Rest of the model refactors

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
Co-authored-by: Lucain Pouget <lucainp@gmail.com>

.circleci/config.yml

@@ -116,7 +116,7 @@ jobs:
                 command: pip freeze | tee installed.txt
             - store_artifacts:
                   path: ~/transformers/installed.txt
-            - run: python -c "from transformers import *" || (echo '🚨 import failed, this means you introduced unprotected imports! 🚨'; exit 1)
+            - run: python -c "from transformers import *" || (echo '🚨 import failed, this means you introduced unprotected imports! Have you added this object to the __all__ object of the module? 🚨'; exit 1)
             - run: ruff check examples tests src utils
             - run: ruff format tests src utils --check
             - run: python utils/custom_init_isort.py --check_only

Makefile

@@ -50,10 +50,9 @@ repo-consistency:
 # this target runs checks on all files
 
 quality:
-	@python -c "from transformers import *" || (echo '🚨 import failed, this means you introduced unprotected imports! 🚨'; exit 1)
+	@python -c "from transformers import *" || (echo '🚨 import failed, this means you introduced unprotected imports! Have you added this object to the __all__ object of the module? 🚨'; exit 1)
 	ruff check $(check_dirs) setup.py conftest.py
 	ruff format --check $(check_dirs) setup.py conftest.py
-	python utils/custom_init_isort.py --check_only
 	python utils/sort_auto_mappings.py --check_only
 	python utils/check_doc_toc.py
 	python utils/check_docstrings.py --check_all
@@ -62,7 +61,6 @@ quality:
 # Format source code automatically and check is there are any problems left that need manual fixing
 
 extra_style_checks:
-	python utils/custom_init_isort.py
 	python utils/sort_auto_mappings.py
 	python utils/check_doc_toc.py --fix_and_overwrite
 

src/transformers/__init__.py

@@ -1499,7 +1499,6 @@ else:
             "BertForQuestionAnswering",
             "BertForSequenceClassification",
             "BertForTokenClassification",
-            "BertLayer",
             "BertLMHeadModel",
             "BertModel",
             "BertPreTrainedModel",
@@ -1523,7 +1522,6 @@ else:
             "BigBirdForQuestionAnswering",
             "BigBirdForSequenceClassification",
             "BigBirdForTokenClassification",
-            "BigBirdLayer",
             "BigBirdModel",
             "BigBirdPreTrainedModel",
             "load_tf_weights_in_big_bird",
@@ -1642,7 +1640,6 @@ else:
             "CanineForQuestionAnswering",
             "CanineForSequenceClassification",
             "CanineForTokenClassification",
-            "CanineLayer",
             "CanineModel",
             "CaninePreTrainedModel",
             "load_tf_weights_in_canine",
@@ -1729,7 +1726,6 @@ else:
             "ConvBertForQuestionAnswering",
             "ConvBertForSequenceClassification",
             "ConvBertForTokenClassification",
-            "ConvBertLayer",
             "ConvBertModel",
             "ConvBertPreTrainedModel",
             "load_tf_weights_in_convbert",
@@ -1958,7 +1954,6 @@ else:
             "QDQBertForQuestionAnswering",
             "QDQBertForSequenceClassification",
             "QDQBertForTokenClassification",
-            "QDQBertLayer",
             "QDQBertLMHeadModel",
             "QDQBertModel",
             "QDQBertPreTrainedModel",
@@ -2210,7 +2205,6 @@ else:
             "FNetForQuestionAnswering",
             "FNetForSequenceClassification",
             "FNetForTokenClassification",
-            "FNetLayer",
             "FNetModel",
             "FNetPreTrainedModel",
         ]
@@ -2311,7 +2305,6 @@ else:
             "GPTNeoXForQuestionAnswering",
             "GPTNeoXForSequenceClassification",
             "GPTNeoXForTokenClassification",
-            "GPTNeoXLayer",
             "GPTNeoXModel",
             "GPTNeoXPreTrainedModel",
         ]
@@ -2319,7 +2312,6 @@ else:
     _import_structure["models.gpt_neox_japanese"].extend(
         [
             "GPTNeoXJapaneseForCausalLM",
-            "GPTNeoXJapaneseLayer",
             "GPTNeoXJapaneseModel",
             "GPTNeoXJapanesePreTrainedModel",
         ]
@@ -2551,7 +2543,6 @@ else:
             "LongformerForTokenClassification",
             "LongformerModel",
             "LongformerPreTrainedModel",
-            "LongformerSelfAttention",
         ]
     )
     _import_structure["models.longt5"].extend(
@@ -2584,7 +2575,6 @@ else:
             "LxmertModel",
             "LxmertPreTrainedModel",
             "LxmertVisualFeatureEncoder",
-            "LxmertXLayer",
         ]
     )
     _import_structure["models.m2m_100"].extend(
@@ -2608,7 +2598,9 @@ else:
             "Mamba2PreTrainedModel",
         ]
     )
-    _import_structure["models.marian"].extend(["MarianForCausalLM", "MarianModel", "MarianMTModel"])
+    _import_structure["models.marian"].extend(
+        ["MarianForCausalLM", "MarianModel", "MarianMTModel", "MarianPreTrainedModel"]
+    )
     _import_structure["models.markuplm"].extend(
         [
             "MarkupLMForQuestionAnswering",
@@ -2691,7 +2683,6 @@ else:
             "MobileBertForQuestionAnswering",
             "MobileBertForSequenceClassification",
             "MobileBertForTokenClassification",
-            "MobileBertLayer",
             "MobileBertModel",
             "MobileBertPreTrainedModel",
             "load_tf_weights_in_mobilebert",
@@ -2737,7 +2728,6 @@ else:
             "MPNetForQuestionAnswering",
             "MPNetForSequenceClassification",
             "MPNetForTokenClassification",
-            "MPNetLayer",
             "MPNetModel",
             "MPNetPreTrainedModel",
         ]
@@ -2827,7 +2817,6 @@ else:
             "NystromformerForQuestionAnswering",
             "NystromformerForSequenceClassification",
             "NystromformerForTokenClassification",
-            "NystromformerLayer",
             "NystromformerModel",
             "NystromformerPreTrainedModel",
         ]
@@ -2941,7 +2930,6 @@ else:
             "PerceiverForMultimodalAutoencoding",
             "PerceiverForOpticalFlow",
             "PerceiverForSequenceClassification",
-            "PerceiverLayer",
             "PerceiverModel",
             "PerceiverPreTrainedModel",
         ]
@@ -3077,11 +3065,9 @@ else:
     )
     _import_structure["models.reformer"].extend(
         [
-            "ReformerAttention",
             "ReformerForMaskedLM",
             "ReformerForQuestionAnswering",
             "ReformerForSequenceClassification",
-            "ReformerLayer",
             "ReformerModel",
             "ReformerModelWithLMHead",
             "ReformerPreTrainedModel",
@@ -3102,7 +3088,6 @@ else:
             "RemBertForQuestionAnswering",
             "RemBertForSequenceClassification",
             "RemBertForTokenClassification",
-            "RemBertLayer",
             "RemBertModel",
             "RemBertPreTrainedModel",
             "load_tf_weights_in_rembert",
@@ -3149,7 +3134,6 @@ else:
             "RoCBertForQuestionAnswering",
             "RoCBertForSequenceClassification",
             "RoCBertForTokenClassification",
-            "RoCBertLayer",
             "RoCBertModel",
             "RoCBertPreTrainedModel",
             "load_tf_weights_in_roc_bert",
@@ -3163,7 +3147,6 @@ else:
             "RoFormerForQuestionAnswering",
             "RoFormerForSequenceClassification",
             "RoFormerForTokenClassification",
-            "RoFormerLayer",
             "RoFormerModel",
             "RoFormerPreTrainedModel",
             "load_tf_weights_in_roformer",
@@ -3220,7 +3203,6 @@ else:
             "SegformerDecodeHead",
             "SegformerForImageClassification",
             "SegformerForSemanticSegmentation",
-            "SegformerLayer",
             "SegformerModel",
             "SegformerPreTrainedModel",
         ]
@@ -3279,7 +3261,6 @@ else:
         [
             "SplinterForPreTraining",
             "SplinterForQuestionAnswering",
-            "SplinterLayer",
             "SplinterModel",
             "SplinterPreTrainedModel",
         ]
@@ -3292,7 +3273,6 @@ else:
             "SqueezeBertForSequenceClassification",
             "SqueezeBertForTokenClassification",
             "SqueezeBertModel",
-            "SqueezeBertModule",
             "SqueezeBertPreTrainedModel",
         ]
     )
@@ -3491,7 +3471,6 @@ else:
             "ViltForMaskedLM",
             "ViltForQuestionAnswering",
             "ViltForTokenClassification",
-            "ViltLayer",
             "ViltModel",
             "ViltPreTrainedModel",
         ]
@@ -3511,7 +3490,6 @@ else:
             "VisualBertForQuestionAnswering",
             "VisualBertForRegionToPhraseAlignment",
             "VisualBertForVisualReasoning",
-            "VisualBertLayer",
             "VisualBertModel",
             "VisualBertPreTrainedModel",
         ]
@@ -3527,7 +3505,6 @@ else:
     _import_structure["models.vit_mae"].extend(
         [
             "ViTMAEForPreTraining",
-            "ViTMAELayer",
             "ViTMAEModel",
             "ViTMAEPreTrainedModel",
         ]
@@ -3707,7 +3684,6 @@ else:
             "YosoForQuestionAnswering",
             "YosoForSequenceClassification",
             "YosoForTokenClassification",
-            "YosoLayer",
             "YosoModel",
             "YosoPreTrainedModel",
         ]
@@ -3854,7 +3830,6 @@ else:
     )
     _import_structure["models.bert"].extend(
         [
-            "TFBertEmbeddings",
             "TFBertForMaskedLM",
             "TFBertForMultipleChoice",
             "TFBertForNextSentencePrediction",
@@ -3920,7 +3895,6 @@ else:
             "TFConvBertForQuestionAnswering",
             "TFConvBertForSequenceClassification",
             "TFConvBertForTokenClassification",
-            "TFConvBertLayer",
             "TFConvBertModel",
             "TFConvBertPreTrainedModel",
         ]
@@ -4151,7 +4125,6 @@ else:
             "TFLongformerForTokenClassification",
             "TFLongformerModel",
             "TFLongformerPreTrainedModel",
-            "TFLongformerSelfAttention",
         ]
     )
     _import_structure["models.lxmert"].extend(
@@ -4252,7 +4225,6 @@ else:
             "TFRemBertForQuestionAnswering",
             "TFRemBertForSequenceClassification",
             "TFRemBertForTokenClassification",
-            "TFRemBertLayer",
             "TFRemBertModel",
             "TFRemBertPreTrainedModel",
         ]
@@ -4298,7 +4270,6 @@ else:
             "TFRoFormerForQuestionAnswering",
             "TFRoFormerForSequenceClassification",
             "TFRoFormerForTokenClassification",
-            "TFRoFormerLayer",
             "TFRoFormerModel",
             "TFRoFormerPreTrainedModel",
         ]
@@ -5827,7 +5798,8 @@ if TYPE_CHECKING:
         from .models.llama import LlamaTokenizer
         from .models.m2m_100 import M2M100Tokenizer
         from .models.marian import MarianTokenizer
-        from .models.mbart import MBart50Tokenizer, MBartTokenizer
+        from .models.mbart import MBartTokenizer
+        from .models.mbart50 import MBart50Tokenizer
         from .models.mluke import MLukeTokenizer
         from .models.mt5 import MT5Tokenizer
         from .models.nllb import NllbTokenizer
@@ -6298,7 +6270,6 @@ if TYPE_CHECKING:
             BertForQuestionAnswering,
             BertForSequenceClassification,
             BertForTokenClassification,
-            BertLayer,
             BertLMHeadModel,
             BertModel,
             BertPreTrainedModel,
@@ -6318,7 +6289,6 @@ if TYPE_CHECKING:
             BigBirdForQuestionAnswering,
             BigBirdForSequenceClassification,
             BigBirdForTokenClassification,
-            BigBirdLayer,
             BigBirdModel,
             BigBirdPreTrainedModel,
             load_tf_weights_in_big_bird,
@@ -6413,7 +6383,6 @@ if TYPE_CHECKING:
             CanineForQuestionAnswering,
             CanineForSequenceClassification,
             CanineForTokenClassification,
-            CanineLayer,
             CanineModel,
             CaninePreTrainedModel,
             load_tf_weights_in_canine,
@@ -6486,7 +6455,6 @@ if TYPE_CHECKING:
             ConvBertForQuestionAnswering,
             ConvBertForSequenceClassification,
             ConvBertForTokenClassification,
-            ConvBertLayer,
             ConvBertModel,
             ConvBertPreTrainedModel,
             load_tf_weights_in_convbert,
@@ -6671,7 +6639,6 @@ if TYPE_CHECKING:
             QDQBertForQuestionAnswering,
             QDQBertForSequenceClassification,
             QDQBertForTokenClassification,
-            QDQBertLayer,
             QDQBertLMHeadModel,
             QDQBertModel,
             QDQBertPreTrainedModel,
@@ -6870,7 +6837,6 @@ if TYPE_CHECKING:
             FNetForQuestionAnswering,
             FNetForSequenceClassification,
             FNetForTokenClassification,
-            FNetLayer,
             FNetModel,
             FNetPreTrainedModel,
         )
@@ -6958,13 +6924,11 @@ if TYPE_CHECKING:
             GPTNeoXForQuestionAnswering,
             GPTNeoXForSequenceClassification,
             GPTNeoXForTokenClassification,
-            GPTNeoXLayer,
             GPTNeoXModel,
             GPTNeoXPreTrainedModel,
         )
         from .models.gpt_neox_japanese import (
             GPTNeoXJapaneseForCausalLM,
-            GPTNeoXJapaneseLayer,
             GPTNeoXJapaneseModel,
             GPTNeoXJapanesePreTrainedModel,
         )
@@ -7140,7 +7104,6 @@ if TYPE_CHECKING:
             LongformerForTokenClassification,
             LongformerModel,
             LongformerPreTrainedModel,
-            LongformerSelfAttention,
         )
         from .models.longt5 import (
             LongT5EncoderModel,
@@ -7167,7 +7130,6 @@ if TYPE_CHECKING:
             LxmertModel,
             LxmertPreTrainedModel,
             LxmertVisualFeatureEncoder,
-            LxmertXLayer,
         )
         from .models.m2m_100 import (
             M2M100ForConditionalGeneration,
@@ -7184,7 +7146,7 @@ if TYPE_CHECKING:
             Mamba2Model,
             Mamba2PreTrainedModel,
         )
-        from .models.marian import MarianForCausalLM, MarianModel, MarianMTModel
+        from .models.marian import MarianForCausalLM, MarianModel, MarianMTModel, MarianPreTrainedModel
         from .models.markuplm import (
             MarkupLMForQuestionAnswering,
             MarkupLMForSequenceClassification,
@@ -7250,7 +7212,6 @@ if TYPE_CHECKING:
             MobileBertForQuestionAnswering,
             MobileBertForSequenceClassification,
             MobileBertForTokenClassification,
-            MobileBertLayer,
             MobileBertModel,
             MobileBertPreTrainedModel,
             load_tf_weights_in_mobilebert,
@@ -7286,7 +7247,6 @@ if TYPE_CHECKING:
             MPNetForQuestionAnswering,
             MPNetForSequenceClassification,
             MPNetForTokenClassification,
-            MPNetLayer,
             MPNetModel,
             MPNetPreTrainedModel,
         )
@@ -7358,7 +7318,6 @@ if TYPE_CHECKING:
             NystromformerForQuestionAnswering,
             NystromformerForSequenceClassification,
             NystromformerForTokenClassification,
-            NystromformerLayer,
             NystromformerModel,
             NystromformerPreTrainedModel,
         )
@@ -7446,7 +7405,6 @@ if TYPE_CHECKING:
             PerceiverForMultimodalAutoencoding,
             PerceiverForOpticalFlow,
             PerceiverForSequenceClassification,
-            PerceiverLayer,
             PerceiverModel,
             PerceiverPreTrainedModel,
         )
@@ -7548,11 +7506,9 @@ if TYPE_CHECKING:
             RecurrentGemmaPreTrainedModel,
         )
         from .models.reformer import (
-            ReformerAttention,
             ReformerForMaskedLM,
             ReformerForQuestionAnswering,
             ReformerForSequenceClassification,
-            ReformerLayer,
             ReformerModel,
             ReformerModelWithLMHead,
             ReformerPreTrainedModel,
@@ -7569,7 +7525,6 @@ if TYPE_CHECKING:
             RemBertForQuestionAnswering,
             RemBertForSequenceClassification,
             RemBertForTokenClassification,
-            RemBertLayer,
             RemBertModel,
             RemBertPreTrainedModel,
             load_tf_weights_in_rembert,
@@ -7608,7 +7563,6 @@ if TYPE_CHECKING:
             RoCBertForQuestionAnswering,
             RoCBertForSequenceClassification,
             RoCBertForTokenClassification,
-            RoCBertLayer,
             RoCBertModel,
             RoCBertPreTrainedModel,
             load_tf_weights_in_roc_bert,
@@ -7620,7 +7574,6 @@ if TYPE_CHECKING:
             RoFormerForQuestionAnswering,
             RoFormerForSequenceClassification,
             RoFormerForTokenClassification,
-            RoFormerLayer,
             RoFormerModel,
             RoFormerPreTrainedModel,
             load_tf_weights_in_roformer,
@@ -7665,7 +7618,6 @@ if TYPE_CHECKING:
             SegformerDecodeHead,
             SegformerForImageClassification,
             SegformerForSemanticSegmentation,
-            SegformerLayer,
             SegformerModel,
             SegformerPreTrainedModel,
         )
@@ -7710,7 +7662,6 @@ if TYPE_CHECKING:
         from .models.splinter import (
             SplinterForPreTraining,
             SplinterForQuestionAnswering,
-            SplinterLayer,
             SplinterModel,
             SplinterPreTrainedModel,
         )
@@ -7721,7 +7672,6 @@ if TYPE_CHECKING:
             SqueezeBertForSequenceClassification,
             SqueezeBertForTokenClassification,
             SqueezeBertModel,
-            SqueezeBertModule,
             SqueezeBertPreTrainedModel,
         )
         from .models.stablelm import (
@@ -7870,7 +7820,6 @@ if TYPE_CHECKING:
             ViltForMaskedLM,
             ViltForQuestionAnswering,
             ViltForTokenClassification,
-            ViltLayer,
             ViltModel,
             ViltPreTrainedModel,
         )
@@ -7886,7 +7835,6 @@ if TYPE_CHECKING:
             VisualBertForQuestionAnswering,
             VisualBertForRegionToPhraseAlignment,
             VisualBertForVisualReasoning,
-            VisualBertLayer,
             VisualBertModel,
             VisualBertPreTrainedModel,
         )
@@ -7898,7 +7846,6 @@ if TYPE_CHECKING:
         )
         from .models.vit_mae import (
             ViTMAEForPreTraining,
-            ViTMAELayer,
             ViTMAEModel,
             ViTMAEPreTrainedModel,
         )
@@ -8040,7 +7987,6 @@ if TYPE_CHECKING:
             YosoForQuestionAnswering,
             YosoForSequenceClassification,
             YosoForTokenClassification,
-            YosoLayer,
             YosoModel,
             YosoPreTrainedModel,
         )
@@ -8174,7 +8120,6 @@ if TYPE_CHECKING:
             TFBartPretrainedModel,
         )
         from .models.bert import (
-            TFBertEmbeddings,
             TFBertForMaskedLM,
             TFBertForMultipleChoice,
             TFBertForNextSentencePrediction,
@@ -8228,7 +8173,6 @@ if TYPE_CHECKING:
             TFConvBertForQuestionAnswering,
             TFConvBertForSequenceClassification,
             TFConvBertForTokenClassification,
-            TFConvBertLayer,
             TFConvBertModel,
             TFConvBertPreTrainedModel,
         )
@@ -8413,7 +8357,6 @@ if TYPE_CHECKING:
             TFLongformerForTokenClassification,
             TFLongformerModel,
             TFLongformerPreTrainedModel,
-            TFLongformerSelfAttention,
         )
         from .models.lxmert import (
             TFLxmertForPreTraining,
@@ -8503,7 +8446,6 @@ if TYPE_CHECKING:
             TFRemBertForQuestionAnswering,
             TFRemBertForSequenceClassification,
             TFRemBertForTokenClassification,
-            TFRemBertLayer,
             TFRemBertModel,
             TFRemBertPreTrainedModel,
         )
@@ -8541,7 +8483,6 @@ if TYPE_CHECKING:
             TFRoFormerForQuestionAnswering,
             TFRoFormerForSequenceClassification,
             TFRoFormerForTokenClassification,
-            TFRoFormerLayer,
             TFRoFormerModel,
             TFRoFormerPreTrainedModel,
         )

