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/lilt/modeling_lilt.py

@@ -763,6 +763,7 @@ class LiltModel(LiltPreTrainedModel):
         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/longformer/modeling_longformer.py

@@ -1707,6 +1707,7 @@ class LongformerModel(LongformerPreTrainedModel):
         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/luke/modeling_luke.py

@@ -1116,6 +1116,7 @@ class LukeModel(LukePreTrainedModel):
         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/lxmert/modeling_lxmert.py

@@ -923,6 +923,7 @@ class LxmertModel(LxmertPreTrainedModel):
         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/m2m_100/modeling_m2m_100.py

@@ -774,6 +774,7 @@ class M2M100Encoder(M2M100PreTrainedModel):
         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/marian/modeling_marian.py

@@ -745,6 +745,7 @@ class MarianEncoder(MarianPreTrainedModel):
         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/markuplm/modeling_markuplm.py

@@ -877,6 +877,7 @@ class MarkupLMModel(MarkupLMPreTrainedModel):
         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/mega/modeling_mega.py

@@ -1533,6 +1533,7 @@ class MegaModel(MegaPreTrainedModel):
         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()
             device = input_ids.device
         elif inputs_embeds is not None:

src/transformers/models/megatron_bert/modeling_megatron_bert.py

@@ -933,6 +933,7 @@ class MegatronBertModel(MegatronBertPreTrainedModel):
         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/mobilebert/modeling_mobilebert.py

@@ -867,6 +867,7 @@ class MobileBertModel(MobileBertPreTrainedModel):
         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/mpnet/modeling_mpnet.py

@@ -533,6 +533,7 @@ class MPNetModel(MPNetPreTrainedModel):
         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/mra/modeling_mra.py

@@ -988,6 +988,7 @@ class MraModel(MraPreTrainedModel):
         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/nezha/modeling_nezha.py

@@ -950,6 +950,7 @@ class NezhaModel(NezhaPreTrainedModel):
         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/nllb_moe/modeling_nllb_moe.py

@@ -1105,6 +1105,7 @@ class NllbMoeEncoder(NllbMoePreTrainedModel):
         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/nystromformer/modeling_nystromformer.py

@@ -599,6 +599,7 @@ class NystromformerModel(NystromformerPreTrainedModel):
         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/openai/modeling_openai.py

@@ -452,6 +452,7 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
         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/pegasus/modeling_pegasus.py

@@ -757,6 +757,7 @@ class PegasusEncoder(PegasusPreTrainedModel):
         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/pegasus_x/modeling_pegasus_x.py

@@ -1013,6 +1013,7 @@ class PegasusXEncoder(PegasusXPreTrainedModel):
         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/qdqbert/modeling_qdqbert.py

@@ -928,6 +928,7 @@ class QDQBertModel(QDQBertPreTrainedModel):
         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()
             batch_size, seq_length = input_shape
         elif inputs_embeds is not None:

src/transformers/models/realm/modeling_realm.py

@@ -1063,6 +1063,7 @@ class RealmBertModel(RealmPreTrainedModel):
         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]