diff --git a/pytorch_lightning/models/trainer.py b/pytorch_lightning/models/trainer.py
index 1f82dd1ed3..434e50609b 100644
--- a/pytorch_lightning/models/trainer.py
+++ b/pytorch_lightning/models/trainer.py
@@ -407,12 +407,11 @@ class Trainer(TrainerIO):
 
         # forward pass
         # return a scalar value and a dic with tqdm metrics
-        output = self.model(data_batch, batch_nb)
-
-        # when DP, we need to aggregate the scalars we received as outputs
-        # use mean as the reduce function
         if self.data_parallel:
+            output = self.model(data_batch, batch_nb)
             output = reduce_distributed_output(output, len(self.data_parallel_device_ids))
+        else:
+            output = self.model.training_step(data_batch, batch_nb)
 
         model_specific_tqdm_metrics_dic = output['tqdm_metrics']
         loss = output['loss']