// 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) {

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(

  tensorflow::FunctionDef fdef;
  status->status = tensorflow::GraphToFunctionDef(
      fn_body->graph, fn_name, append_hash_to_fn_name != 0,
      /*set_stateful_from_nodes=*/true,
      /*copy_placeholder_attrs_from_nodes=*/true, body_nodes, input_tensors,
      output_tensors, output_names_vec, control_output_nodes,
      control_output_names_vec, description, &fdef);
  if (TF_GetCode(status) != TF_OK) {
    return nullptr;

                                /*transpose_a = */ true,
                                /*transpose_b = */ true, name_grad_A.c_str()));

      TF_RETURN_IF_ERROR(MatMul(ctx, upstream_grad, A.get(), &matmul_B_output,
                                /*transpose_a = */ true,
                                /*transpose_b = */ true, name_grad_B.c_str()));
    }

    // Gradient for A

        //
        // TODO(allenl): We should have an async API that works with the
        // parallel device.
        device_(device),
        executor_(
            TFE_NewExecutor(is_async, /*enable_streaming_enqueue=*/true,
                            /*in_flight_nodes_limit=*/in_flight_nodes_limit)),
        op_(nullptr),
        thread_(tensorflow::Env::Default()->StartThread(
            tensorflow::ThreadOptions(), "parallel_device_execute",

      opts->session_options.options,
      static_cast<tensorflow::ContextDevicePlacementPolicy>(
          opts->device_placement_policy),
      opts->async, device_mgr.release(),
      /*device_mgr_owned*/ true, std::move(r),
      /*cluster_flr=*/nullptr,
      /*collective_executor_mgr=*/nullptr,
      /*run_eager_op_as_function=*/opts->run_eager_op_as_function,
      /*jit_compile_rewrite=*/opts->jit_compile_rewrite);

static bool register_tracing = [] {
  RegisterTracingEngineFactory("graphdef", GraphTracingFactory);
  SetDefaultTracingEngine("graphdef").IgnoreError();
  return true;
}();

}  // namespace graph
}  // namespace tracing

    if (dev_type == device_type) {
      *device_name = dev_name;
      LOG(INFO) << "Found " << device_type << " device " << *device_name;
      TF_DeleteDeviceList(devices);
      return true;
    }
  }
  TF_DeleteDeviceList(devices);
  return false;
}

tensorflow::ServerDef GetServerDef(const string& job_name, int num_tasks) {
  tensorflow::ServerDef server_def;