using tensorflow::NodeDef;

static void BoolDeallocator(void* data, size_t, void* arg) {
  delete[] static_cast<bool*>(data);
}

static void Int32Deallocator(void* data, size_t, void* arg) {
  delete[] static_cast<int32_t*>(data);
}

static void DoubleDeallocator(void* data, size_t, void* arg) {
  delete[] static_cast<double*>(data);
}

static void FloatDeallocator(void* data, size_t, void* arg) {
  delete[] static_cast<float*>(data);

}

void TF_AppendWritableFile(TF_WritableFileHandle* handle, const char* data,
                           size_t length, TF_Status* status) {
  auto* cc_file = reinterpret_cast<::tensorflow::WritableFile*>(handle);
  TF_SetStatus(status, TF_OK, "");
  ::tensorflow::Set_TF_Status_from_Status(
      status, cc_file->Append(::tensorflow::StringPiece{data, length}));
}

void TF_DeleteFile(const char* filename, TF_Status* status) {

#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/op_def.pb.h"
#include "tensorflow/core/grappler/costs/graph_properties.h"
#include "tensorflow/core/grappler/costs/op_performance_data.pb.h"
#include "tensorflow/core/grappler/grappler_item.h"
#include "tensorflow/core/grappler/optimizers/custom_graph_optimizer_registry.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/status.h"

  return absl::OkStatus();
}

Status SetAttrString(AbstractOperation* op_, const char* attr_name,
                     const char* data, size_t length,
                     ForwardOperation* forward_op_) {
  forward_op_->attrs.Set(attr_name, StringPiece(data, length));
  return op_->SetAttrString(attr_name, data, length);
}
Status SetAttrInt(AbstractOperation* op_, const char* attr_name, int64_t value,

  GetDims(theta_tensor, theta_dims.data());

  // Initialize auxilary data structures.
  vector<float> thetaPlus_data(num_elems);
  vector<float> thetaMinus_data(num_elems);
  AbstractTensorHandle* f_outputs[1];

  // Numerical Grad Check
  for (int i = 0; i < num_elems; i++) {
    // Get relative epsilon value
    float epsilon = theta_data[i] == 0 ? 1e-4 : std::abs(theta_data[i] * 1e-4);

  CHECK_OK(pjrt_client.status());
  auto c_api_client = down_cast<xla::PjRtCApiClient*>(pjrt_client->get());
  std::vector<int32_t> data(1, 0);
  xla::Shape shape = xla::ShapeUtil::MakeShape(xla::S32, {1});

  auto buffer = c_api_client->pjrt_c_client()->client->BufferFromHostBuffer(
      data.data(), shape.element_type(), shape.dimensions(),
      /*byte_strides=*/std::nullopt,

void ParallelTensorDeallocator(void* data) {
  delete reinterpret_cast<ParallelTensor*>(data);
}

// Used as an argument to TFE_NewCustomDeviceTensorHandle, for computing the
// number of dimensions of a parallel tensor.
int ParallelTensorNumDims(void* data, TF_Status* status) {
  const std::vector<int64_t>* shape;
  Status s = reinterpret_cast<ParallelTensor*>(data)->Shape(&shape);
  if (!s.ok()) {

  TF_Buffer* result = TF_NewBuffer();
  const std::string& value_str = *value;
  void* data = malloc(value_str.length());
  value_str.copy(static_cast<char*>(data), value_str.length(), 0);
  result->data = data;
  result->length = value_str.length();
  result->data_deallocator = [](void* data, size_t length) { free(data); };
  return result;
}

  TFE_Context* ctx = TFE_NewContext(opts, status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteContextOptions(opts);

  TFE_ContextSetServerDef(ctx, 0, serialized.data(), serialized.size(), status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

  const char task0_name[] = "/job:localhost/replica:0/task:0/device:CPU:0";

      TF_CreateConfig(
          /*enable_xla_compilation=*/false,
          /*gpu_memory_allow_growth=*/true, /*num_cpu_devices=*/2),
      TF_DeleteBuffer);
  TFE_ContextOptionsSetConfig(opts.get(), config->data, config->length,
                              status.get());
  std::unique_ptr<TFE_Context, decltype(&TFE_DeleteContext)> context(
      TFE_NewContext(opts.get(), status.get()), TFE_DeleteContext);