From 91b869d04300bda906306a15f018520e745a75ff Mon Sep 17 00:00:00 2001
From: William Falcon <falc@fb.com>
Date: Fri, 12 Jul 2019 15:55:28 -0400
Subject: [PATCH] fixed nccl init

---
 pytorch_lightning/models/trainer.py | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/pytorch_lightning/models/trainer.py b/pytorch_lightning/models/trainer.py
index 64fbf16b..fb08a337 100644
--- a/pytorch_lightning/models/trainer.py
+++ b/pytorch_lightning/models/trainer.py
@@ -375,7 +375,10 @@ 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}'