Remove legacy dead code in DDP script launch (#11678)

Co-authored-by: Jirka <[email protected]>
Co-authored-by: Carlos Mocholi <[email protected]>

Author: 3 people

pytorch_lightning/strategies/ddp.py

@@ -57,7 +57,7 @@
     sync_ddp_if_available,
 )
 from pytorch_lightning.utilities.enums import _StrategyType
-from pytorch_lightning.utilities.exceptions import DeadlockDetectedException, MisconfigurationException
+from pytorch_lightning.utilities.exceptions import DeadlockDetectedException
 from pytorch_lightning.utilities.seed import reset_seed
 from pytorch_lightning.utilities.types import STEP_OUTPUT
 
@@ -211,13 +211,6 @@ def _call_children_scripts(self):
         else:  # Script called as `python -m a.b.c`
             command = [sys.executable, "-m", __main__.__spec__.name] + sys.argv[1:]
 
-        # the visible devices tell us how many GPUs we want to use.
-        # when the trainer script was called the device has already been scoped by the time
-        # code reaches this point. so, to call the scripts, we need to leave cuda visible devices alone
-        # but forward the GPUs selected via environment variables
-        if self.parallel_devices is None:
-            raise MisconfigurationException("you selected (distribute_backend = ddp) but did not set Trainer(gpus=?)")
-
         os.environ["WORLD_SIZE"] = f"{self.num_processes * self.num_nodes}"
 
         self.interactive_ddp_procs = []