Add warning for missing attention mask when pad tokens are detected (#25345)

* Add attention mask and pad token warning to many of the models

* Remove changes under examples/research_projects

These files are not maintained by HG.

* Skip the warning check during torch.fx or JIT tracing

* Switch ordering for the warning and input shape assignment

This ordering is a little cleaner for some of the cases.

* Add missing line break in one of the files

src/transformers/models/reformer/modeling_reformer.py

@@ -2031,6 +2031,7 @@ class ReformerModel(ReformerPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()  # noqa: F841
             device = input_ids.device
         elif inputs_embeds is not None:

src/transformers/models/rembert/modeling_rembert.py

@@ -837,6 +837,7 @@ class RemBertModel(RemBertPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]

src/transformers/models/roberta/modeling_roberta.py

@@ -788,8 +788,8 @@ class RobertaModel(RobertaPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
-            input_shape = input_ids.size()
             self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
+            input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]
         else:

src/transformers/models/roberta_prelayernorm/modeling_roberta_prelayernorm.py

@@ -790,6 +790,7 @@ class RobertaPreLayerNormModel(RobertaPreLayerNormPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]

src/transformers/models/roc_bert/modeling_roc_bert.py

@@ -993,6 +993,7 @@ class RoCBertModel(RoCBertPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]

src/transformers/models/roformer/modeling_roformer.py

@@ -874,6 +874,7 @@ class RoFormerModel(RoFormerPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]

src/transformers/models/splinter/modeling_splinter.py

@@ -700,6 +700,7 @@ class SplinterModel(SplinterPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]

src/transformers/models/squeezebert/modeling_squeezebert.py

@@ -596,6 +596,7 @@ class SqueezeBertModel(SqueezeBertPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]

src/transformers/models/tapas/modeling_tapas.py

@@ -932,6 +932,7 @@ class TapasModel(TapasPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]

src/transformers/models/vilt/modeling_vilt.py

@@ -798,6 +798,7 @@ class ViltModel(ViltPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]

src/transformers/models/visual_bert/modeling_visual_bert.py

@@ -773,6 +773,7 @@ class VisualBertModel(VisualBertPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]

src/transformers/models/xglm/modeling_xglm.py

@@ -606,6 +606,7 @@ class XGLMModel(XGLMPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
             input_ids = input_ids.view(-1, input_shape[-1])
         elif inputs_embeds is not None:

src/transformers/models/xlm_roberta/modeling_xlm_roberta.py

@@ -790,8 +790,8 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
-            input_shape = input_ids.size()
             self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
+            input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]
         else:

src/transformers/models/xlm_roberta_xl/modeling_xlm_roberta_xl.py

@@ -756,8 +756,8 @@ class XLMRobertaXLModel(XLMRobertaXLPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
-            input_shape = input_ids.size()
             self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
+            input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]
         else:

src/transformers/models/xmod/modeling_xmod.py

@@ -886,6 +886,7 @@ class XmodModel(XmodPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]

src/transformers/models/yoso/modeling_yoso.py

@@ -790,6 +790,7 @@ class YosoModel(YosoPreTrainedModel):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]

templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/modeling_{{cookiecutter.lowercase_modelname}}.py

@@ -845,6 +845,7 @@ class {{cookiecutter.camelcase_modelname}}Model({{cookiecutter.camelcase_modelna
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]
@@ -2271,6 +2272,7 @@ class {{cookiecutter.camelcase_modelname}}Encoder({{cookiecutter.camelcase_model
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
             input_ids = input_ids.view(-1, input_shape[-1])
         elif inputs_embeds is not None: