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:

src/transformers/models/deberta/modeling_deberta.py

@@ -959,6 +959,7 @@ class DebertaModel(DebertaPreTrainedModel):
         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/deberta_v2/modeling_deberta_v2.py

@@ -1059,6 +1059,7 @@ class DebertaV2Model(DebertaV2PreTrainedModel):
         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/decision_transformer/modeling_decision_transformer.py

@@ -530,6 +530,7 @@ class DecisionTransformerGPT2Model(DecisionTransformerGPT2PreTrainedModel):
         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/distilbert/modeling_distilbert.py

@@ -589,6 +589,7 @@ class DistilBertModel(DistilBertPreTrainedModel):
         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/dpr/modeling_dpr.py

@@ -562,6 +562,7 @@ class DPRQuestionEncoder(DPRPretrainedQuestionEncoder):
         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]
@@ -650,6 +651,7 @@ class DPRReader(DPRPretrainedReader):
         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/electra/modeling_electra.py

@@ -868,6 +868,7 @@ class ElectraModel(ElectraPreTrainedModel):
         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/ernie/modeling_ernie.py

@@ -895,6 +895,7 @@ class ErnieModel(ErniePreTrainedModel):
         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/esm/modeling_esm.py

@@ -876,6 +876,7 @@ class EsmModel(EsmPreTrainedModel):
         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/funnel/modeling_funnel.py

@@ -966,6 +966,7 @@ class FunnelBaseModel(FunnelPreTrainedModel):
         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]
@@ -1041,6 +1042,7 @@ class FunnelModel(FunnelPreTrainedModel):
         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/git/modeling_git.py

@@ -1213,6 +1213,7 @@ class GitModel(GitPreTrainedModel):
         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/gpt2/modeling_gpt2.py

@@ -776,6 +776,7 @@ class GPT2Model(GPT2PreTrainedModel):
         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/gpt_bigcode/modeling_gpt_bigcode.py

@@ -560,6 +560,7 @@ class GPTBigCodeModel(GPTBigCodePreTrainedModel):
         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/gpt_neo/modeling_gpt_neo.py

@@ -525,6 +525,7 @@ class GPTNeoModel(GPTNeoPreTrainedModel):
         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/gpt_neox/modeling_gpt_neox.py

@@ -571,6 +571,7 @@ class GPTNeoXModel(GPTNeoXPreTrainedModel):
         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/gpt_neox_japanese/modeling_gpt_neox_japanese.py

@@ -508,6 +508,7 @@ class GPTNeoXJapaneseModel(GPTNeoXJapanesePreTrainedModel):
         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/gptj/modeling_gptj.py

@@ -577,6 +577,7 @@ class GPTJModel(GPTJPreTrainedModel):
         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/ibert/modeling_ibert.py

@@ -796,6 +796,7 @@ class IBertModel(IBertPreTrainedModel):
         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/imagegpt/modeling_imagegpt.py

@@ -715,6 +715,7 @@ class ImageGPTModel(ImageGPTPreTrainedModel):
         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/layoutlm/modeling_layoutlm.py

@@ -795,6 +795,7 @@ class LayoutLMModel(LayoutLMPreTrainedModel):
         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/layoutlmv2/modeling_layoutlmv2.py

@@ -1047,6 +1047,7 @@ class LayoutLMv2ForSequenceClassification(LayoutLMv2PreTrainedModel):
         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/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]

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: