fixed nccl init

2019-07-12 15:55:28 -04:00 · 2019-07-12 15:55:28 -04:00 · 6c02afefca
parent 4696e12641
commit 6c02afefca
1 changed files with 4 additions and 1 deletions
--- a/pytorch_lightning/models/trainer.py
+++ b/pytorch_lightning/models/trainer.py
@ -379,7 +379,10 @@ class Trainer(TrainerIO):
        :param tries:
        :return:
        """
-        sleep(self.proc_rank*2)
+        # hack to get nccl to stop throwing error... seems to be an nccl race condition
+        if self.proc_rank > 0:
+            sleep(10.0)
+
        root_node = os.environ['SLURM_NODELIST'].split(' ')[0]
        os.environ['MASTER_ADDR'] = root_node
        os.environ['MASTER_PORT'] = f'{port}'