src/transformers/models/albert/__init__.py

@@ -11,165 +11,21 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 from typing import TYPE_CHECKING
 
-from ...utils import (
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    is_flax_available,
-    is_sentencepiece_available,
-    is_tf_available,
-    is_tokenizers_available,
-    is_torch_available,
-)
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
-_import_structure = {
-    "configuration_albert": ["AlbertConfig", "AlbertOnnxConfig"],
-}
-
-try:
-    if not is_sentencepiece_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_albert"] = ["AlbertTokenizer"]
-
-try:
-    if not is_tokenizers_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_albert_fast"] = ["AlbertTokenizerFast"]
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_albert"] = [
-        "AlbertForMaskedLM",
-        "AlbertForMultipleChoice",
-        "AlbertForPreTraining",
-        "AlbertForQuestionAnswering",
-        "AlbertForSequenceClassification",
-        "AlbertForTokenClassification",
-        "AlbertModel",
-        "AlbertPreTrainedModel",
-        "load_tf_weights_in_albert",
-    ]
-
-try:
-    if not is_tf_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_tf_albert"] = [
-        "TFAlbertForMaskedLM",
-        "TFAlbertForMultipleChoice",
-        "TFAlbertForPreTraining",
-        "TFAlbertForQuestionAnswering",
-        "TFAlbertForSequenceClassification",
-        "TFAlbertForTokenClassification",
-        "TFAlbertMainLayer",
-        "TFAlbertModel",
-        "TFAlbertPreTrainedModel",
-    ]
-
-try:
-    if not is_flax_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_flax_albert"] = [
-        "FlaxAlbertForMaskedLM",
-        "FlaxAlbertForMultipleChoice",
-        "FlaxAlbertForPreTraining",
-        "FlaxAlbertForQuestionAnswering",
-        "FlaxAlbertForSequenceClassification",
-        "FlaxAlbertForTokenClassification",
-        "FlaxAlbertModel",
-        "FlaxAlbertPreTrainedModel",
-    ]
-
 if TYPE_CHECKING:
-    from .configuration_albert import AlbertConfig, AlbertOnnxConfig
-
-    try:
-        if not is_sentencepiece_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_albert import AlbertTokenizer
-
-    try:
-        if not is_tokenizers_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_albert_fast import AlbertTokenizerFast
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_albert import (
-            AlbertForMaskedLM,
-            AlbertForMultipleChoice,
-            AlbertForPreTraining,
-            AlbertForQuestionAnswering,
-            AlbertForSequenceClassification,
-            AlbertForTokenClassification,
-            AlbertModel,
-            AlbertPreTrainedModel,
-            load_tf_weights_in_albert,
-        )
-
-    try:
-        if not is_tf_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_tf_albert import (
-            TFAlbertForMaskedLM,
-            TFAlbertForMultipleChoice,
-            TFAlbertForPreTraining,
-            TFAlbertForQuestionAnswering,
-            TFAlbertForSequenceClassification,
-            TFAlbertForTokenClassification,
-            TFAlbertMainLayer,
-            TFAlbertModel,
-            TFAlbertPreTrainedModel,
-        )
-
-    try:
-        if not is_flax_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_flax_albert import (
-            FlaxAlbertForMaskedLM,
-            FlaxAlbertForMultipleChoice,
-            FlaxAlbertForPreTraining,
-            FlaxAlbertForQuestionAnswering,
-            FlaxAlbertForSequenceClassification,
-            FlaxAlbertForTokenClassification,
-            FlaxAlbertModel,
-            FlaxAlbertPreTrainedModel,
-        )
+    from .configuration_albert import *
+    from .modeling_albert import *
+    from .modeling_flax_albert import *
+    from .modeling_tf_albert import *
+    from .tokenization_albert import *
+    from .tokenization_albert_fast import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/albert/configuration_albert.py

