Merge pull request #156 from danpovey/ddp_master_port

Make master port configurable

egs/librispeech/asr/simple_v1/mmi_att_transformer_train.py

@@ -286,6 +286,11 @@ def get_parser():
         type=int,
         default=1,
         help='Number of GPUs for DDP training.')
+    parser.add_argument(
+        '--master-port',
+        type=int,
+        default=12354,
+        help='Master port to use for DDP training.')
     parser.add_argument(
         '--model-type',
         type=str,
@@ -362,7 +367,7 @@ def run(rank, world_size, args):
     att_rate = args.att_rate
 
     fix_random_seed(42)
-    setup_dist(rank, world_size)
+    setup_dist(rank, world_size, args.master_port)
 
     exp_dir = Path('exp-' + model_type + '-noam-mmi-att-musan-sa')
     setup_logger(f'{exp_dir}/log/log-train-{rank}')

snowfall/dist.py

@@ -3,9 +3,14 @@
 from torch import distributed as dist
 
 
-def setup_dist(rank, world_size):
+def setup_dist(rank, world_size, master_port = None):
     os.environ['MASTER_ADDR'] = 'localhost'
+<<<<<<< HEAD
     os.environ['MASTER_PORT'] = '12355'
+=======
+    os.environ['MASTER_PORT'] = ('12354' if master_port is None
+                                 else str(master_port))
+>>>>>>> 8f51e68... Make master port command-line configurable
     dist.init_process_group("nccl", rank=rank, world_size=world_size)
     torch.cuda.set_device(rank)