TFE_OpGetAttrs(const TFE_Op* op);
// Add attributes in `attrs` to `op`.
//
// Does not overwrite or update existing attributes, but adds new ones.
TF_CAPI_EXPORT extern void TFE_OpAddAttrs(TFE_Op* op, const TFE_OpAttrs* attrs);

// Serialize `attrs` as a tensorflow::NameAttrList protocol buffer (into `buf`),

        output->Ref();
      }
    }
  }

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    AbstractTensorHandle* upstream_grad = grad_outputs[0];
    AbstractTensorHandle* activations = forward_outputs_[0];

    // Calculate Grad
    std::string name = "relu_grad";

 public:
  explicit GraphFunction(FunctionDef fdef);
  ~GraphFunction() override;

  // GraphFunction maybe stay alive for the duration of the returned
  // FunctionDef.
  Status GetFunctionDef(const FunctionDef** fdef) override;

  // Returns a shared reference to the wrapped function.
  absl::StatusOr<core::RefCountPtr<FunctionRecord>> GetFunctionRecord()
      override {
    return func_record_.GetNewRef();
  }

  // Builds a tensor filled with ones with the same shape and dtype as `t`.
  Status BuildOnesLike(const TapeTensor& t,
                       AbstractTensorHandle** result) const override;

  // Looks up the ID of a Gradient.
  int64_t TensorId(AbstractTensorHandle* tensor) const override;

  // Converts a Gradient to a TapeTensor.
  TapeTensor TapeTensorFromGradient(AbstractTensorHandle* g) const override;

  outputs[0] = temp_outputs[1];
  temp_outputs[0]->Unref();
  return absl::OkStatus();
}

class CppGradients
    : public ::testing::TestWithParam<std::tuple<const char*, bool, bool>> {
 protected:
  void SetUp() override {
    TF_StatusPtr status(TF_NewStatus());
    TF_SetTracingImplementation(std::get<0>(GetParam()), status.get());
    status_ = StatusFromTF_Status(status.get());
    ASSERT_EQ(errors::OK, status_.code()) << status_.message();

    return absl::OkStatus();
  }
  const string& Name() const override { return op_type_; }
  const string& DeviceName() const override { return device_name_; }

  Status SetDeviceName(const char* name) override {
    // TODO(srbs): Implement this.
    device_name_ = name;
    return absl::OkStatus();
  }

  Status AddInput(AbstractTensorHandle* input) override {
    GraphTensor* t = dyn_cast<GraphTensor>(input);

                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    // TODO(b/161805092): Support broadcasting.

    DCHECK(grad_outputs[0]);
    grad_inputs[0] = grad_outputs[0];
    grad_inputs[1] = grad_outputs[0];

    grad_inputs[0]->Ref();
    grad_inputs[1]->Ref();
    return absl::OkStatus();
  }
  ~AddGradientFunction() override {}
};

class ExpGradientFunction : public GradientFunction {

        methods_(methods) {}

  ~CAPICustomDeviceTensorHandle() override { methods_.deallocator(data_); }
  void* DevicePointer() const override { return data_; }
  Status NumDims(int* num_dims) const override {
    TF_Status s;
    *num_dims = methods_.num_dims(data_, &s);
    return s.status;
  }
  Status Dim(int dim_index, int64_t* dim) const override {
    TF_Status s;
    *dim = methods_.dim(data_, dim_index, &s);