[fix] Add a cluster environment teardown to clean up environment state

CHANGELOG.md

@@ -9,6 +9,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Added
 
+- Added a `teardown` hook to `ClusterEnvironment` ([#6942](https://github.com/PyTorchLightning/pytorch-lightning/pull/6942))
+
+
 - Added utils for NaN/Inf detection for gradients and parameters ([#6834](https://github.com/PyTorchLightning/pytorch-lightning/pull/6834/))
 
 
@@ -196,6 +199,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
+- 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))
+
+
 - Set better defaults for `rank_zero_only.rank` when training is launched with SLURM and torchelastic:
     * Support SLURM and torchelastic global rank environment variables ([#5715](https://github.com/PyTorchLightning/pytorch-lightning/pull/5715))
     * Remove hardcoding of local rank in accelerator connector ([#6878](https://github.com/PyTorchLightning/pytorch-lightning/pull/6878))
@@ -243,7 +249,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed `--gpus` default for parser returned by `Trainer.add_argparse_args` ([#6898](https://github.com/PyTorchLightning/pytorch-lightning/pull/6898))
 
 
-- Fixed pickle error checker to now check for `pickle.PickleError` to catch all pickle errors ([#6917](https://github.com/PyTorchLightning/pytorch-lightning/pull/6917)) 
+- Fixed pickle error checker to now check for `pickle.PickleError` to catch all pickle errors ([#6917](https://github.com/PyTorchLightning/pytorch-lightning/pull/6917))
 
 
 - Fixed `AttributeError` for `require_backward_grad_sync` when running manual optimization with sharded plugin ([#6915](https://github.com/PyTorchLightning/pytorch-lightning/pull/6915))

pytorch_lightning/plugins/environments/cluster_environment.py

@@ -52,3 +52,7 @@ def local_rank(self) -> int:
     @abstractmethod
     def node_rank(self) -> int:
         """ The rank (index) of the node on which the current process runs. """
+
+    def teardown(self) -> None:
+        """ Clean up any state set after execution finishes. """
+        pass

pytorch_lightning/plugins/environments/lightning_environment.py

@@ -68,6 +68,10 @@ def node_rank(self) -> int:
         group_rank = os.environ.get("GROUP_RANK", 0)
         return int(os.environ.get("NODE_RANK", group_rank))
 
+    def teardown(self) -> None:
+        if "WORLD_SIZE" in os.environ:
+            del os.environ["WORLD_SIZE"]
+
 
 def find_free_network_port() -> int:
     """

pytorch_lightning/plugins/training_type/ddp.py

@@ -280,9 +280,8 @@ def pre_dispatch(self):
 
         self.barrier()
 
-    def post_dispatch(self):
-        if "WORLD_SIZE" in os.environ:
-            del os.environ["WORLD_SIZE"]
+    def post_dispatch(self) -> None:
+        self.cluster_environment.teardown()
 
     def barrier(self, *args, **kwargs):
         if torch_distrib.is_initialized():

tests/plugins/environments/test_lightning_environment.py

@@ -55,3 +55,14 @@ def test_random_master_port():
     assert isinstance(port, int)
     # repeated calls do not generate a new port number
     assert env.master_port() == port
+
+
+@mock.patch.dict(os.environ, {
+    "WORLD_SIZE": "1",
+})
+def test_teardown():
+    """ Test that the GROUP_RANK substitutes NODE_RANK. """
+    env = LightningEnvironment()
+    assert "WORLD_SIZE" in os.environ
+    env.teardown()
+    assert "WORLD_SIZE" not in os.environ