/*xla*/ 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);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  server_def.set_task_index(1);

  std::unique_ptr<tensorflow::GrpcServer> worker_server;
  ASSERT_TRUE(tensorflow::GrpcServer::Create(
                  server_def, tensorflow::Env::Default(), &worker_server)
                  .ok());
  ASSERT_TRUE(worker_server->Start().ok());

  TF_Status* status = TF_NewStatus();
  TFE_ContextOptions* opts = TFE_NewContextOptions();

  }

  status_ = registry_.Register("MatMul", MatMulRegisterer);
  ASSERT_EQ(errors::OK, status_.code()) << status_.message();

  bool transpose_a_vals[] = {false, false, true, true};
  bool transpose_b_vals[] = {false, true, false, true};
  float dA_vals[4][9] = {{24, 15, 6, 24, 15, 6, 24, 15, 6},
                         {18, 15, 12, 18, 15, 12, 18, 15, 12},
                         {24, 24, 24, 15, 15, 15, 6, 6, 6},

    // false, it is safe to keep these flag values as is.
    tensorflow::MarkForCompilationPassFlags* flags =
        tensorflow::GetMarkForCompilationPassFlags();
    flags->tf_xla_cpu_global_jit = true;
    flags->tf_xla_min_cluster_size = 1;
  } else {
    optimizer_options->set_global_jit_level(tensorflow::OptimizerOptions::OFF);
  }
}

unsigned char TF_SetXlaEnableLazyCompilation(unsigned char enable) {

  for (const auto& attr : node_def.attr()) {
    if (attr.first == "dtype") {
      if (attr.second.type() == tensorflow::DT_INT32) {
        found_dtype = true;
      } else {
        return false;
      }
    } else if (attr.first == "shape") {
      found_shape = true;
    }
  }
  return found_dtype && found_shape;
}

bool IsScalarConst(const tensorflow::NodeDef& node_def, int v) {

bool IsValidStrideCompactRowMajorData(int64_t* shape_arr, int64_t* stride_arr,
                                      int ndim) {
  bool valid = true;
  int64_t expected_stride = 1;
  for (int i = ndim - 1; i >= 0; --i) {
    // Empty tensors are always compact regardless of strides.
    if (shape_arr[i] == 0) return true;
    // Note that dimensions with size=1 can have any stride.
    if (shape_arr[i] != 1 && stride_arr[i] != expected_stride) {

                                   TensorShape::MaxDimensions(), ".");
  }
  TensorShapeProto proto;
  if (num_dims < 0) {
    proto.set_unknown_rank(true);
  } else {
    for (int d = 0; d < num_dims; ++d) {
      proto.add_dim()->set_size(dims[d]);
    }
  }

  forward_op_->attrs.Set(attr_name, proto);
  return op_->SetAttrShape(attr_name, dims, num_dims);

  std::unique_ptr<tensorflow::GrpcServer> worker_server1;
  ASSERT_TRUE(tensorflow::GrpcServer::Create(
                  server_def, tensorflow::Env::Default(), &worker_server1)
                  .ok());
  ASSERT_TRUE(worker_server1->Start().ok());

  server_def.set_task_index(2);
  std::unique_ptr<tensorflow::GrpcServer> worker_server2;
  ASSERT_TRUE(tensorflow::GrpcServer::Create(

                         status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  // Create a variable handle (uninitialized to start) placed on the parallel
  // device.
  std::function<void(Variable*)> variable_deleter = [&](Variable* to_delete) {
    to_delete->Destroy(context, status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

      TFE_NewContextOptions(), TFE_DeleteContextOptions);
  std::unique_ptr<TF_Buffer, decltype(&TF_DeleteBuffer)> config(
      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());