Blenderbot (#7418)

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

tests/test_modeling_common.py

@@ -752,6 +752,10 @@ class ModelTesterMixin:
 
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
+        def set_nan_tensor_to_zero(t):
+            t[t != t] = 0
+            return t
+
         def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
             with torch.no_grad():
                 tuple_output = model(**tuple_inputs, return_dict=False, **additional_kwargs)
@@ -765,7 +769,9 @@ class ModelTesterMixin:
                         return
                     else:
                         self.assertTrue(
-                            torch.allclose(tuple_object, dict_object, atol=1e-5),
+                            torch.allclose(
+                                set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
+                            ),
                             msg=f"Tuple and dict output are not equal. Difference: {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`: {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}.",
                         )