with `mask_zero=True` to automatically infer a correct padding mask.
        *   Added a `use_causal_mask` call time argument to the layer. Passing
            `use_causal_mask=True` will compute a causal attention mask, and
            optionally combine it with any `attention_mask` passed in directly
            when calling the layer.

-  extern __typeof (name) aliasname __attribute__ ((weak, alias (#name)));
+  extern __typeof (name) aliasname __attribute__ ((weak, alias (#name))) \
+    __attribute_copy__ (name);
 
 /* Same as WEAK_ALIAS, but mark symbol as hidden.  */
 # define weak_hidden_alias(name, aliasname) \
   _weak_hidden_alias (name, aliasname)
 # define _weak_hidden_alias(name, aliasname) \
   extern __typeof (name) aliasname \

}

}  // end extern "C"

namespace tensorflow {

void RecordMutation(TF_Graph* graph, const TF_Operation& op,
                    const char* mutation_type) {
  // If any session has already run this node_id, mark this session as
  // unrunnable.
  for (auto it : graph->sessions) {
    mutex_lock session_lock(it.first->mu);
    if (it.first->last_num_graph_nodes > op.node.id()) {
      it.second = strings::StrCat(

// custom devices will fail.
//
// `device_name` must not name an existing physical or custom device. It must
// follow the format:
//
//    /job:<name>/replica:<replica>/task:<task>/device:<type>:<device_num>
//
// If the device is successfully registered, `status` is set to TF_OK. Otherwise
// the device is not usable. In case of a bad status, `device.delete_device` is

    ask_for_var: string for how to ask for user input.
    var_default: default value string.

  Returns:
    string value for var_name
  """
  var = environ_cp.get(var_name)
  if not var:
    var = get_input(ask_for_var)
    print('\n')
  if not var:
    var = var_default
  return var


def prompt_loop_or_load_from_env(environ_cp,

    TFE_Context* ctx, const char* task, int64_t timeout_in_ms,
    TF_Status* status) {
  tensorflow::EagerContext* context =
      tensorflow::ContextFromInterface(tensorflow::unwrap(ctx));
  auto collective_executor_handle = context->GetCollectiveExecutorHandle();
  tensorflow::Notification done;
  collective_executor_handle->get()->remote_access()->CheckPeerHealth(
      task, timeout_in_ms, [&done, status](const Status& s) {

// gradients available initially).
//
// Then we repeatedly pop an entry from the stack, run its backprop, and update
// the gradients of its inputs. Once we have computed all gradients for a single
// input we can mark this input as done, and this can trigger adding an entry to
// the stack if all outputs of that entry are now done.
//
// When the stack is empty we have gradients for all tensors we're interested
// in.

namespace {

    // During execution:
    //   * task:0 processes main function `VariableAddFunctionWithGraphError`
    //     and places the 'read0_maybe_with_graph_error' op on task:2
    //   * task:0 partitions the main function with a subgraph containing
    //     'read0_maybe_with_graph_error' sent to task:2
    //   * task:2 graph pass reports an error when it sees

                           int64_t init_timeout_in_ms);

tensorflow::ServerDef ReplaceTaskInServerDef(
    const tensorflow::ServerDef& server_def, int task_index);

void ReplaceTaskInServerDef(tensorflow::ServerDef* server_def, int task_index,
                            const std::string& host, int port);

std::vector<std::string> ListDeviceNames(TFE_Context* ctx);

// A Server instance encapsulates a set of devices and a Session target that
// can participate in distributed training. A server belongs to a cluster
// (specified by a ClusterSpec), and corresponds to a particular task in a
// named job. The server can communicate with any other server in the same
// cluster.

// In-process TensorFlow server.
typedef struct TF_Server TF_Server;