[fix] Add barrier to accelerator's teardown

CHANGELOG.md

@@ -230,6 +230,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
+- Added a barrier in the accelerator `teardown` to synchronize processes before execution finishes ([#6814](https://github.com/PyTorchLightning/pytorch-lightning/pull/6814))
+
+
 - 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))
 
 

pytorch_lightning/accelerators/accelerator.py

@@ -140,8 +140,11 @@ def teardown(self) -> None:
         """
         This method is called to teardown the training process.
         It is the right place to release memory and free other ressources.
+
+        By default we add a barrier here to synchronize processes before returning
+        control back to the caller.
         """
-        pass
+        self.barrier("teardown")
 
     def batch_to_device(self, batch: Any, device: Optional[torch.device] = None) -> Any:
         """Moves the batch to the correct device.

tests/trainer/test_trainer.py

@@ -1923,3 +1923,32 @@ def test_model_in_correct_mode_during_stages(tmpdir, accelerator, num_processes)
     trainer.validate(model)
     trainer.test(model)
     trainer.predict(model, model.val_dataloader())
+
+
+class TestDummyModelForCheckpoint(BoringModel):
+
+    def validation_step(self, batch, batch_idx):
+        output = self.layer(batch)
+        loss = self.loss(batch, output)
+        self.log('x', loss)
+
+    def validation_epoch_end(self, outputs) -> None:
+        pass
+
+
+@RunIf(skip_windows=True)
+def test_fit_test_synchronization(tmpdir):
+    """Test that the trainer synchronizes processes before returning control back to the caller. """
+    tutils.set_random_master_port()
+    model = TestDummyModelForCheckpoint()
+    checkpoint = ModelCheckpoint(dirpath=tmpdir, monitor='x', mode='min', save_top_k=1)
+    trainer = Trainer(
+        default_root_dir=tmpdir,
+        max_epochs=2,
+        accelerator='ddp_cpu',
+        num_processes=2,
+        callbacks=[checkpoint],
+    )
+    trainer.fit(model)
+    assert os.path.exists(checkpoint.best_model_path), f'Could not find checkpoint at rank {trainer.global_rank}'
+    trainer.test()