[gradient checkpointing] lower tolerance for test (huggingface#652)

* lowe tolerance

* put model in eval mode

Author: patil-suraj

tests/test_models_unet.py

@@ -271,7 +271,7 @@ def prepare_init_args_and_inputs_for_common(self):
     def test_gradient_checkpointing(self):
         # enable deterministic behavior for gradient checkpointing
         init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
-        model = self.model_class(**init_dict)
+        model = self.model_class(**init_dict).eval()
         model.to(torch_device)
 
         assert not model.is_gradient_checkpointing and model.training