introduce logger.warning_once and use it for grad checkpointing code (#21804)

* logger.warning_once

* style

src/transformers/models/altclip/modeling_altclip.py

@@ -638,7 +638,7 @@ class AltRobertaEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/bart/modeling_bart.py

@@ -1085,7 +1085,7 @@ class BartDecoder(BartPretrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/bert/modeling_bert.py

@@ -585,7 +585,7 @@ class BertEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/bert_generation/modeling_bert_generation.py

@@ -395,7 +395,7 @@ class BertEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/big_bird/modeling_big_bird.py

@@ -1606,7 +1606,7 @@ class BigBirdEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py

@@ -2265,7 +2265,7 @@ class BigBirdPegasusDecoder(BigBirdPegasusPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/biogpt/modeling_biogpt.py

@@ -557,7 +557,7 @@ class BioGptModel(BioGptPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/blenderbot/modeling_blenderbot.py

@@ -1016,7 +1016,7 @@ class BlenderbotDecoder(BlenderbotPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/blenderbot_small/modeling_blenderbot_small.py

@@ -1012,7 +1012,7 @@ class BlenderbotSmallDecoder(BlenderbotSmallPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/bloom/modeling_bloom.py

@@ -757,7 +757,7 @@ class BloomModel(BloomPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/bridgetower/modeling_bridgetower.py

@@ -769,7 +769,7 @@ class BridgeTowerTextEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/camembert/modeling_camembert.py

@@ -516,7 +516,7 @@ class CamembertEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/chinese_clip/modeling_chinese_clip.py

@@ -901,7 +901,7 @@ class ChineseCLIPTextEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/clap/modeling_clap.py

@@ -1588,7 +1588,7 @@ class ClapTextEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/codegen/modeling_codegen.py

@@ -548,7 +548,7 @@ class CodeGenModel(CodeGenPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
                         "`use_cache=False`..."
                     )

src/transformers/models/data2vec/modeling_data2vec_text.py

@@ -502,7 +502,7 @@ class Data2VecTextEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/decision_transformer/modeling_decision_transformer.py

@@ -609,7 +609,7 @@ class DecisionTransformerGPT2Model(DecisionTransformerGPT2PreTrainedModel):
 
         if self.gradient_checkpointing and self.training:
             if use_cache:
-                logger.warning(
+                logger.warning_once(
                     "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                 )
                 use_cache = False

src/transformers/models/electra/modeling_electra.py

@@ -563,7 +563,7 @@ class ElectraEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/ernie/modeling_ernie.py

@@ -498,7 +498,7 @@ class ErnieEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/esm/modeling_esm.py

@@ -597,7 +597,7 @@ class EsmEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
                         "`use_cache=False`..."
                     )

src/transformers/models/git/modeling_git.py

@@ -444,7 +444,7 @@ class GitEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/gpt2/modeling_gpt2.py

@@ -853,7 +853,7 @@ class GPT2Model(GPT2PreTrainedModel):
 
         if self.gradient_checkpointing and self.training:
             if use_cache:
-                logger.warning(
+                logger.warning_once(
                     "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                 )
                 use_cache = False

src/transformers/models/gpt_neo/modeling_gpt_neo.py

@@ -589,7 +589,7 @@ class GPTNeoModel(GPTNeoPreTrainedModel):
 
         if self.gradient_checkpointing and self.training:
             if use_cache:
-                logger.warning(
+                logger.warning_once(
                     "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                 )
                 use_cache = False

src/transformers/models/gptj/modeling_gptj.py

@@ -653,7 +653,7 @@ class GPTJModel(GPTJPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/imagegpt/modeling_imagegpt.py

@@ -812,7 +812,7 @@ class ImageGPTModel(ImageGPTPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/layoutlm/modeling_layoutlm.py

@@ -479,7 +479,7 @@ class LayoutLMEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/led/modeling_led.py

@@ -2136,7 +2136,7 @@ class LEDDecoder(LEDPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/m2m_100/modeling_m2m_100.py

@@ -1055,7 +1055,7 @@ class M2M100Decoder(M2M100PreTrainedModel):
 
                 if self.gradient_checkpointing and self.training:
                     if use_cache:
-                        logger.warning(
+                        logger.warning_once(
                             "`use_cache=True` is incompatible with gradient checkpointing. Setting"
                             " `use_cache=False`..."
                         )

src/transformers/models/marian/modeling_marian.py

@@ -1020,7 +1020,7 @@ class MarianDecoder(MarianPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/markuplm/modeling_markuplm.py

@@ -641,7 +641,7 @@ class MarkupLMEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/mbart/modeling_mbart.py

@@ -1069,7 +1069,7 @@ class MBartDecoder(MBartPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing`. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/megatron_bert/modeling_megatron_bert.py