@@ -165,3 +165,6 @@ class AlbertOnnxConfig(OnnxConfig):
                 ("token_type_ids", dynamic_axis),
             ]
         )
+
+
+__all__ = ["AlbertConfig", "AlbertOnnxConfig"]

src/transformers/models/albert/modeling_albert.py

@@ -1466,3 +1466,16 @@ class AlbertForMultipleChoice(AlbertPreTrainedModel):
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
         )
+
+
+__all__ = [
+    "load_tf_weights_in_albert",
+    "AlbertPreTrainedModel",
+    "AlbertModel",
+    "AlbertForPreTraining",
+    "AlbertForMaskedLM",
+    "AlbertForSequenceClassification",
+    "AlbertForTokenClassification",
+    "AlbertForQuestionAnswering",
+    "AlbertForMultipleChoice",
+]

src/transformers/models/albert/modeling_flax_albert.py

@@ -1119,3 +1119,14 @@ append_call_sample_docstring(
     FlaxQuestionAnsweringModelOutput,
     _CONFIG_FOR_DOC,
 )
+
+__all__ = [
+    "FlaxAlbertPreTrainedModel",
+    "FlaxAlbertModel",
+    "FlaxAlbertForPreTraining",
+    "FlaxAlbertForMaskedLM",
+    "FlaxAlbertForSequenceClassification",
+    "FlaxAlbertForMultipleChoice",
+    "FlaxAlbertForTokenClassification",
+    "FlaxAlbertForQuestionAnswering",
+]

src/transformers/models/albert/modeling_tf_albert.py

@@ -1558,3 +1558,16 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
         if getattr(self, "classifier", None) is not None:
             with tf.name_scope(self.classifier.name):
                 self.classifier.build([None, None, self.config.hidden_size])
+
+
+__all__ = [
+    "TFAlbertPreTrainedModel",
+    "TFAlbertModel",
+    "TFAlbertForPreTraining",
+    "TFAlbertForMaskedLM",
+    "TFAlbertForSequenceClassification",
+    "TFAlbertForTokenClassification",
+    "TFAlbertForQuestionAnswering",
+    "TFAlbertForMultipleChoice",
+    "TFAlbertMainLayer",
+]

src/transformers/models/albert/tokenization_albert.py

@@ -23,6 +23,7 @@ import sentencepiece as spm
 
 from ...tokenization_utils import AddedToken, PreTrainedTokenizer
 from ...utils import logging
+from ...utils.import_utils import export
 
 
 logger = logging.get_logger(__name__)
@@ -32,6 +33,7 @@ VOCAB_FILES_NAMES = {"vocab_file": "spiece.model"}
 SPIECE_UNDERLINE = "▁"
 
 
+@export(backends=("sentencepiece",))
 class AlbertTokenizer(PreTrainedTokenizer):
     """
     Construct an ALBERT tokenizer. Based on [SentencePiece](https://github.com/google/sentencepiece).
@@ -343,3 +345,6 @@ class AlbertTokenizer(PreTrainedTokenizer):
                 fi.write(content_spiece_model)
 
         return (out_vocab_file,)
+
+
+__all__ = ["AlbertTokenizer"]

src/transformers/models/albert/tokenization_albert_fast.py

@@ -207,3 +207,6 @@ class AlbertTokenizerFast(PreTrainedTokenizerFast):
             copyfile(self.vocab_file, out_vocab_file)
 
         return (out_vocab_file,)
+
+
+__all__ = ["AlbertTokenizerFast"]

src/transformers/models/align/__init__.py

@@ -13,57 +13,16 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import (
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    is_torch_available,
-)
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
-_import_structure = {
-    "configuration_align": [
-        "AlignConfig",
-        "AlignTextConfig",
-        "AlignVisionConfig",
-    ],
-    "processing_align": ["AlignProcessor"],
-}
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_align"] = [
-        "AlignModel",
-        "AlignPreTrainedModel",
-        "AlignTextModel",
-        "AlignVisionModel",
-    ]
-
 if TYPE_CHECKING:
-    from .configuration_align import (
-        AlignConfig,
-        AlignTextConfig,
-        AlignVisionConfig,
-    )
-    from .processing_align import AlignProcessor
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_align import (
-            AlignModel,
-            AlignPreTrainedModel,
-            AlignTextModel,
-            AlignVisionModel,
-        )
-
+    from .configuration_align import *
+    from .modeling_align import *
+    from .processing_align import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/align/configuration_align.py

@@ -378,3 +378,6 @@ class AlignConfig(PretrainedConfig):
         """
 
         return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
+
+
+__all__ = ["AlignTextConfig", "AlignVisionConfig", "AlignConfig"]

src/transformers/models/align/modeling_align.py

@@ -1636,3 +1636,6 @@ class AlignModel(AlignPreTrainedModel):
             text_model_output=text_outputs,
             vision_model_output=vision_outputs,
         )
+
+
+__all__ = ["AlignPreTrainedModel", "AlignTextModel", "AlignVisionModel", "AlignModel"]

src/transformers/models/align/processing_align.py

@@ -162,3 +162,6 @@ class AlignProcessor(ProcessorMixin):
         tokenizer_input_names = self.tokenizer.model_input_names
         image_processor_input_names = self.image_processor.model_input_names
         return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+
+
+__all__ = ["AlignProcessor"]

src/transformers/models/altclip/__init__.py

@@ -13,55 +13,16 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
-
-
-_import_structure = {
-    "configuration_altclip": [
-        "AltCLIPConfig",
-        "AltCLIPTextConfig",
-        "AltCLIPVisionConfig",
-    ],
-    "processing_altclip": ["AltCLIPProcessor"],
-}
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_altclip"] = [
-        "AltCLIPPreTrainedModel",
-        "AltCLIPModel",
-        "AltCLIPTextModel",
-        "AltCLIPVisionModel",
-    ]
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
 if TYPE_CHECKING:
-    from .configuration_altclip import (
-        AltCLIPConfig,
-        AltCLIPTextConfig,
-        AltCLIPVisionConfig,
-    )
-    from .processing_altclip import AltCLIPProcessor
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_altclip import (
-            AltCLIPModel,
-            AltCLIPPreTrainedModel,
-            AltCLIPTextModel,
-            AltCLIPVisionModel,
-        )
-
-
+    from .configuration_altclip import *
+    from .modeling_altclip import *
+    from .processing_altclip import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/altclip/configuration_altclip.py

@@ -398,3 +398,6 @@ class AltCLIPConfig(PretrainedConfig):
         """
 
         return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
+
+
+__all__ = ["AltCLIPTextConfig", "AltCLIPVisionConfig", "AltCLIPConfig"]

src/transformers/models/altclip/modeling_altclip.py

@@ -1694,3 +1694,6 @@ def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_l
     mask = input_ids.ne(padding_idx).int()
     incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask
     return incremental_indices.long() + padding_idx
+
+
+__all__ = ["AltCLIPPreTrainedModel", "AltCLIPVisionModel", "AltCLIPTextModel", "AltCLIPModel"]

src/transformers/models/altclip/processing_altclip.py

@@ -130,3 +130,6 @@ class AltCLIPProcessor(ProcessorMixin):
         tokenizer_input_names = self.tokenizer.model_input_names
         image_processor_input_names = self.image_processor.model_input_names
         return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+
+
+__all__ = ["AltCLIPProcessor"]

src/transformers/models/audio_spectrogram_transformer/__init__.py

@@ -13,47 +13,16 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
-
-
-_import_structure = {
-    "configuration_audio_spectrogram_transformer": ["ASTConfig"],
-    "feature_extraction_audio_spectrogram_transformer": ["ASTFeatureExtractor"],
-}
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_audio_spectrogram_transformer"] = [
-        "ASTForAudioClassification",
-        "ASTModel",
-        "ASTPreTrainedModel",
-    ]
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
 if TYPE_CHECKING:
-    from .configuration_audio_spectrogram_transformer import (
-        ASTConfig,
-    )
-    from .feature_extraction_audio_spectrogram_transformer import ASTFeatureExtractor
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_audio_spectrogram_transformer import (
-            ASTForAudioClassification,
-            ASTModel,
-            ASTPreTrainedModel,
-        )
-
-
+    from .configuration_audio_spectrogram_transformer import *
+    from .feature_extraction_audio_spectrogram_transformer import *
+    from .modeling_audio_spectrogram_transformer import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/audio_spectrogram_transformer/configuration_audio_spectrogram_transformer.py

@@ -126,3 +126,6 @@ class ASTConfig(PretrainedConfig):
     # generative parameters deprecation cycle, overwriting this function prevents this from happening.
     def _get_non_default_generation_parameters(self) -> Dict[str, Any]:
         return {}
+
+
+__all__ = ["ASTConfig"]

src/transformers/models/audio_spectrogram_transformer/feature_extraction_audio_spectrogram_transformer.py

@@ -234,3 +234,6 @@ class ASTFeatureExtractor(SequenceFeatureExtractor):
             padded_inputs = padded_inputs.convert_to_tensors(return_tensors)
 
         return padded_inputs
+
+
+__all__ = ["ASTFeatureExtractor"]

src/transformers/models/audio_spectrogram_transformer/modeling_audio_spectrogram_transformer.py

@@ -654,3 +654,6 @@ class ASTForAudioClassification(ASTPreTrainedModel):
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
         )
+
+
+__all__ = ["ASTPreTrainedModel", "ASTModel", "ASTForAudioClassification"]

src/transformers/models/autoformer/__init__.py

@@ -13,45 +13,15 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-# rely on isort to merge the imports
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
-
-
-_import_structure = {
-    "configuration_autoformer": ["AutoformerConfig"],
-}
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_autoformer"] = [
-        "AutoformerForPrediction",
-        "AutoformerModel",
-        "AutoformerPreTrainedModel",
-    ]
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
 if TYPE_CHECKING:
-    from .configuration_autoformer import (
-        AutoformerConfig,
-    )
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_autoformer import (
-            AutoformerForPrediction,
-            AutoformerModel,
-            AutoformerPreTrainedModel,
-        )
-
+    from .configuration_autoformer import *
+    from .modeling_autoformer import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/autoformer/configuration_autoformer.py

