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/modeling_utils.py

@@ -74,7 +74,7 @@ from .utils import (
     replace_return_docstrings,
 )
 from .utils.hub import convert_file_size_to_int, get_checkpoint_shard_files
-from .utils.import_utils import ENV_VARS_TRUE_VALUES, is_sagemaker_mp_enabled
+from .utils.import_utils import ENV_VARS_TRUE_VALUES, is_sagemaker_mp_enabled, is_torch_fx_proxy
 from .utils.quantization_config import BitsAndBytesConfig
 from .utils.versions import require_version_core
 
@@ -3527,6 +3527,11 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         """
         Shows a one-time warning if the input_ids appear to contain padding and no attention mask was given.
         """
+
+        # Skip the check during tracing.
+        if is_torch_fx_proxy(input_ids) or torch.jit.is_tracing():
+            return
+
         if (attention_mask is not None) or (self.config.pad_token_id is None):
             return
 

src/transformers/models/albert/modeling_albert.py

@@ -701,6 +701,7 @@ class AlbertModel(AlbertPreTrainedModel):
         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/align/modeling_align.py

@@ -1267,6 +1267,7 @@ class AlignTextModel(AlignPreTrainedModel):
         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/altclip/modeling_altclip.py

@@ -1305,8 +1305,8 @@ class AltRobertaModel(AltCLIPPreTrainedModel):
         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/bert/modeling_bert.py

@@ -966,8 +966,8 @@ class BertModel(BertPreTrainedModel):
         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/bert_generation/modeling_bert_generation.py

@@ -768,6 +768,7 @@ class BertGenerationEncoder(BertGenerationPreTrainedModel):
         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/big_bird/modeling_big_bird.py

@@ -2035,6 +2035,7 @@ class BigBirdModel(BigBirdPreTrainedModel):
         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/bigbird_pegasus/modeling_bigbird_pegasus.py

@@ -1857,6 +1857,7 @@ class BigBirdPegasusEncoder(BigBirdPegasusPreTrainedModel):
         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/blenderbot/modeling_blenderbot.py

@@ -732,6 +732,7 @@ class BlenderbotEncoder(BlenderbotPreTrainedModel):
         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/blenderbot_small/modeling_blenderbot_small.py

@@ -729,6 +729,7 @@ class BlenderbotSmallEncoder(BlenderbotSmallPreTrainedModel):
         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/blip/modeling_blip_text.py

@@ -717,6 +717,7 @@ class BlipTextModel(BlipTextPreTrainedModel):
         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
             device = input_ids.device

src/transformers/models/bridgetower/modeling_bridgetower.py

@@ -1118,8 +1118,8 @@ class BridgeTowerTextModel(BridgeTowerPreTrainedModel):
         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/camembert/modeling_camembert.py

@@ -841,8 +841,8 @@ class CamembertModel(CamembertPreTrainedModel):
         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/canine/modeling_canine.py

@@ -1126,6 +1126,7 @@ class CanineModel(CaninePreTrainedModel):
         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/chinese_clip/modeling_chinese_clip.py

@@ -1208,6 +1208,7 @@ class ChineseCLIPTextModel(ChineseCLIPPreTrainedModel):
         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/clap/modeling_clap.py

@@ -1853,8 +1853,8 @@ class ClapTextModel(ClapPreTrainedModel):
         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/codegen/modeling_codegen.py

@@ -460,6 +460,7 @@ class CodeGenModel(CodeGenPreTrainedModel):
         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])
             batch_size = input_ids.shape[0]

src/transformers/models/convbert/modeling_convbert.py

@@ -818,6 +818,7 @@ class ConvBertModel(ConvBertPreTrainedModel):
         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/ctrl/modeling_ctrl.py

@@ -397,6 +397,7 @@ class CTRLModel(CTRLPreTrainedModel):
         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])
             batch_size = input_ids.shape[0]

src/transformers/models/data2vec/modeling_data2vec_text.py

@@ -790,8 +790,8 @@ class Data2VecTextModel(Data2VecTextPreTrainedModel):
         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: