diff --git a/pytorch_lightning/models/trainer.py b/pytorch_lightning/models/trainer.py
index bb63e35b..3daa759d 100644
--- a/pytorch_lightning/models/trainer.py
+++ b/pytorch_lightning/models/trainer.py
@@ -283,7 +283,6 @@ class Trainer(TrainerIO):
                 model, self.optimizers, opt_level=self.amp_level,
             )
             self.optimizers = optimizers
-            model.trainer = self
 
         # when using gpus, first thing we do is spawn a new process between each worker
         # applies to single gpu, multi-gpu and multi-nodes
@@ -385,6 +384,8 @@ class Trainer(TrainerIO):
         if self.on_gpu:
             ref_model = model.module
 
+        ref_model.trainer = self
+
         # set local properties on the model
         ref_model.on_gpu = self.on_gpu