[testing] ensure concurrent pytest workers use a unique port for torch.dist (#12166)

* ensure concurrent pytest workers use a unique port for torch.distributed.launch

* reword

tests/test_trainer_distributed.py

@@ -16,7 +16,12 @@ import sys
 from typing import Dict
 
 from transformers import EvalPrediction, HfArgumentParser, TrainingArguments, is_torch_available
-from transformers.testing_utils import TestCasePlus, execute_subprocess_async, require_torch_multi_gpu
+from transformers.testing_utils import (
+    TestCasePlus,
+    execute_subprocess_async,
+    get_torch_dist_unique_port,
+    require_torch_multi_gpu,
+)
 from transformers.utils import logging
 
 
@@ -64,6 +69,7 @@ class TestTrainerDistributed(TestCasePlus):
         distributed_args = f"""
             -m torch.distributed.launch
             --nproc_per_node={torch.cuda.device_count()}
+            --master_port={get_torch_dist_unique_port()}
             {self.test_file_dir}/test_trainer_distributed.py
         """.split()
         output_dir = self.get_auto_remove_tmp_dir()