debug

pytorch_lightning/core/optimizer.py

@@ -206,8 +206,7 @@ class LightningOptimizer:
         else:
             # make sure to call optimizer_closure when accumulating
             with trainer.profiler.profile("closure"):
-                with trainer.train_loop.block_ddp_sync_behaviour():
-                    closure()
+                closure()
 
     def __repr__(self):
         groups = [

pytorch_lightning/trainer/training_loop.py

@@ -752,15 +752,7 @@ class TrainLoop:
 
     @contextmanager
     def block_ddp_sync_behaviour(self):
-        """
-        Blocks ddp sync gradients behaviour on backwards pass.
-        This is useful for skipping sync when accumulating gradients, reducing communication overhead
-        Returns: context manager with sync behaviour off
-        """
-        if self.trainer.accelerator_backend is not None:
-            yield self.trainer.accelerator_backend.block_ddp_plugin_sync_behaviour()
-        else:
-            yield
+        yield
 
     def _process_closure_result(
         self, batch_outputs: list, opt_idx: int