Fast imports part 3 (#9474)

* New intermediate inits

* Update template

* Avoid importing torch/tf/flax in tokenization unless necessary

* Styling

* Shutup flake8

* Better python version check

src/transformers/file_utils.py

@@ -51,7 +51,7 @@ from .utils import logging
 
 
 # The package importlib_metadata is in a different place, depending on the python version.
-if version.parse(sys.version) < version.parse("3.8"):
+if sys.version_info < (3, 8):
     import importlib_metadata
 else:
     import importlib.metadata as importlib_metadata

src/transformers/models/albert/__init__.py

@@ -16,9 +16,58 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_sentencepiece_available, is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_albert import ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, AlbertConfig
 
+from ...file_utils import (
+    _BaseLazyModule,
+    is_sentencepiece_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_albert": ["ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "AlbertConfig"],
+}
+
+if is_sentencepiece_available():
+    _import_structure["tokenization_albert"] = ["AlbertTokenizer"]
+
+if is_tokenizers_available():
+    _import_structure["tokenization_albert_fast"] = ["AlbertTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_albert"] = [
+        "ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "AlbertForMaskedLM",
+        "AlbertForMultipleChoice",
+        "AlbertForPreTraining",
+        "AlbertForQuestionAnswering",
+        "AlbertForSequenceClassification",
+        "AlbertForTokenClassification",
+        "AlbertModel",
+        "AlbertPreTrainedModel",
+        "load_tf_weights_in_albert",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_albert"] = [
+        "TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFAlbertForMaskedLM",
+        "TFAlbertForMultipleChoice",
+        "TFAlbertForPreTraining",
+        "TFAlbertForQuestionAnswering",
+        "TFAlbertForSequenceClassification",
+        "TFAlbertForTokenClassification",
+        "TFAlbertMainLayer",
+        "TFAlbertModel",
+        "TFAlbertPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_albert import ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, AlbertConfig
 
     if is_sentencepiece_available():
         from .tokenization_albert import AlbertTokenizer
@@ -53,3 +102,21 @@ if is_tf_available():
             TFAlbertModel,
             TFAlbertPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/auto/__init__.py

@@ -16,11 +16,77 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_flax_available, is_tf_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_flax_available, is_tf_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_auto": ["ALL_PRETRAINED_CONFIG_ARCHIVE_MAP", "CONFIG_MAPPING", "MODEL_NAMES_MAPPING", "AutoConfig"],
+    "tokenization_auto": ["TOKENIZER_MAPPING", "AutoTokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_auto"] = [
+        "MODEL_FOR_CAUSAL_LM_MAPPING",
+        "MODEL_FOR_MASKED_LM_MAPPING",
+        "MODEL_FOR_MULTIPLE_CHOICE_MAPPING",
+        "MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING",
+        "MODEL_FOR_PRETRAINING_MAPPING",
+        "MODEL_FOR_QUESTION_ANSWERING_MAPPING",
+        "MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING",
+        "MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING",
+        "MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING",
+        "MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING",
+        "MODEL_MAPPING",
+        "MODEL_WITH_LM_HEAD_MAPPING",
+        "AutoModel",
+        "AutoModelForCausalLM",
+        "AutoModelForMaskedLM",
+        "AutoModelForMultipleChoice",
+        "AutoModelForNextSentencePrediction",
+        "AutoModelForPreTraining",
+        "AutoModelForQuestionAnswering",
+        "AutoModelForSeq2SeqLM",
+        "AutoModelForSequenceClassification",
+        "AutoModelForTableQuestionAnswering",
+        "AutoModelForTokenClassification",
+        "AutoModelWithLMHead",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_auto"] = [
+        "TF_MODEL_FOR_CAUSAL_LM_MAPPING",
+        "TF_MODEL_FOR_MASKED_LM_MAPPING",
+        "TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING",
+        "TF_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING",
+        "TF_MODEL_FOR_PRETRAINING_MAPPING",
+        "TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING",
+        "TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING",
+        "TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING",
+        "TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING",
+        "TF_MODEL_MAPPING",
+        "TF_MODEL_WITH_LM_HEAD_MAPPING",
+        "TFAutoModel",
+        "TFAutoModelForCausalLM",
+        "TFAutoModelForMaskedLM",
+        "TFAutoModelForMultipleChoice",
+        "TFAutoModelForPreTraining",
+        "TFAutoModelForQuestionAnswering",
+        "TFAutoModelForSeq2SeqLM",
+        "TFAutoModelForSequenceClassification",
+        "TFAutoModelForTokenClassification",
+        "TFAutoModelWithLMHead",
+    ]
+
+if is_flax_available():
+    _import_structure["modeling_flax_auto"] = ["FLAX_MODEL_MAPPING", "FlaxAutoModel"]
+
+
+if TYPE_CHECKING:
     from .configuration_auto import ALL_PRETRAINED_CONFIG_ARCHIVE_MAP, CONFIG_MAPPING, MODEL_NAMES_MAPPING, AutoConfig
     from .tokenization_auto import TOKENIZER_MAPPING, AutoTokenizer
 
-
     if is_torch_available():
         from .modeling_auto import (
             MODEL_FOR_CAUSAL_LM_MAPPING,
@@ -76,3 +142,21 @@ if is_tf_available():
 
     if is_flax_available():
         from .modeling_flax_auto import FLAX_MODEL_MAPPING, FlaxAutoModel
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/bart/__init__.py

