fixed nccl init

This commit is contained in:
William Falcon
2019-07-12 15:59:12 -04:00
parent 91b869d043
commit d99b121379
+2 -1
View File
@@ -382,7 +382,8 @@ class Trainer(TrainerIO):
root_node = os.environ['SLURM_NODELIST'].split(' ')[0]
os.environ['MASTER_ADDR'] = root_node
os.environ['MASTER_PORT'] = f'{port}'
dist.init_process_group("nccl", rank=self.proc_rank, world_size=self.world_size)
# dist.init_process_group("nccl", rank=self.proc_rank, world_size=self.world_size)
dist.init_process_group("nccl")
def __run_pretrain_routine(self, model):
"""