auto ReluGradModel = BuildGradModel(ReluModel, registry_);

  float X_vals[] = {1.0f, 2.0f, 3.0f, -5.0f, -4.0f, -3.0f, 2.0f, 10.0f, -1.0f};
  int64_t X_dims[] = {3, 3};
  AbstractTensorHandlePtr X;
  {
    AbstractTensorHandle* X_raw;
    status_ = TestTensorHandleWithDims<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), X_vals, X_dims, 2, &X_raw);

TEST(CAPI, TestFromProto) {
  Tensor t_cc(DT_FLOAT, TensorShape({2, 3}));
  t_cc.flat<float>().setConstant(1.0);
  tensorflow::TensorProto t_proto;
  t_cc.AsProtoField(&t_proto);

  TF_Buffer* t_buffer = TF_NewBuffer();
  TF_CHECK_OK(MessageToBuffer(t_proto, t_buffer));

  const int num_bytes = 6 * sizeof(float);
  float* values =
      reinterpret_cast<float*>(tensorflow::cpu_allocator()->AllocateRaw(

      TF_AbstractTensorGetEagerTensor(result, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
  TF_Tensor* result_tensor = TFE_TensorHandleResolve(result_t, status.get());
  float* result_value = static_cast<float*>(TF_TensorData(result_tensor));
  EXPECT_EQ(*result_value, 4.0);

  TF_DeleteTensor(result_tensor);
  TF_DeleteAbstractTensor(result);
  TF_DeleteOutputList(o);
  TF_DeleteExecutionContext(ctx);

  ASSERT_EQ(1, num_retvals);
  EXPECT_EQ(TF_FLOAT, TFE_TensorHandleDataType(value_handle));
  EXPECT_EQ(0, TFE_TensorHandleNumDims(value_handle, status));
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  float value = 0.0f;
  TF_Tensor* t = TFE_TensorHandleResolve(value_handle, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  ASSERT_EQ(sizeof(float), TF_TensorByteSize(t));

    TF_Tensor* resolved = TFE_TensorHandleResolve(on_host, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    const float* resolved_data =
        static_cast<const float*>(TF_TensorData(resolved));
    EXPECT_EQ(0, memcmp(m_float, resolved_data, 4 * sizeof(float)));
    TF_DeleteTensor(resolved);
    TFE_DeleteTensorHandle(copy_aliased);  // Note that this will delete copy.
    TFE_DeleteTensorHandle(on_host);

}
Status TapeOperation::SetAttrFloatList(const char* attr_name,
                                       const float* values, int num_values) {
  forward_op_.attrs.Set(attr_name,
                        gtl::ArraySlice<const float>(values, num_values));
  return parent_op_->SetAttrFloatList(attr_name, values, num_values);
}
Status TapeOperation::SetAttrIntList(const char* attr_name,

  virtual void GetNameAttrList(
      tensorflow::NameAttrList* name_and_attrs) const = 0;

  virtual bool GetInt(absl::string_view, int64_t* result) const = 0;
  virtual bool GetFloat(absl::string_view attr_name, float* result) const = 0;
  virtual bool GetBool(absl::string_view attr_name, bool* result) const = 0;
  virtual bool GetType(absl::string_view attr_name, DataType* result) const = 0;
  virtual Status GetTypeList(

        forward path for data of type `tf.float32` but not for data of type
        `tf.float64` (for which there was no GPU implementation). In this
        current release, GPU support for other floating-point types
        (`tf.float16`, `tf.float64`, `tf.complex64`, and `tf.complex128`) has
        been added for this op. If you were relying on the determinism of the

    status_ = TestScalarTensorHandle<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), 2.0f, &x_raw);
    ASSERT_EQ(errors::OK, status_.code()) << status_.message();
    x.reset(x_raw);
  }

  AbstractTensorHandlePtr y;
  {
    AbstractTensorHandle* y_raw = nullptr;
    status_ = TestScalarTensorHandle<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), 2.0f, &y_raw);

    ASSERT_EQ(num_elem_numerical, num_elem_analytical);

    float* dnumerical = new float[num_elem_numerical]{0};
    memcpy(&dnumerical[0], TF_TensorData(numerical_tensor),
           TF_TensorByteSize(numerical_tensor));
    float* danalytical = new float[num_elem_analytical]{0};
    memcpy(&danalytical[0], TF_TensorData(analytical_tensor),
           TF_TensorByteSize(analytical_tensor));