@@ -544,7 +544,7 @@ class MegatronBertEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/mt5/modeling_mt5.py

@@ -1008,7 +1008,7 @@ class MT5Stack(MT5PreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/mvp/modeling_mvp.py

@@ -1212,7 +1212,7 @@ class MvpDecoder(MvpPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/nezha/modeling_nezha.py

@@ -571,7 +571,7 @@ class NezhaEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/opt/modeling_opt.py

@@ -671,7 +671,7 @@ class OPTDecoder(OPTPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/pegasus/modeling_pegasus.py

@@ -1070,7 +1070,7 @@ class PegasusDecoder(PegasusPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/pegasus_x/modeling_pegasus_x.py

@@ -1311,7 +1311,7 @@ class PegasusXDecoder(PegasusXPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/plbart/modeling_plbart.py

@@ -1048,7 +1048,7 @@ class PLBartDecoder(PLBartPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/prophetnet/modeling_prophetnet.py

@@ -1572,7 +1572,7 @@ class ProphetNetDecoder(ProphetNetPreTrainedModel):
 
         if self.gradient_checkpointing and self.training:
             if use_cache:
-                logger.warning(
+                logger.warning_once(
                     "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                 )
                 use_cache = False

src/transformers/models/qdqbert/modeling_qdqbert.py

@@ -575,7 +575,7 @@ class QDQBertEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/realm/modeling_realm.py

@@ -578,7 +578,7 @@ class RealmEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/rembert/modeling_rembert.py

@@ -536,7 +536,7 @@ class RemBertEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/roberta/modeling_roberta.py

@@ -502,7 +502,7 @@ class RobertaEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/roberta_prelayernorm/modeling_roberta_prelayernorm.py

@@ -504,7 +504,7 @@ class RobertaPreLayerNormEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/roc_bert/modeling_roc_bert.py

@@ -636,7 +636,7 @@ class RoCBertEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/roformer/modeling_roformer.py

@@ -573,7 +573,7 @@ class RoFormerEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/speecht5/modeling_speecht5.py

@@ -1692,7 +1692,7 @@ class SpeechT5Decoder(SpeechT5PreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/splinter/modeling_splinter.py

@@ -451,7 +451,7 @@ class SplinterEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/switch_transformers/modeling_switch_transformers.py

@@ -1057,7 +1057,7 @@ class SwitchTransformersStack(SwitchTransformersPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/t5/modeling_t5.py

@@ -1037,7 +1037,7 @@ class T5Stack(T5PreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/time_series_transformer/modeling_time_series_transformer.py

@@ -1471,7 +1471,7 @@ class TimeSeriesTransformerDecoder(TimeSeriesTransformerPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/trajectory_transformer/modeling_trajectory_transformer.py

@@ -543,7 +543,7 @@ class TrajectoryTransformerModel(TrajectoryTransformerPreTrainedModel):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/xlm_prophetnet/modeling_xlm_prophetnet.py

@@ -1595,7 +1595,7 @@ class XLMProphetNetDecoder(XLMProphetNetPreTrainedModel):
 
         if self.gradient_checkpointing and self.training:
             if use_cache:
-                logger.warning(
+                logger.warning_once(
                     "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                 )
                 use_cache = False

src/transformers/models/xlm_roberta/modeling_xlm_roberta.py

@@ -503,7 +503,7 @@ class XLMRobertaEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/xlm_roberta_xl/modeling_xlm_roberta_xl.py

@@ -492,7 +492,7 @@ class XLMRobertaXLEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/models/xmod/modeling_xmod.py

@@ -566,7 +566,7 @@ class XmodEncoder(nn.Module):
 
             if self.gradient_checkpointing and self.training:
                 if use_cache:
-                    logger.warning(
+                    logger.warning_once(
                         "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                     )
                     use_cache = False

src/transformers/utils/logging.py

@@ -14,6 +14,8 @@
 # limitations under the License.
 """ Logging utilities."""
 
+
+import functools
 import logging
 import os
 import sys
@@ -281,6 +283,21 @@ def warning_advice(self, *args, **kwargs):
 logging.Logger.warning_advice = warning_advice
 
 
+@functools.lru_cache(None)
+def warning_once(self, *args, **kwargs):
+    """
+    This method is identical to `logger.warning()`, but will emit the warning with the same message only once
+
+    Note: The cache is for the function arguments, so 2 different callers using the same arguments will hit the cache.
+    The assumption here is that all warning messages are unique across the code. If they aren't then need to switch to
+    another type of cache that includes the caller frame information in the hashing function.
+    """
+    self.warning(*args, **kwargs)
+
+
+logging.Logger.warning_once = warning_once
+
+
 class EmptyTqdm:
     """Dummy tqdm which doesn't do anything."""