fixed nccl init

pytorch_lightning/models/trainer.py

@@ -367,7 +367,7 @@ class Trainer(TrainerIO):
         # continue training routine
         self.__run_pretrain_routine(model)
 
-    def __init_tcp_connection(self, port=12000, tries=0):
+    def __init_tcp_connection(self, port=12975, tries=0):
         """
         Connect all procs in the world using the env:// init
         Use the first node as the root address
@@ -379,19 +379,6 @@ class Trainer(TrainerIO):
         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)
-        # if tries > 20:
-        #     raise RuntimeError('Failed to connect using 20 different ip addresses')
-        #
-        # try:
-        #     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)
-        #
-        # except RuntimeError as e:
-        #     # port taken
-        #     warnings.warn(f'port {port} taken, trying port {port}...')
-        #     self.__init_tcp_connection(port + 1, tries + 1)
 
     def __run_pretrain_routine(self, model):
         """