Fix DPP + SyncBN (Lightning-AI#6838)

* Fix DPP + SyncBN

Ensure that model is already on correct GPU before applying SyncBN conversion

* Fix order of SyncBN for ddp_spawn

Author: BloodAxe

pytorch_lightning/plugins/training_type/ddp.py

@@ -257,12 +257,12 @@ def pre_dispatch(self):
         self.dist.rank = self.global_rank
         self.dist.device = self.root_device
 
-        if self.sync_batchnorm:
-            self.model = self.configure_sync_batchnorm(self.model)
-
         # move the model to the correct device
         self.model_to_device()
 
+        if self.sync_batchnorm:
+            self.model = self.configure_sync_batchnorm(self.model)
+
         self.configure_ddp()
 
         self.barrier()

pytorch_lightning/plugins/training_type/ddp_spawn.py

@@ -148,12 +148,12 @@ def new_process(self, process_idx, trainer, mp_queue):
         self.dist.rank = self.global_rank
         self.dist.device = self.root_device
 
-        if self.sync_batchnorm:
-            self.model = self.configure_sync_batchnorm(self.model)
-
         # move the model to the correct device
         self.model_to_device()
 
+        if self.sync_batchnorm:
+            self.model = self.configure_sync_batchnorm(self.model)
+
         self.configure_ddp()
 
         self.barrier()

tests/trainer/test_dataloaders.py

@@ -721,8 +721,6 @@ def __len__(self):
     assert has_len(dataloader)
     assert has_iterable_dataset(dataloader)
     trainer = Trainer(default_root_dir=tmpdir, max_steps=3)
-    with pytest.warns(UserWarning, match='Your `IterableDataset` has `__len__` defined.'):
-        trainer.validate(model, val_dataloaders=[dataloader])
     with pytest.warns(UserWarning, match='Your `IterableDataset` has `__len__` defined.'):
         trainer.fit(model, train_dataloader=dataloader, val_dataloaders=[dataloader])
     with pytest.warns(UserWarning, match='Your `IterableDataset` has `__len__` defined.'):
@@ -735,7 +733,6 @@ def __len__(self):
     assert not has_len(dataloader)
     assert has_iterable_dataset(dataloader)
     trainer = Trainer(default_root_dir=tmpdir, max_steps=3)
-    trainer.validate(model, val_dataloaders=dataloader)
     trainer.fit(model, train_dataloader=dataloader, val_dataloaders=[dataloader])
     trainer.test(model, test_dataloaders=dataloader)
     trainer.predict(model, dataloaders=dataloader)