@@ -240,3 +240,6 @@ class AutoformerConfig(PretrainedConfig):
             + self.num_static_real_features
             + self.input_size * 2  # the log1p(abs(loc)) and log(scale) features
         )
+
+
+__all__ = ["AutoformerConfig"]

src/transformers/models/autoformer/modeling_autoformer.py

@@ -2150,3 +2150,6 @@ class AutoformerForPrediction(AutoformerPreTrainedModel):
                 (-1, num_parallel_samples, self.config.prediction_length) + self.target_shape,
             )
         )
+
+
+__all__ = ["AutoformerPreTrainedModel", "AutoformerModel", "AutoformerForPrediction"]

src/transformers/models/bark/__init__.py

@@ -13,63 +13,16 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import (
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    is_torch_available,
-)
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
-_import_structure = {
-    "configuration_bark": [
-        "BarkCoarseConfig",
-        "BarkConfig",
-        "BarkFineConfig",
-        "BarkSemanticConfig",
-    ],
-    "processing_bark": ["BarkProcessor"],
-}
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_bark"] = [
-        "BarkFineModel",
-        "BarkSemanticModel",
-        "BarkCoarseModel",
-        "BarkModel",
-        "BarkPreTrainedModel",
-        "BarkCausalModel",
-    ]
-
 if TYPE_CHECKING:
-    from .configuration_bark import (
-        BarkCoarseConfig,
-        BarkConfig,
-        BarkFineConfig,
-        BarkSemanticConfig,
-    )
-    from .processing_bark import BarkProcessor
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_bark import (
-            BarkCausalModel,
-            BarkCoarseModel,
-            BarkFineModel,
-            BarkModel,
-            BarkPreTrainedModel,
-            BarkSemanticModel,
-        )
-
+    from .configuration_bark import *
+    from .modeling_bark import *
+    from .processing_bark import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/bark/configuration_bark.py

@@ -323,3 +323,6 @@ class BarkConfig(PretrainedConfig):
             codec_config=codec_config.to_dict(),
             **kwargs,
         )
+
+
+__all__ = ["BarkSemanticConfig", "BarkCoarseConfig", "BarkFineConfig", "BarkConfig"]

src/transformers/models/bark/modeling_bark.py

@@ -546,6 +546,8 @@ BARK_CAUSAL_MODEL_INPUTS_DOCSTRING = r"""
 
 
 # GPT2-like autoregressive model
+
+
 class BarkCausalModel(BarkPreTrainedModel):
     config_class = BarkSubModelConfig
 
@@ -1811,3 +1813,13 @@ class BarkModel(BarkPreTrainedModel):
         config.coarse_acoustics_config._attn_implementation = config._attn_implementation
         config.fine_acoustics_config._attn_implementation = config._attn_implementation
         return config
+
+
+__all__ = [
+    "BarkPreTrainedModel",
+    "BarkCausalModel",
+    "BarkFineModel",
+    "BarkCoarseModel",
+    "BarkSemanticModel",
+    "BarkModel",
+]

src/transformers/models/bark/processing_bark.py

@@ -285,3 +285,6 @@ class BarkProcessor(ProcessorMixin):
             encoded_text["history_prompt"] = voice_preset
 
         return encoded_text
+
+
+__all__ = ["BarkProcessor"]

src/transformers/models/bart/__init__.py

@@ -13,134 +13,19 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import (
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    is_flax_available,
-    is_tf_available,
-    is_tokenizers_available,
-    is_torch_available,
-)
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
-_import_structure = {
-    "configuration_bart": ["BartConfig", "BartOnnxConfig"],
-    "tokenization_bart": ["BartTokenizer"],
-}
-
-try:
-    if not is_tokenizers_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_bart_fast"] = ["BartTokenizerFast"]
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_bart"] = [
-        "BartForCausalLM",
-        "BartForConditionalGeneration",
-        "BartForQuestionAnswering",
-        "BartForSequenceClassification",
-        "BartModel",
-        "BartPreTrainedModel",
-        "BartPretrainedModel",
-        "PretrainedBartModel",
-    ]
-
-try:
-    if not is_tf_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_tf_bart"] = [
-        "TFBartForConditionalGeneration",
-        "TFBartForSequenceClassification",
-        "TFBartModel",
-        "TFBartPretrainedModel",
-    ]
-
-try:
-    if not is_flax_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_flax_bart"] = [
-        "FlaxBartDecoderPreTrainedModel",
-        "FlaxBartForCausalLM",
-        "FlaxBartForConditionalGeneration",
-        "FlaxBartForQuestionAnswering",
-        "FlaxBartForSequenceClassification",
-        "FlaxBartModel",
-        "FlaxBartPreTrainedModel",
-    ]
-
 if TYPE_CHECKING:
-    from .configuration_bart import BartConfig, BartOnnxConfig
-    from .tokenization_bart import BartTokenizer
-
-    try:
-        if not is_tokenizers_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_bart_fast import BartTokenizerFast
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_bart import (
-            BartForCausalLM,
-            BartForConditionalGeneration,
-            BartForQuestionAnswering,
-            BartForSequenceClassification,
-            BartModel,
-            BartPreTrainedModel,
-            BartPretrainedModel,
-            PretrainedBartModel,
-        )
-
-    try:
-        if not is_tf_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_tf_bart import (
-            TFBartForConditionalGeneration,
-            TFBartForSequenceClassification,
-            TFBartModel,
-            TFBartPretrainedModel,
-        )
-
-    try:
-        if not is_flax_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_flax_bart import (
-            FlaxBartDecoderPreTrainedModel,
-            FlaxBartForCausalLM,
-            FlaxBartForConditionalGeneration,
-            FlaxBartForQuestionAnswering,
-            FlaxBartForSequenceClassification,
-            FlaxBartModel,
-            FlaxBartPreTrainedModel,
-        )
-
+    from .configuration_bart import *
+    from .modeling_bart import *
+    from .modeling_flax_bart import *
+    from .modeling_tf_bart import *
+    from .tokenization_bart import *
+    from .tokenization_bart_fast import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/bart/configuration_bart.py

@@ -18,10 +18,10 @@ import warnings
 from collections import OrderedDict
 from typing import Any, Mapping, Optional
 
-from ... import PreTrainedTokenizer
 from ...configuration_utils import PretrainedConfig
 from ...onnx import OnnxConfig, OnnxConfigWithPast, OnnxSeq2SeqConfigWithPast
 from ...onnx.utils import compute_effective_axis_dimension
+from ...tokenization_utils import PreTrainedTokenizer
 from ...utils import TensorType, is_torch_available, logging
 
 
@@ -400,3 +400,6 @@ class BartOnnxConfig(OnnxSeq2SeqConfigWithPast):
             flattened_output = super(OnnxSeq2SeqConfigWithPast, self)._flatten_past_key_values_(
                 flattened_output, name, idx, t
             )
+
+
+__all__ = ["BartConfig", "BartOnnxConfig"]

src/transformers/models/bart/modeling_bart.py

@@ -2222,3 +2222,16 @@ class BartForCausalLM(BartPreTrainedModel):
                 tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
             )
         return reordered_past
+
+
+__all__ = [
+    "BartPreTrainedModel",
+    "PretrainedBartModel",
+    "BartPretrainedModel",
+    "BartDecoder",
+    "BartModel",
+    "BartForConditionalGeneration",
+    "BartForSequenceClassification",
+    "BartForQuestionAnswering",
+    "BartForCausalLM",
+]

src/transformers/models/bart/modeling_flax_bart.py

@@ -1993,3 +1993,13 @@ append_call_sample_docstring(
     FlaxCausalLMOutputWithCrossAttentions,
     _CONFIG_FOR_DOC,
 )
+
+__all__ = [
+    "FlaxBartPreTrainedModel",
+    "FlaxBartModel",
+    "FlaxBartForConditionalGeneration",
+    "FlaxBartForSequenceClassification",
+    "FlaxBartForQuestionAnswering",
+    "FlaxBartDecoderPreTrainedModel",
+    "FlaxBartForCausalLM",
+]

src/transformers/models/bart/modeling_tf_bart.py

@@ -1709,3 +1709,12 @@ class TFBartForSequenceClassification(TFBartPretrainedModel, TFSequenceClassific
         if getattr(self, "classification_head", None) is not None:
             with tf.name_scope(self.classification_head.name):
                 self.classification_head.build(None)
+
+
+__all__ = [
+    "TFBartPretrainedModel",
+    "TFBartModel",
+    "TFBartForConditionalGeneration",
+    "TFBartForSequenceClassification",
+    "TFBartMainLayer",
+]

src/transformers/models/bart/tokenization_bart.py

@@ -388,3 +388,6 @@ class BartTokenizer(PreTrainedTokenizer):
         if (is_split_into_words or add_prefix_space) and (len(text) > 0 and not text[0].isspace()):
             text = " " + text
         return (text, kwargs)
+
+
+__all__ = ["BartTokenizer"]

src/transformers/models/bart/tokenization_bart_fast.py

@@ -274,3 +274,6 @@ class BartTokenizerFast(PreTrainedTokenizerFast):
         if token_ids_1 is None:
             return len(cls + token_ids_0 + sep) * [0]
         return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]
+
+
+__all__ = ["BartTokenizerFast"]

src/transformers/models/barthez/__init__.py

@@ -11,49 +11,17 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 from typing import TYPE_CHECKING
 
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_sentencepiece_available, is_tokenizers_available
-
-
-_import_structure = {}
-
-try:
-    if not is_sentencepiece_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_barthez"] = ["BarthezTokenizer"]
-
-try:
-    if not is_tokenizers_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_barthez_fast"] = ["BarthezTokenizerFast"]
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
 if TYPE_CHECKING:
-    try:
-        if not is_sentencepiece_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_barthez import BarthezTokenizer
-
-    try:
-        if not is_tokenizers_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_barthez_fast import BarthezTokenizerFast
-
+    from .tokenization_barthez import *
+    from .tokenization_barthez_fast import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/barthez/tokenization_barthez.py

@@ -22,6 +22,7 @@ import sentencepiece as spm
 
 from ...tokenization_utils import AddedToken, PreTrainedTokenizer
 from ...utils import logging
+from ...utils.import_utils import export
 
 
 logger = logging.get_logger(__name__)
@@ -34,6 +35,7 @@ SPIECE_UNDERLINE = "▁"
 # TODO this class is useless. This is the most standard sentencpiece model. Let's find which one is closest and nuke this.
 
 
+@export(backends=("sentencepiece",))
 class BarthezTokenizer(PreTrainedTokenizer):
     """
     Adapted from [`CamembertTokenizer`] and [`BartTokenizer`]. Construct a BARThez tokenizer. Based on
@@ -284,3 +286,6 @@ class BarthezTokenizer(PreTrainedTokenizer):
                 fi.write(content_spiece_model)
 
         return (out_vocab_file,)
+
+
+__all__ = ["BarthezTokenizer"]

src/transformers/models/barthez/tokenization_barthez_fast.py

@@ -192,3 +192,6 @@ class BarthezTokenizerFast(PreTrainedTokenizerFast):
             copyfile(self.vocab_file, out_vocab_file)
 
         return (out_vocab_file,)
+
+
+__all__ = ["BarthezTokenizerFast"]

src/transformers/models/bartpho/__init__.py

@@ -11,32 +11,16 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 from typing import TYPE_CHECKING
 
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_sentencepiece_available
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
-_import_structure = {}
-
-try:
-    if not is_sentencepiece_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_bartpho"] = ["BartphoTokenizer"]
-
 if TYPE_CHECKING:
-    try:
-        if not is_sentencepiece_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_bartpho import BartphoTokenizer
-
+    from .tokenization_bartpho import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/bartpho/tokenization_bartpho.py

@@ -22,6 +22,7 @@ import sentencepiece as spm
 
 from ...tokenization_utils import AddedToken, PreTrainedTokenizer
 from ...utils import logging
+from ...utils.import_utils import export
 
 
 logger = logging.get_logger(__name__)
@@ -31,6 +32,7 @@ SPIECE_UNDERLINE = "▁"
 VOCAB_FILES_NAMES = {"vocab_file": "sentencepiece.bpe.model", "monolingual_vocab_file": "dict.txt"}
 
 
+@export(backends=("sentencepiece",))
 class BartphoTokenizer(PreTrainedTokenizer):
     """
     Adapted from [`XLMRobertaTokenizer`]. Based on [SentencePiece](https://github.com/google/sentencepiece).
@@ -311,3 +313,6 @@ class BartphoTokenizer(PreTrainedTokenizer):
                         fp.write(f"{str(token)} \n")
 
         return out_vocab_file, out_monolingual_vocab_file
+
+
+__all__ = ["BartphoTokenizer"]

src/transformers/models/beit/__init__.py

@@ -11,100 +11,20 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 from typing import TYPE_CHECKING
 
-from ...utils import (
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    is_flax_available,
-    is_torch_available,
-    is_vision_available,
-)
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
-_import_structure = {"configuration_beit": ["BeitConfig", "BeitOnnxConfig"]}
-
-try:
-    if not is_vision_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["feature_extraction_beit"] = ["BeitFeatureExtractor"]
-    _import_structure["image_processing_beit"] = ["BeitImageProcessor"]
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_beit"] = [
-        "BeitForImageClassification",
-        "BeitForMaskedImageModeling",
-        "BeitForSemanticSegmentation",
-        "BeitModel",
-        "BeitPreTrainedModel",
-        "BeitBackbone",
-    ]
-
-
-try:
-    if not is_flax_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_flax_beit"] = [
-        "FlaxBeitForImageClassification",
-        "FlaxBeitForMaskedImageModeling",
-        "FlaxBeitModel",
-        "FlaxBeitPreTrainedModel",
-    ]
-
 if TYPE_CHECKING:
-    from .configuration_beit import BeitConfig, BeitOnnxConfig
-
-    try:
-        if not is_vision_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .feature_extraction_beit import BeitFeatureExtractor
-        from .image_processing_beit import BeitImageProcessor
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_beit import (
-            BeitBackbone,
-            BeitForImageClassification,
-            BeitForMaskedImageModeling,
-            BeitForSemanticSegmentation,
-            BeitModel,
-            BeitPreTrainedModel,
-        )
-
-    try:
-        if not is_flax_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_flax_beit import (
-            FlaxBeitForImageClassification,
-            FlaxBeitForMaskedImageModeling,
-            FlaxBeitModel,
-            FlaxBeitPreTrainedModel,
-        )
-
-
+    from .configuration_beit import *
+    from .feature_extraction_beit import *
+    from .image_processing_beit import *
+    from .modeling_beit import *
+    from .modeling_flax_beit import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/beit/configuration_beit.py

@@ -224,3 +224,6 @@ class BeitOnnxConfig(OnnxConfig):
     @property
     def atol_for_validation(self) -> float:
         return 1e-4
+
+
+__all__ = ["BeitConfig", "BeitOnnxConfig"]

src/transformers/models/beit/feature_extraction_beit.py

@@ -17,12 +17,14 @@
 import warnings
 
 from ...utils import logging
+from ...utils.import_utils import export
 from .image_processing_beit import BeitImageProcessor
 
 
 logger = logging.get_logger(__name__)
 
 
+@export(backends=("vision",))
 class BeitFeatureExtractor(BeitImageProcessor):
     def __init__(self, *args, **kwargs) -> None:
         warnings.warn(
@@ -31,3 +33,6 @@ class BeitFeatureExtractor(BeitImageProcessor):
             FutureWarning,
         )
         super().__init__(*args, **kwargs)
+
+
+__all__ = ["BeitFeatureExtractor"]

src/transformers/models/beit/image_processing_beit.py

@@ -42,6 +42,7 @@ from ...utils import (
     logging,
 )
 from ...utils.deprecation import deprecate_kwarg
+from ...utils.import_utils import export
 
 
 if is_vision_available():
@@ -54,6 +55,7 @@ if is_torch_available():
 logger = logging.get_logger(__name__)
 
 
+@export(backends=("vision",))
 class BeitImageProcessor(BaseImageProcessor):
     r"""
     Constructs a BEiT image processor.
@@ -510,3 +512,6 @@ class BeitImageProcessor(BaseImageProcessor):
             semantic_segmentation = [semantic_segmentation[i] for i in range(semantic_segmentation.shape[0])]
 
         return semantic_segmentation
+
+
+__all__ = ["BeitImageProcessor"]

src/transformers/models/beit/modeling_beit.py

@@ -1576,3 +1576,13 @@ class BeitBackbone(BeitPreTrainedModel, BackboneMixin):
             hidden_states=outputs.hidden_states if output_hidden_states else None,
             attentions=outputs.attentions,
         )
+
+
+__all__ = [
+    "BeitPreTrainedModel",
+    "BeitModel",
+    "BeitForMaskedImageModeling",
+    "BeitForImageClassification",
+    "BeitForSemanticSegmentation",
+    "BeitBackbone",
+]

src/transformers/models/beit/modeling_flax_beit.py

@@ -946,3 +946,10 @@ overwrite_call_docstring(FlaxBeitForImageClassification, FLAX_BEIT_CLASSIF_DOCST
 append_replace_return_docstrings(
     FlaxBeitForImageClassification, output_type=FlaxSequenceClassifierOutput, config_class=BeitConfig
 )
+
+__all__ = [
+    "FlaxBeitPreTrainedModel",
+    "FlaxBeitModel",
+    "FlaxBeitForMaskedImageModeling",
+    "FlaxBeitForImageClassification",
+]

src/transformers/models/bert/__init__.py

@@ -11,183 +11,22 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 from typing import TYPE_CHECKING
 
-from ...utils import (
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    is_flax_available,
-    is_tensorflow_text_available,
-    is_tf_available,
-    is_tokenizers_available,
-    is_torch_available,
-)
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
-_import_structure = {
-    "configuration_bert": ["BertConfig", "BertOnnxConfig"],
-    "tokenization_bert": ["BasicTokenizer", "BertTokenizer", "WordpieceTokenizer"],
-}
-
-try:
-    if not is_tokenizers_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_bert_fast"] = ["BertTokenizerFast"]
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_bert"] = [
-        "BertForMaskedLM",
-        "BertForMultipleChoice",
-        "BertForNextSentencePrediction",
-        "BertForPreTraining",
-        "BertForQuestionAnswering",
-        "BertForSequenceClassification",
-        "BertForTokenClassification",
-        "BertLayer",
-        "BertLMHeadModel",
-        "BertModel",
-        "BertPreTrainedModel",
-        "load_tf_weights_in_bert",
-    ]
-
-try:
-    if not is_tf_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_tf_bert"] = [
-        "TFBertEmbeddings",
-        "TFBertForMaskedLM",
-        "TFBertForMultipleChoice",
-        "TFBertForNextSentencePrediction",
-        "TFBertForPreTraining",
-        "TFBertForQuestionAnswering",
-        "TFBertForSequenceClassification",
-        "TFBertForTokenClassification",
-        "TFBertLMHeadModel",
-        "TFBertMainLayer",
-        "TFBertModel",
-        "TFBertPreTrainedModel",
-    ]
-try:
-    if not is_tensorflow_text_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_bert_tf"] = ["TFBertTokenizer"]
-
-try:
-    if not is_flax_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_flax_bert"] = [
-        "FlaxBertForCausalLM",
-        "FlaxBertForMaskedLM",
-        "FlaxBertForMultipleChoice",
-        "FlaxBertForNextSentencePrediction",
-        "FlaxBertForPreTraining",
-        "FlaxBertForQuestionAnswering",
-        "FlaxBertForSequenceClassification",
-        "FlaxBertForTokenClassification",
-        "FlaxBertModel",
-        "FlaxBertPreTrainedModel",
-    ]
-
 if TYPE_CHECKING:
-    from .configuration_bert import BertConfig, BertOnnxConfig
-    from .tokenization_bert import BasicTokenizer, BertTokenizer, WordpieceTokenizer
-
-    try:
-        if not is_tokenizers_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_bert_fast import BertTokenizerFast
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_bert import (
-            BertForMaskedLM,
-            BertForMultipleChoice,
-            BertForNextSentencePrediction,
-            BertForPreTraining,
-            BertForQuestionAnswering,
-            BertForSequenceClassification,
-            BertForTokenClassification,
-            BertLayer,
-            BertLMHeadModel,
-            BertModel,
-            BertPreTrainedModel,
-            load_tf_weights_in_bert,
-        )
-
-    try:
-        if not is_tf_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_tf_bert import (
-            TFBertEmbeddings,
-            TFBertForMaskedLM,
-            TFBertForMultipleChoice,
-            TFBertForNextSentencePrediction,
-            TFBertForPreTraining,
-            TFBertForQuestionAnswering,
-            TFBertForSequenceClassification,
-            TFBertForTokenClassification,
-            TFBertLMHeadModel,
-            TFBertMainLayer,
-            TFBertModel,
-            TFBertPreTrainedModel,
-        )
-
-    try:
-        if not is_tensorflow_text_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_bert_tf import TFBertTokenizer
-
-    try:
-        if not is_flax_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_flax_bert import (
-            FlaxBertForCausalLM,
-            FlaxBertForMaskedLM,
-            FlaxBertForMultipleChoice,
-            FlaxBertForNextSentencePrediction,
-            FlaxBertForPreTraining,
-            FlaxBertForQuestionAnswering,
-            FlaxBertForSequenceClassification,
-            FlaxBertForTokenClassification,
-            FlaxBertModel,
-            FlaxBertPreTrainedModel,
-        )
-
+    from .configuration_bert import *
+    from .modeling_bert import *
+    from .modeling_flax_bert import *
+    from .modeling_tf_bert import *
+    from .tokenization_bert import *
+    from .tokenization_bert_fast import *
+    from .tokenization_bert_tf import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/bert/configuration_bert.py

@@ -149,3 +149,6 @@ class BertOnnxConfig(OnnxConfig):
                 ("token_type_ids", dynamic_axis),
             ]
         )
+
+
+__all__ = ["BertConfig", "BertOnnxConfig"]

src/transformers/models/bert/modeling_bert.py

@@ -2021,3 +2021,18 @@ class BertForQuestionAnswering(BertPreTrainedModel):
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
         )
+
+
+__all__ = [
+    "load_tf_weights_in_bert",
+    "BertPreTrainedModel",
+    "BertModel",
+    "BertForPreTraining",
+    "BertLMHeadModel",
+    "BertForMaskedLM",
+    "BertForNextSentencePrediction",
+    "BertForSequenceClassification",
+    "BertForMultipleChoice",
+    "BertForTokenClassification",
+    "BertForQuestionAnswering",
+]

src/transformers/models/bert/modeling_flax_bert.py

@@ -1711,3 +1711,16 @@ append_call_sample_docstring(
     FlaxCausalLMOutputWithCrossAttentions,
     _CONFIG_FOR_DOC,
 )
+
+__all__ = [
+    "FlaxBertPreTrainedModel",
+    "FlaxBertModel",
+    "FlaxBertForPreTraining",
+    "FlaxBertForMaskedLM",
+    "FlaxBertForNextSentencePrediction",
+    "FlaxBertForSequenceClassification",
+    "FlaxBertForMultipleChoice",
+    "FlaxBertForTokenClassification",
+    "FlaxBertForQuestionAnswering",
+    "FlaxBertForCausalLM",
+]

src/transformers/models/bert/modeling_tf_bert.py

@@ -2108,3 +2108,18 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel, TFQuestionAnsweringLoss)
         if getattr(self, "qa_outputs", None) is not None:
             with tf.name_scope(self.qa_outputs.name):
                 self.qa_outputs.build([None, None, self.config.hidden_size])
+
+
+__all__ = [
+    "TFBertPreTrainedModel",
+    "TFBertModel",
+    "TFBertForPreTraining",
+    "TFBertForMaskedLM",
+    "TFBertLMHeadModel",
+    "TFBertForNextSentencePrediction",
+    "TFBertForSequenceClassification",
+    "TFBertForMultipleChoice",
+    "TFBertForTokenClassification",
+    "TFBertForQuestionAnswering",
+    "TFBertMainLayer",
+]

src/transformers/models/bert/tokenization_bert.py

@@ -497,3 +497,6 @@ class WordpieceTokenizer:
             else:
                 output_tokens.extend(sub_tokens)
         return output_tokens
+
+
+__all__ = ["BertTokenizer", "BasicTokenizer", "WordpieceTokenizer"]

src/transformers/models/bert/tokenization_bert_fast.py

@@ -170,3 +170,6 @@ class BertTokenizerFast(PreTrainedTokenizerFast):
     def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
         files = self._tokenizer.model.save(save_directory, name=filename_prefix)
         return tuple(files)
+
+
+__all__ = ["BertTokenizerFast"]

src/transformers/models/bert/tokenization_bert_tf.py

@@ -6,9 +6,11 @@ from tensorflow_text import BertTokenizer as BertTokenizerLayer
 from tensorflow_text import FastBertTokenizer, ShrinkLongestTrimmer, case_fold_utf8, combine_segments, pad_model_inputs
 
 from ...modeling_tf_utils import keras
+from ...utils.import_utils import export
 from .tokenization_bert import BertTokenizer
 
