fixed nccl init

2019-07-12 16:16:46 -04:00 · 2019-07-12 16:16:46 -04:00 · 7e37f68a5b
parent 960937ebe9
commit 7e37f68a5b
1 changed files with 2 additions and 0 deletions
--- a/pytorch_lightning/models/trainer.py
+++ b/pytorch_lightning/models/trainer.py
@ -385,6 +385,8 @@ class Trainer(TrainerIO):
            port = 12910
            os.environ['MASTER_PORT'] = f'{port}'

+        sleep(self.proc_rank * 2)
+
        root_node = os.environ['SLURM_NODELIST'].split(' ')[0]
        os.environ['MASTER_ADDR'] = root_node
        dist.init_process_group("nccl", rank=self.proc_rank, world_size=self.world_size)