[fix] Fix multi-node DDP launch by using local rank instead of global rank for main process (#7061)

* Update ddp.py

* Update CHANGELOG.md

CHANGELOG.md

@@ -202,6 +202,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
+- Fixed multi-node DDP sub-process launch by using `local_rank` instead of `global_rank` for main process assertion ([#7061](https://github.com/PyTorchLightning/pytorch-lightning/pull/7061))
+
+
 - Fixed incorrect removal of `WORLD_SIZE` environment variable in DDP training when launching with torch distributed/torchelastic ([#6942](https://github.com/PyTorchLightning/pytorch-lightning/pull/6942))
 
 

pytorch_lightning/plugins/training_type/ddp.py

@@ -110,7 +110,7 @@ class DDPPlugin(ParallelPlugin):
     def _call_children_scripts(self):
 
         # bookkeeping of spawned processes
-        assert self.global_rank == 0
+        assert self.local_rank == 0
         self._check_can_spawn_children()
         self._has_spawned_children = True