 
+@export(backends=("tf",))
 class TFBertTokenizer(keras.layers.Layer):
     """
     This is an in-graph tokenizer for BERT. It should be initialized similarly to other tokenizers, using the
@@ -252,3 +254,6 @@ class TFBertTokenizer(keras.layers.Layer):
             "sep_token_id": self.sep_token_id,
             "pad_token_id": self.pad_token_id,
         }
+
+
+__all__ = ["TFBertTokenizer"]

src/transformers/models/bert_generation/__init__.py

@@ -11,61 +11,18 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 from typing import TYPE_CHECKING
 
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_sentencepiece_available, is_torch_available
-
-
-_import_structure = {"configuration_bert_generation": ["BertGenerationConfig"]}
-
-try:
-    if not is_sentencepiece_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_bert_generation"] = ["BertGenerationTokenizer"]
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_bert_generation"] = [
-        "BertGenerationDecoder",
-        "BertGenerationEncoder",
-        "BertGenerationPreTrainedModel",
-        "load_tf_weights_in_bert_generation",
-    ]
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
 if TYPE_CHECKING:
-    from .configuration_bert_generation import BertGenerationConfig
-
-    try:
-        if not is_sentencepiece_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_bert_generation import BertGenerationTokenizer
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_bert_generation import (
-            BertGenerationDecoder,
-            BertGenerationEncoder,
-            BertGenerationPreTrainedModel,
-            load_tf_weights_in_bert_generation,
-        )
-
+    from .configuration_bert_generation import *
+    from .modeling_bert_generation import *
+    from .tokenization_bert_generation import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/bert_generation/configuration_bert_generation.py

@@ -122,3 +122,6 @@ class BertGenerationConfig(PretrainedConfig):
         self.layer_norm_eps = layer_norm_eps
         self.position_embedding_type = position_embedding_type
         self.use_cache = use_cache
+
+
+__all__ = ["BertGenerationConfig"]

src/transformers/models/bert_generation/modeling_bert_generation.py

@@ -1018,3 +1018,11 @@ class BertGenerationDecoder(BertGenerationPreTrainedModel):
                 tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
             )
         return reordered_past
+
+
+__all__ = [
+    "load_tf_weights_in_bert_generation",
+    "BertGenerationPreTrainedModel",
+    "BertGenerationEncoder",
+    "BertGenerationDecoder",
+]

src/transformers/models/bert_generation/tokenization_bert_generation.py

@@ -22,6 +22,7 @@ import sentencepiece as spm
 
 from ...tokenization_utils import PreTrainedTokenizer
 from ...utils import logging
+from ...utils.import_utils import export
 
 
 logger = logging.get_logger(__name__)
@@ -29,6 +30,7 @@ logger = logging.get_logger(__name__)
 VOCAB_FILES_NAMES = {"vocab_file": "spiece.model"}
 
 
+@export(backends=("sentencepiece",))
 class BertGenerationTokenizer(PreTrainedTokenizer):
     """
     Construct a BertGeneration tokenizer. Based on [SentencePiece](https://github.com/google/sentencepiece).
@@ -170,3 +172,6 @@ class BertGenerationTokenizer(PreTrainedTokenizer):
                 fi.write(content_spiece_model)
 
         return (out_vocab_file,)
+
+
+__all__ = ["BertGenerationTokenizer"]

src/transformers/models/bert_japanese/__init__.py

@@ -11,19 +11,16 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 from typing import TYPE_CHECKING
 
 from ...utils import _LazyModule
-
-
-_import_structure = {"tokenization_bert_japanese": ["BertJapaneseTokenizer", "CharacterTokenizer", "MecabTokenizer"]}
+from ...utils.import_utils import define_import_structure
 
 
 if TYPE_CHECKING:
-    from .tokenization_bert_japanese import BertJapaneseTokenizer, CharacterTokenizer, MecabTokenizer
-
+    from .tokenization_bert_japanese import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/bert_japanese/tokenization_bert_japanese.py

@@ -977,3 +977,6 @@ class SentencepieceTokenizer:
                 new_pieces.append(piece)
 
         return new_pieces
+
+
+__all__ = ["BertJapaneseTokenizer", "CharacterTokenizer", "MecabTokenizer"]

src/transformers/models/bertweet/__init__.py

@@ -11,19 +11,16 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 from typing import TYPE_CHECKING
 
 from ...utils import _LazyModule
-
-
-_import_structure = {"tokenization_bertweet": ["BertweetTokenizer"]}
+from ...utils.import_utils import define_import_structure
 
 
 if TYPE_CHECKING:
-    from .tokenization_bertweet import BertweetTokenizer
-
+    from .tokenization_bertweet import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/bertweet/tokenization_bertweet.py

@@ -764,3 +764,5 @@ def casual_tokenize(text, preserve_case=True, reduce_len=False, strip_handles=Fa
 
 
 ###############################################################################
+
+__all__ = ["BertweetTokenizer"]

src/transformers/models/big_bird/__init__.py

@@ -13,133 +13,18 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import (
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    is_flax_available,
-    is_sentencepiece_available,
-    is_tf_available,
-    is_tokenizers_available,
-    is_torch_available,
-)
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
-_import_structure = {
-    "configuration_big_bird": ["BigBirdConfig", "BigBirdOnnxConfig"],
-}
-
-try:
-    if not is_sentencepiece_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_big_bird"] = ["BigBirdTokenizer"]
-
-try:
-    if not is_tokenizers_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_big_bird_fast"] = ["BigBirdTokenizerFast"]
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_big_bird"] = [
-        "BigBirdForCausalLM",
-        "BigBirdForMaskedLM",
-        "BigBirdForMultipleChoice",
-        "BigBirdForPreTraining",
-        "BigBirdForQuestionAnswering",
-        "BigBirdForSequenceClassification",
-        "BigBirdForTokenClassification",
-        "BigBirdLayer",
-        "BigBirdModel",
-        "BigBirdPreTrainedModel",
-        "load_tf_weights_in_big_bird",
-    ]
-
-try:
-    if not is_flax_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_flax_big_bird"] = [
-        "FlaxBigBirdForCausalLM",
-        "FlaxBigBirdForMaskedLM",
-        "FlaxBigBirdForMultipleChoice",
-        "FlaxBigBirdForPreTraining",
-        "FlaxBigBirdForQuestionAnswering",
-        "FlaxBigBirdForSequenceClassification",
-        "FlaxBigBirdForTokenClassification",
-        "FlaxBigBirdModel",
-        "FlaxBigBirdPreTrainedModel",
-    ]
-
 if TYPE_CHECKING:
-    from .configuration_big_bird import BigBirdConfig, BigBirdOnnxConfig
-
-    try:
-        if not is_sentencepiece_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_big_bird import BigBirdTokenizer
-
-    try:
-        if not is_tokenizers_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_big_bird_fast import BigBirdTokenizerFast
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_big_bird import (
-            BigBirdForCausalLM,
-            BigBirdForMaskedLM,
-            BigBirdForMultipleChoice,
-            BigBirdForPreTraining,
-            BigBirdForQuestionAnswering,
-            BigBirdForSequenceClassification,
-            BigBirdForTokenClassification,
-            BigBirdLayer,
-            BigBirdModel,
-            BigBirdPreTrainedModel,
-            load_tf_weights_in_big_bird,
-        )
-
-    try:
-        if not is_flax_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_flax_big_bird import (
-            FlaxBigBirdForCausalLM,
-            FlaxBigBirdForMaskedLM,
-            FlaxBigBirdForMultipleChoice,
-            FlaxBigBirdForPreTraining,
-            FlaxBigBirdForQuestionAnswering,
-            FlaxBigBirdForSequenceClassification,
-            FlaxBigBirdForTokenClassification,
-            FlaxBigBirdModel,
-            FlaxBigBirdPreTrainedModel,
-        )
-
+    from .configuration_big_bird import *
+    from .modeling_big_bird import *
+    from .modeling_flax_big_bird import *
+    from .tokenization_big_bird import *
+    from .tokenization_big_bird_fast import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/big_bird/configuration_big_bird.py

@@ -171,3 +171,6 @@ class BigBirdOnnxConfig(OnnxConfig):
                 ("attention_mask", dynamic_axis),
             ]
         )
+
+
+__all__ = ["BigBirdConfig", "BigBirdOnnxConfig"]

src/transformers/models/big_bird/modeling_big_bird.py

@@ -3147,3 +3147,17 @@ class BigBirdForQuestionAnswering(BigBirdPreTrainedModel):
         mask.unsqueeze_(0)  # -> (1, maxlen)
         mask = torch.where(mask < q_lengths, 1, 0)
         return mask
+
+
+__all__ = [
+    "load_tf_weights_in_big_bird",
+    "BigBirdPreTrainedModel",
+    "BigBirdModel",
+    "BigBirdForPreTraining",
+    "BigBirdForMaskedLM",
+    "BigBirdForCausalLM",
+    "BigBirdForSequenceClassification",
+    "BigBirdForMultipleChoice",
+    "BigBirdForTokenClassification",
+    "BigBirdForQuestionAnswering",
+]

src/transformers/models/big_bird/modeling_flax_big_bird.py

@@ -2633,3 +2633,15 @@ append_call_sample_docstring(
     FlaxCausalLMOutputWithCrossAttentions,
     _CONFIG_FOR_DOC,
 )
+
+__all__ = [
+    "FlaxBigBirdPreTrainedModel",
+    "FlaxBigBirdModel",
+    "FlaxBigBirdForPreTraining",
+    "FlaxBigBirdForMaskedLM",
+    "FlaxBigBirdForSequenceClassification",
+    "FlaxBigBirdForMultipleChoice",
+    "FlaxBigBirdForTokenClassification",
+    "FlaxBigBirdForQuestionAnswering",
+    "FlaxBigBirdForCausalLM",
+]

src/transformers/models/big_bird/tokenization_big_bird.py

@@ -23,6 +23,7 @@ import sentencepiece as spm
 
 from ...tokenization_utils import AddedToken, PreTrainedTokenizer
 from ...utils import logging
+from ...utils.import_utils import export
 
 
 logger = logging.get_logger(__name__)
@@ -30,6 +31,7 @@ logger = logging.get_logger(__name__)
 VOCAB_FILES_NAMES = {"vocab_file": "spiece.model"}
 
 
+@export(backends=("sentencepiece",))
 class BigBirdTokenizer(PreTrainedTokenizer):
     """
     Construct a BigBird tokenizer. Based on [SentencePiece](https://github.com/google/sentencepiece).
@@ -319,3 +321,6 @@ class BigBirdTokenizer(PreTrainedTokenizer):
         if token_ids_1 is None:
             return len(cls + token_ids_0 + sep) * [0]
         return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+
+__all__ = ["BigBirdTokenizer"]

src/transformers/models/big_bird/tokenization_big_bird_fast.py

@@ -227,3 +227,6 @@ class BigBirdTokenizerFast(PreTrainedTokenizerFast):
             copyfile(self.vocab_file, out_vocab_file)
 
         return (out_vocab_file,)
+
+
+__all__ = ["BigBirdTokenizerFast"]

src/transformers/models/bigbird_pegasus/__init__.py

@@ -13,55 +13,15 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
-
-
-_import_structure = {
-    "configuration_bigbird_pegasus": [
-        "BigBirdPegasusConfig",
-        "BigBirdPegasusOnnxConfig",
-    ],
-}
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_bigbird_pegasus"] = [
-        "BigBirdPegasusForCausalLM",
-        "BigBirdPegasusForConditionalGeneration",
-        "BigBirdPegasusForQuestionAnswering",
-        "BigBirdPegasusForSequenceClassification",
-        "BigBirdPegasusModel",
-        "BigBirdPegasusPreTrainedModel",
-    ]
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
 if TYPE_CHECKING:
-    from .configuration_bigbird_pegasus import (
-        BigBirdPegasusConfig,
-        BigBirdPegasusOnnxConfig,
-    )
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_bigbird_pegasus import (
-            BigBirdPegasusForCausalLM,
-            BigBirdPegasusForConditionalGeneration,
-            BigBirdPegasusForQuestionAnswering,
-            BigBirdPegasusForSequenceClassification,
-            BigBirdPegasusModel,
-            BigBirdPegasusPreTrainedModel,
-        )
-
-
+    from .configuration_bigbird_pegasus import *
+    from .modeling_bigbird_pegasus import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/bigbird_pegasus/configuration_bigbird_pegasus.py

