Fix training_args.py barrier for torch_xla (#12464)

torch_xla currently has its own synchronization primitives, so use
xm.rendezvous(tag) instead.

src/transformers/training_args.py

@@ -1027,13 +1027,19 @@ class TrainingArguments:
                 if not is_main_process:
                     # tell all replicas to wait
                     logger.debug(f"{self.process_index}: waiting for the {main_process_desc} to perform {desc}")
-                    torch.distributed.barrier()
+                    if is_torch_tpu_available():
+                        xm.rendezvous(desc)
+                    else:
+                        torch.distributed.barrier()
                 yield
             finally:
                 if is_main_process:
                     # the wait is over
                     logger.debug(f"{self.process_index}: {main_process_desc} completed {desc}, releasing all replicas")
-                    torch.distributed.barrier()
+                    if is_torch_tpu_available():
+                        xm.rendezvous(desc)
+                    else:
+                        torch.distributed.barrier()
         else:
             yield