Compatibility with new_group() API

apex/parallel/__init__.py

 import torch
 
+# Backward compatibility hack around
+# https://github.com/pytorch/pytorch/pull/14767
+if hasattr(torch.distributed, 'get_default_group'):
+    group_creator = torch.distributed.get_default_group
+else:
+    group_creator = torch.distributed.new_group
+
 from .distributed import DistributedDataParallel, Reducer
 try:
     import syncbn

apex/parallel/optimized_sync_batchnorm_kernel.py

@@ -18,7 +18,7 @@ class SyncBatchnormFunction(Function):
                 if process_group:
                     world_size = torch.distributed.get_world_size(process_group)
                 else:
-                    process_group = torch.distributed.get_default_group()
+                    process_group = group_creator()
                     world_size = torch.distributed.get_world_size()
                 mean_all = torch.empty(world_size, mean.size(0), dtype=mean.dtype, device=mean.device)
                 var_all = torch.empty(world_size, var.size(0), dtype=var.dtype, device=var.device)

apex/parallel/sync_batchnorm.py

@@ -65,7 +65,7 @@ class SyncBatchNorm(_BatchNorm):
             if self.process_group:
                 world_size = torch.distributed.get_world_size(process_group)
             else:
-                process_group = torch.distributed.get_default_group()
+                process_group = group_creator()
                 world_size = torch.distributed.get_world_size()
             self.num_batches_tracked += 1
             with torch.no_grad():