@@ -17,10 +17,10 @@
 from collections import OrderedDict
 from typing import Any, Mapping, Optional
 
-from ... import PreTrainedTokenizer
 from ...configuration_utils import PretrainedConfig
 from ...onnx import OnnxConfig, OnnxConfigWithPast, OnnxSeq2SeqConfigWithPast
 from ...onnx.utils import compute_effective_axis_dimension
+from ...tokenization_utils import PreTrainedTokenizer
 from ...utils import TensorType, is_torch_available, logging
 
 
@@ -407,3 +407,6 @@ class BigBirdPegasusOnnxConfig(OnnxSeq2SeqConfigWithPast):
             flattened_output = super(OnnxSeq2SeqConfigWithPast, self)._flatten_past_key_values_(
                 flattened_output, name, idx, t
             )
+
+
+__all__ = ["BigBirdPegasusConfig", "BigBirdPegasusOnnxConfig"]

src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py

@@ -3083,3 +3083,13 @@ class BigBirdPegasusForCausalLM(BigBirdPegasusPreTrainedModel):
                 tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
             )
         return reordered_past
+
+
+__all__ = [
+    "BigBirdPegasusPreTrainedModel",
+    "BigBirdPegasusModel",
+    "BigBirdPegasusForConditionalGeneration",
+    "BigBirdPegasusForSequenceClassification",
+    "BigBirdPegasusForQuestionAnswering",
+    "BigBirdPegasusForCausalLM",
+]

src/transformers/models/biogpt/__init__.py

@@ -13,49 +13,16 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
-
-
-_import_structure = {
-    "configuration_biogpt": ["BioGptConfig"],
-    "tokenization_biogpt": ["BioGptTokenizer"],
-}
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_biogpt"] = [
-        "BioGptForCausalLM",
-        "BioGptForTokenClassification",
-        "BioGptForSequenceClassification",
-        "BioGptModel",
-        "BioGptPreTrainedModel",
-    ]
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
 if TYPE_CHECKING:
-    from .configuration_biogpt import BioGptConfig
-    from .tokenization_biogpt import BioGptTokenizer
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_biogpt import (
-            BioGptForCausalLM,
-            BioGptForSequenceClassification,
-            BioGptForTokenClassification,
-            BioGptModel,
-            BioGptPreTrainedModel,
-        )
-
-
+    from .configuration_biogpt import *
+    from .modeling_biogpt import *
+    from .tokenization_biogpt import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/biogpt/configuration_biogpt.py

@@ -129,3 +129,6 @@ class BioGptConfig(PretrainedConfig):
         self.layerdrop = layerdrop
         self.activation_dropout = activation_dropout
         super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+
+__all__ = ["BioGptConfig"]

src/transformers/models/biogpt/modeling_biogpt.py

@@ -934,3 +934,12 @@ class BioGptForSequenceClassification(BioGptPreTrainedModel):
 
     def set_input_embeddings(self, value):
         self.biogpt.embed_tokens = value
+
+
+__all__ = [
+    "BioGptPreTrainedModel",
+    "BioGptModel",
+    "BioGptForCausalLM",
+    "BioGptForTokenClassification",
+    "BioGptForSequenceClassification",
+]

src/transformers/models/biogpt/tokenization_biogpt.py

@@ -356,3 +356,6 @@ class BioGptTokenizer(PreTrainedTokenizer):
             )
 
         self.sm = sacremoses
+
+
+__all__ = ["BioGptTokenizer"]

src/transformers/models/bit/__init__.py

@@ -13,59 +13,16 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
-
-
-_import_structure = {"configuration_bit": ["BitConfig", "BitOnnxConfig"]}
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_bit"] = [
-        "BitForImageClassification",
-        "BitModel",
-        "BitPreTrainedModel",
-        "BitBackbone",
-    ]
-
-
-try:
-    if not is_vision_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["image_processing_bit"] = ["BitImageProcessor"]
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
 if TYPE_CHECKING:
-    from .configuration_bit import BitConfig, BitOnnxConfig
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_bit import (
-            BitBackbone,
-            BitForImageClassification,
-            BitModel,
-            BitPreTrainedModel,
-        )
-
-    try:
-        if not is_vision_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .image_processing_bit import BitImageProcessor
-
+    from .configuration_bit import *
+    from .image_processing_bit import *
+    from .modeling_bit import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/bit/configuration_bit.py

@@ -131,3 +131,6 @@ class BitConfig(BackboneConfigMixin, PretrainedConfig):
         self._out_features, self._out_indices = get_aligned_output_features_output_indices(
             out_features=out_features, out_indices=out_indices, stage_names=self.stage_names
         )
+
+
+__all__ = ["BitConfig"]

src/transformers/models/bit/image_processing_bit.py

@@ -39,6 +39,7 @@ from ...image_utils import (
     validate_preprocess_arguments,
 )
 from ...utils import TensorType, filter_out_non_signature_kwargs, is_vision_available, logging
+from ...utils.import_utils import export
 
 
 logger = logging.get_logger(__name__)
@@ -48,6 +49,7 @@ if is_vision_available():
     import PIL
 
 
+@export(backends=("vision",))
 class BitImageProcessor(BaseImageProcessor):
     r"""
     Constructs a BiT image processor.
@@ -319,3 +321,6 @@ class BitImageProcessor(BaseImageProcessor):
 
         data = {"pixel_values": images}
         return BatchFeature(data=data, tensor_type=return_tensors)
+
+
+__all__ = ["BitImageProcessor"]

src/transformers/models/bit/modeling_bit.py

@@ -901,3 +901,6 @@ class BitBackbone(BitPreTrainedModel, BackboneMixin):
             hidden_states=outputs.hidden_states if output_hidden_states else None,
             attentions=None,
         )
+
+
+__all__ = ["BitPreTrainedModel", "BitModel", "BitForImageClassification", "BitBackbone"]

src/transformers/models/blenderbot/__init__.py

@@ -11,128 +11,21 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 from typing import TYPE_CHECKING
 
-from ...utils import (
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    is_flax_available,
-    is_tf_available,
-    is_tokenizers_available,
-    is_torch_available,
-)
-
-
-_import_structure = {
-    "configuration_blenderbot": [
-        "BlenderbotConfig",
-        "BlenderbotOnnxConfig",
-    ],
-    "tokenization_blenderbot": ["BlenderbotTokenizer"],
-}
-
-try:
-    if not is_tokenizers_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_blenderbot_fast"] = ["BlenderbotTokenizerFast"]
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_blenderbot"] = [
-        "BlenderbotForCausalLM",
-        "BlenderbotForConditionalGeneration",
-        "BlenderbotModel",
-        "BlenderbotPreTrainedModel",
-    ]
-
-
-try:
-    if not is_tf_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_tf_blenderbot"] = [
-        "TFBlenderbotForConditionalGeneration",
-        "TFBlenderbotModel",
-        "TFBlenderbotPreTrainedModel",
-    ]
-
-
-try:
-    if not is_flax_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_flax_blenderbot"] = [
-        "FlaxBlenderbotForConditionalGeneration",
-        "FlaxBlenderbotModel",
-        "FlaxBlenderbotPreTrainedModel",
-    ]
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
 if TYPE_CHECKING:
-    from .configuration_blenderbot import (
-        BlenderbotConfig,
-        BlenderbotOnnxConfig,
-    )
-    from .tokenization_blenderbot import BlenderbotTokenizer
-
-    try:
-        if not is_tokenizers_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_blenderbot_fast import BlenderbotTokenizerFast
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_blenderbot import (
-            BlenderbotForCausalLM,
-            BlenderbotForConditionalGeneration,
-            BlenderbotModel,
-            BlenderbotPreTrainedModel,
-        )
-
-    try:
-        if not is_tf_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_tf_blenderbot import (
-            TFBlenderbotForConditionalGeneration,
-            TFBlenderbotModel,
-            TFBlenderbotPreTrainedModel,
-        )
-
-    try:
-        if not is_flax_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_flax_blenderbot import (
-            FlaxBlenderbotForConditionalGeneration,
-            FlaxBlenderbotModel,
-            FlaxBlenderbotPreTrainedModel,
-        )
-
+    from .configuration_blenderbot import *
+    from .modeling_blenderbot import *
+    from .modeling_flax_blenderbot import *
+    from .modeling_tf_blenderbot import *
+    from .tokenization_blenderbot import *
+    from .tokenization_blenderbot_fast import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/blenderbot/configuration_blenderbot.py

@@ -17,11 +17,11 @@
 from collections import OrderedDict
 from typing import Any, Mapping, Optional
 
-from ... import PreTrainedTokenizer
 from ...configuration_utils import PretrainedConfig
 from ...file_utils import TensorType, is_torch_available
 from ...onnx import OnnxConfig, OnnxConfigWithPast, OnnxSeq2SeqConfigWithPast
 from ...onnx.utils import compute_effective_axis_dimension
+from ...tokenization_utils import PreTrainedTokenizer
 from ...utils import logging
 
 
@@ -390,3 +390,6 @@ class BlenderbotOnnxConfig(OnnxSeq2SeqConfigWithPast):
             inputs_or_outputs[f"{name}.{i}.decoder.value"] = {0: "batch", 2: decoder_sequence}
             inputs_or_outputs[f"{name}.{i}.encoder.key"] = {0: "batch", 2: encoder_sequence}
             inputs_or_outputs[f"{name}.{i}.encoder.value"] = {0: "batch", 2: encoder_sequence}
+
+
+__all__ = ["BlenderbotConfig", "BlenderbotOnnxConfig"]

src/transformers/models/blenderbot/modeling_blenderbot.py

@@ -1609,3 +1609,11 @@ class BlenderbotForCausalLM(BlenderbotPreTrainedModel):
                 tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
             )
         return reordered_past
+
+
+__all__ = [
+    "BlenderbotPreTrainedModel",
+    "BlenderbotModel",
+    "BlenderbotForConditionalGeneration",
+    "BlenderbotForCausalLM",
+]

src/transformers/models/blenderbot/modeling_flax_blenderbot.py

@@ -1503,3 +1503,5 @@ overwrite_call_docstring(
 append_replace_return_docstrings(
     FlaxBlenderbotForConditionalGeneration, output_type=FlaxSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC
 )
+
+__all__ = ["FlaxBlenderbotPreTrainedModel", "FlaxBlenderbotModel", "FlaxBlenderbotForConditionalGeneration"]

src/transformers/models/blenderbot/modeling_tf_blenderbot.py

@@ -1553,3 +1553,11 @@ class TFBlenderbotForConditionalGeneration(TFBlenderbotPreTrainedModel, TFCausal
         if getattr(self, "bias_layer", None) is not None:
             with tf.name_scope(self.bias_layer.name):
                 self.bias_layer.build(None)
+
+
+__all__ = [
+    "TFBlenderbotPreTrainedModel",
+    "TFBlenderbotModel",
+    "TFBlenderbotForConditionalGeneration",
+    "TFBlenderbotMainLayer",
+]

