namespace gradients {
namespace internal {

void CompareNumericalAndAutodiffGradients(
    Model model, Model grad_model, AbstractContext* ctx,
    absl::Span<AbstractTensorHandle* const> inputs, bool use_function,
    double abs_error) {
  auto num_inputs = inputs.size();
  std::vector<AbstractTensorHandle*> outputs(num_inputs);
  auto s = RunModel(grad_model, ctx, inputs, absl::MakeSpan(outputs),
                    /*use_function=*/use_function);

  return TF_NewTensor(TF_INT32, nullptr, 0, values, num_bytes,
                      &Int32Deallocator, nullptr);
}

TF_Tensor* DoubleTensor(double v) {
  const int num_bytes = sizeof(double);
  double* values = new double[1];
  values[0] = v;
  return TF_NewTensor(TF_DOUBLE, nullptr, 0, values, num_bytes,
                      &DoubleDeallocator, nullptr);
}

TF_Tensor* FloatTensor(float v) {

  TF_DeleteStatus(s);
}

TEST(CAPI, DataTypeEnum) {
  EXPECT_EQ(TF_FLOAT, static_cast<TF_DataType>(tensorflow::DT_FLOAT));
  EXPECT_EQ(TF_DOUBLE, static_cast<TF_DataType>(tensorflow::DT_DOUBLE));
  EXPECT_EQ(TF_INT32, static_cast<TF_DataType>(tensorflow::DT_INT32));
  EXPECT_EQ(TF_UINT8, static_cast<TF_DataType>(tensorflow::DT_UINT8));

  switch (data_type) {
    case TF_DataType::TF_BOOL:
      dtype.code = DLDataTypeCode::kDLBool;
      break;
    case TF_DataType::TF_HALF:
    case TF_DataType::TF_FLOAT:
    case TF_DataType::TF_DOUBLE:
      dtype.code = DLDataTypeCode::kDLFloat;
      break;
    case TF_DataType::TF_INT8:
    case TF_DataType::TF_INT16:
    case TF_DataType::TF_INT32:
    case TF_DataType::TF_INT64:

  return StatusFromTF_Status(plugin_status.get());
}

Status ModularFileSystem::SetOption(const std::string& name,
                                    const std::vector<double>& values) {
  if (ops_->set_filesystem_configuration == nullptr) {
    return errors::Unimplemented(
        "Filesystem does not support SetConfiguration()");
  }
  if (values.empty()) {

TEST_P(UnifiedCAPI, TestBasicGraph) {
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);

  // Start a new function / execution context.
  string fn_name = "double";
  TF_ExecutionContext* graph_ctx =
      TF_CreateFunction(fn_name.c_str(), status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  auto* placeholder_t =

void CompareNumericalAndManualGradients(
    Model model, AbstractContext* ctx,
    absl::Span<AbstractTensorHandle* const> inputs, int input_index,
    float* expected_grad, int num_grad, bool use_function,
    double abs_error = 1e-2) {
  Status s;
  AbstractTensorHandlePtr numerical_grad;
  {
    AbstractTensorHandle* numerical_grad_raw;
    s = CalcNumericalGrad(ctx, model, inputs, input_index, use_function,

  TF_DeleteStatus(status);
  return th;
}

TFE_TensorHandle* DoubleTestMatrixTensorHandle(TFE_Context* ctx) {
  int64_t dims[] = {2, 2};
  double data[] = {1.0, 2.0, 3.0, 4.0};
  TF_Status* status = TF_NewStatus();
  TF_Tensor* t = TFE_AllocateHostTensor(ctx, TF_DOUBLE, &dims[0],
                                        sizeof(dims) / sizeof(int64_t), status);
  memcpy(TF_TensorData(t), &data[0], TF_TensorByteSize(t));

  buf->data_deallocator = [](void* data, size_t length) {
    tensorflow::port::Free(data);
  };
}

TFE_MonitoringBuckets* TFE_MonitoringNewExponentialBuckets(double scale,
                                                           double growth_factor,
                                                           int bucket_count) {
  return new TFE_MonitoringBuckets([scale, growth_factor, bucket_count]() {