src/transformers/models/blenderbot/tokenization_blenderbot.py

@@ -405,3 +405,6 @@ class BlenderbotTokenizer(PreTrainedTokenizer):
             `List[int]`: list of [input IDs](../glossary#input-ids) with the appropriate special tokens.
         """
         return token_ids_0 + [self.eos_token_id]
+
+
+__all__ = ["BlenderbotTokenizer"]

src/transformers/models/blenderbot/tokenization_blenderbot_fast.py

@@ -287,3 +287,6 @@ class BlenderbotTokenizerFast(PreTrainedTokenizerFast):
             `List[int]`: list of [input IDs](../glossary#input-ids) with the appropriate special tokens.
         """
         return token_ids_0 + [self.eos_token_id]
+
+
+__all__ = ["BlenderbotTokenizerFast"]

src/transformers/models/blenderbot_small/__init__.py

@@ -13,122 +13,19 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import (
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    is_flax_available,
-    is_tf_available,
-    is_tokenizers_available,
-    is_torch_available,
-)
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
-_import_structure = {
-    "configuration_blenderbot_small": [
-        "BlenderbotSmallConfig",
-        "BlenderbotSmallOnnxConfig",
-    ],
-    "tokenization_blenderbot_small": ["BlenderbotSmallTokenizer"],
-}
-
-try:
-    if not is_tokenizers_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_blenderbot_small_fast"] = ["BlenderbotSmallTokenizerFast"]
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_blenderbot_small"] = [
-        "BlenderbotSmallForCausalLM",
-        "BlenderbotSmallForConditionalGeneration",
-        "BlenderbotSmallModel",
-        "BlenderbotSmallPreTrainedModel",
-    ]
-
-try:
-    if not is_tf_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_tf_blenderbot_small"] = [
-        "TFBlenderbotSmallForConditionalGeneration",
-        "TFBlenderbotSmallModel",
-        "TFBlenderbotSmallPreTrainedModel",
-    ]
-
-try:
-    if not is_flax_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_flax_blenderbot_small"] = [
-        "FlaxBlenderbotSmallForConditionalGeneration",
-        "FlaxBlenderbotSmallModel",
-        "FlaxBlenderbotSmallPreTrainedModel",
-    ]
-
 if TYPE_CHECKING:
-    from .configuration_blenderbot_small import (
-        BlenderbotSmallConfig,
-        BlenderbotSmallOnnxConfig,
-    )
-    from .tokenization_blenderbot_small import BlenderbotSmallTokenizer
-
-    try:
-        if not is_tokenizers_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_blenderbot_small_fast import BlenderbotSmallTokenizerFast
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_blenderbot_small import (
-            BlenderbotSmallForCausalLM,
-            BlenderbotSmallForConditionalGeneration,
-            BlenderbotSmallModel,
-            BlenderbotSmallPreTrainedModel,
-        )
-
-    try:
-        if not is_tf_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_tf_blenderbot_small import (
-            TFBlenderbotSmallForConditionalGeneration,
-            TFBlenderbotSmallModel,
-            TFBlenderbotSmallPreTrainedModel,
-        )
-
-    try:
-        if not is_flax_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_flax_blenderbot_small import (
-            FlaxBlenderbotSmallForConditionalGeneration,
-            FlaxBlenderbotSmallModel,
-            FlaxBlenderbotSmallPreTrainedModel,
-        )
-
+    from .configuration_blenderbot_small import *
+    from .modeling_blenderbot_small import *
+    from .modeling_flax_blenderbot_small import *
+    from .modeling_tf_blenderbot_small import *
+    from .tokenization_blenderbot_small import *
+    from .tokenization_blenderbot_small_fast import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/blenderbot_small/configuration_blenderbot_small.py

@@ -17,11 +17,11 @@
 from collections import OrderedDict
 from typing import Any, Mapping, Optional
 
-from ... import PreTrainedTokenizer
 from ...configuration_utils import PretrainedConfig
 from ...file_utils import TensorType, is_torch_available
 from ...onnx import OnnxConfig, OnnxConfigWithPast, OnnxSeq2SeqConfigWithPast
 from ...onnx.utils import compute_effective_axis_dimension
+from ...tokenization_utils import PreTrainedTokenizer
 from ...utils import logging
 
 
@@ -385,3 +385,6 @@ class BlenderbotSmallOnnxConfig(OnnxSeq2SeqConfigWithPast):
             flattened_output = super(OnnxSeq2SeqConfigWithPast, self)._flatten_past_key_values_(
                 flattened_output, name, idx, t
             )
+
+
+__all__ = ["BlenderbotSmallConfig", "BlenderbotSmallOnnxConfig"]

src/transformers/models/blenderbot_small/modeling_blenderbot_small.py

@@ -1561,3 +1561,11 @@ class BlenderbotSmallForCausalLM(BlenderbotSmallPreTrainedModel):
                 tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
             )
         return reordered_past
+
+
+__all__ = [
+    "BlenderbotSmallPreTrainedModel",
+    "BlenderbotSmallModel",
+    "BlenderbotSmallForConditionalGeneration",
+    "BlenderbotSmallForCausalLM",
+]

src/transformers/models/blenderbot_small/modeling_flax_blenderbot_small.py

@@ -1519,3 +1519,9 @@ overwrite_call_docstring(
 append_replace_return_docstrings(
     FlaxBlenderbotSmallForConditionalGeneration, output_type=FlaxSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC
 )
+
+__all__ = [
+    "FlaxBlenderbotSmallPreTrainedModel",
+    "FlaxBlenderbotSmallModel",
+    "FlaxBlenderbotSmallForConditionalGeneration",
+]

src/transformers/models/blenderbot_small/modeling_tf_blenderbot_small.py

@@ -1523,3 +1523,11 @@ class TFBlenderbotSmallForConditionalGeneration(TFBlenderbotSmallPreTrainedModel
         if getattr(self, "bias_layer", None) is not None:
             with tf.name_scope(self.bias_layer.name):
                 self.bias_layer.build(None)
+
+
+__all__ = [
+    "TFBlenderbotSmallPreTrainedModel",
+    "TFBlenderbotSmallModel",
+    "TFBlenderbotSmallForConditionalGeneration",
+    "TFBlenderbotSmallMainLayer",
+]

src/transformers/models/blenderbot_small/tokenization_blenderbot_small.py

@@ -217,3 +217,6 @@ class BlenderbotSmallTokenizer(PreTrainedTokenizer):
                 index += 1
 
         return vocab_file, merge_file
+
+
+__all__ = ["BlenderbotSmallTokenizer"]

src/transformers/models/blenderbot_small/tokenization_blenderbot_small_fast.py

@@ -98,3 +98,6 @@ class BlenderbotSmallTokenizerFast(PreTrainedTokenizerFast):
         if token_ids_1 is None:
             return len(cls + token_ids_0 + sep) * [0]
         return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]
+
+
+__all__ = ["BlenderbotSmallTokenizerFast"]

src/transformers/models/blip/__init__.py

@@ -13,110 +13,20 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import (
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    is_tf_available,
-    is_torch_available,
-    is_vision_available,
-)
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
 
 
-_import_structure = {
-    "configuration_blip": [
-        "BlipConfig",
-        "BlipTextConfig",
-        "BlipVisionConfig",
-    ],
-    "processing_blip": ["BlipProcessor"],
-}
-
-try:
-    if not is_vision_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["image_processing_blip"] = ["BlipImageProcessor"]
-
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_blip"] = [
-        "BlipModel",
-        "BlipPreTrainedModel",
-        "BlipForConditionalGeneration",
-        "BlipForQuestionAnswering",
-        "BlipVisionModel",
-        "BlipTextModel",
-        "BlipForImageTextRetrieval",
-    ]
-
-try:
-    if not is_tf_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_tf_blip"] = [
-        "TFBlipModel",
-        "TFBlipPreTrainedModel",
-        "TFBlipForConditionalGeneration",
-        "TFBlipForQuestionAnswering",
-        "TFBlipVisionModel",
-        "TFBlipTextModel",
-        "TFBlipForImageTextRetrieval",
-    ]
-
 if TYPE_CHECKING:
-    from .configuration_blip import BlipConfig, BlipTextConfig, BlipVisionConfig
-    from .processing_blip import BlipProcessor
-
-    try:
-        if not is_vision_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .image_processing_blip import BlipImageProcessor
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_blip import (
-            BlipForConditionalGeneration,
-            BlipForImageTextRetrieval,
-            BlipForQuestionAnswering,
-            BlipModel,
-            BlipPreTrainedModel,
-            BlipTextModel,
-            BlipVisionModel,
-        )
-
-    try:
-        if not is_tf_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_tf_blip import (
-            TFBlipForConditionalGeneration,
-            TFBlipForImageTextRetrieval,
-            TFBlipForQuestionAnswering,
-            TFBlipModel,
-            TFBlipPreTrainedModel,
-            TFBlipTextModel,
-            TFBlipVisionModel,
-        )
-
+    from .configuration_blip import *
+    from .image_processing_blip import *
+    from .modeling_blip import *
+    from .modeling_blip_text import *
+    from .modeling_tf_blip import *
+    from .modeling_tf_blip_text import *
+    from .processing_blip import *
 else:
     import sys
 
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/blip/configuration_blip.py

@@ -360,3 +360,6 @@ class BlipConfig(PretrainedConfig):
         """
 
         return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
+
+
+__all__ = ["BlipTextConfig", "BlipVisionConfig", "BlipConfig"]

src/transformers/models/blip/image_processing_blip.py

@@ -34,6 +34,7 @@ from ...image_utils import (
     validate_preprocess_arguments,
 )
 from ...utils import TensorType, filter_out_non_signature_kwargs, is_vision_available, logging
+from ...utils.import_utils import export
 
 
 if is_vision_available():
@@ -43,6 +44,7 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
+@export(backends=("vision",))
 class BlipImageProcessor(BaseImageProcessor):
     r"""
     Constructs a BLIP image processor.
@@ -292,3 +294,6 @@ class BlipImageProcessor(BaseImageProcessor):
         encoded_outputs = BatchFeature(data={"pixel_values": images}, tensor_type=return_tensors)
 
         return encoded_outputs
+
+
+__all__ = ["BlipImageProcessor"]

src/transformers/models/blip/modeling_blip.py

@@ -1563,3 +1563,13 @@ class BlipForImageTextRetrieval(BlipPreTrainedModel):
             attentions=vision_outputs.attentions,
             question_embeds=question_embeds,
         )
+
+
+__all__ = [
+    "BlipPreTrainedModel",
+    "BlipVisionModel",
+    "BlipModel",
+    "BlipForConditionalGeneration",
+    "BlipForQuestionAnswering",
+    "BlipForImageTextRetrieval",
+]

src/transformers/models/blip/modeling_blip_text.py

@@ -568,6 +568,8 @@ class BlipTextPreTrainedModel(PreTrainedModel):
 
 
 # Adapted from https://github.com/salesforce/BLIP/blob/3a29b7410476bf5f2ba0955827390eb6ea1f4f9d/models/med.py#L571
+
+
 class BlipTextModel(BlipTextPreTrainedModel):
     """
     The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
@@ -948,3 +950,6 @@ class BlipTextLMHeadModel(BlipTextPreTrainedModel):
                 tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
             )
         return reordered_past
+
+
+__all__ = ["BlipTextModel"]

src/transformers/models/blip/modeling_tf_blip.py

@@ -1696,3 +1696,14 @@ class TFBlipForImageTextRetrieval(TFBlipPreTrainedModel):
         if getattr(self, "itm_head", None) is not None:
             with tf.name_scope(self.itm_head.name):
                 self.itm_head.build([None, None, self.config.text_config.hidden_size])
+
+
+__all__ = [
+    "TFBlipPreTrainedModel",
+    "TFBlipVisionModel",
+    "TFBlipModel",
+    "TFBlipForConditionalGeneration",
+    "TFBlipForQuestionAnswering",
+    "TFBlipForImageTextRetrieval",
+    "TFBlipMainLayer",
+]