@@ -15,11 +15,38 @@
 # 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 ...file_utils import is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_bart": ["BART_PRETRAINED_CONFIG_ARCHIVE_MAP", "BartConfig"],
+    "tokenization_bart": ["BartTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_bart_fast"] = ["BartTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_bart"] = [
+        "BART_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "BartForConditionalGeneration",
+        "BartForQuestionAnswering",
+        "BartForSequenceClassification",
+        "BartModel",
+        "BartPretrainedModel",
+        "PretrainedBartModel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_bart"] = ["TFBartForConditionalGeneration", "TFBartModel", "TFBartPretrainedModel"]
+
+
+if TYPE_CHECKING:
     from .configuration_bart import BART_PRETRAINED_CONFIG_ARCHIVE_MAP, BartConfig
     from .tokenization_bart import BartTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_bart_fast import BartTokenizerFast
 
@@ -36,3 +63,21 @@ if is_torch_available():
 
     if is_tf_available():
         from .modeling_tf_bart import TFBartForConditionalGeneration, TFBartModel, TFBartPretrainedModel
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/barthez/__init__.py

@@ -16,11 +16,42 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_sentencepiece_available, is_tokenizers_available
+from typing import TYPE_CHECKING
 
+from ...file_utils import _BaseLazyModule, is_sentencepiece_available, is_tokenizers_available
+
+
+_import_structure = {}
+
+if is_sentencepiece_available():
+    _import_structure["tokenization_barthez"] = ["BarthezTokenizer"]
+
+if is_tokenizers_available():
+    _import_structure["tokenization_barthez_fast"] = ["BarthezTokenizerFast"]
+
+
+if TYPE_CHECKING:
 
     if is_sentencepiece_available():
         from .tokenization_barthez import BarthezTokenizer
 
     if is_tokenizers_available():
         from .tokenization_barthez_fast import BarthezTokenizerFast
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/bert/__init__.py

@@ -16,11 +16,67 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import (
+    _BaseLazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_bert": ["BERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "BertConfig"],
+    "tokenization_bert": ["BasicTokenizer", "BertTokenizer", "WordpieceTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_bert_fast"] = ["BertTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_bert"] = [
+        "BERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "BertForMaskedLM",
+        "BertForMultipleChoice",
+        "BertForNextSentencePrediction",
+        "BertForPreTraining",
+        "BertForQuestionAnswering",
+        "BertForSequenceClassification",
+        "BertForTokenClassification",
+        "BertLayer",
+        "BertLMHeadModel",
+        "BertModel",
+        "BertPreTrainedModel",
+        "load_tf_weights_in_bert",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_bert"] = [
+        "TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFBertEmbeddings",
+        "TFBertForMaskedLM",
+        "TFBertForMultipleChoice",
+        "TFBertForNextSentencePrediction",
+        "TFBertForPreTraining",
+        "TFBertForQuestionAnswering",
+        "TFBertForSequenceClassification",
+        "TFBertForTokenClassification",
+        "TFBertLMHeadModel",
+        "TFBertMainLayer",
+        "TFBertModel",
+        "TFBertPreTrainedModel",
+    ]
+
+if is_flax_available():
+    _import_structure["modeling_flax_bert"] = ["FlaxBertForMaskedLM", "FlaxBertModel"]
+
+
+if TYPE_CHECKING:
     from .configuration_bert import BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, BertConfig
     from .tokenization_bert import BasicTokenizer, BertTokenizer, WordpieceTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_bert_fast import BertTokenizerFast
 
@@ -60,3 +116,21 @@ if is_tf_available():
 
     if is_flax_available():
         from .modeling_flax_bert import FlaxBertForMaskedLM, FlaxBertModel
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/bert_generation/__init__.py

@@ -16,9 +16,28 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_sentencepiece_available, is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_bert_generation import BertGenerationConfig
 
+from ...file_utils import _BaseLazyModule, is_sentencepiece_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_bert_generation": ["BertGenerationConfig"],
+}
+
+if is_sentencepiece_available():
+    _import_structure["tokenization_bert_generation"] = ["BertGenerationTokenizer"]
+
+if is_torch_available():
+    _import_structure["modeling_bert_generation"] = [
+        "BertGenerationDecoder",
+        "BertGenerationEncoder",
+        "load_tf_weights_in_bert_generation",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_bert_generation import BertGenerationConfig
 
     if is_sentencepiece_available():
         from .tokenization_bert_generation import BertGenerationTokenizer
@@ -29,3 +48,21 @@ if is_torch_available():
             BertGenerationEncoder,
             load_tf_weights_in_bert_generation,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/bert_japanese/__init__.py

@@ -16,4 +16,33 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule
+
+
+_import_structure = {
+    "tokenization_bert_japanese": ["BertJapaneseTokenizer", "CharacterTokenizer", "MecabTokenizer"],
+}
+
+
+if TYPE_CHECKING:
     from .tokenization_bert_japanese import BertJapaneseTokenizer, CharacterTokenizer, MecabTokenizer
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/bertweet/__init__.py

@@ -16,4 +16,33 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule
+
+
+_import_structure = {
+    "tokenization_bertweet": ["BertweetTokenizer"],
+}
+
+
+if TYPE_CHECKING:
     from .tokenization_bertweet import BertweetTokenizer
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/blenderbot/__init__.py

@@ -16,11 +16,33 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_blenderbot": ["BLENDERBOT_PRETRAINED_CONFIG_ARCHIVE_MAP", "BlenderbotConfig"],
+    "tokenization_blenderbot": ["BlenderbotTokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_blenderbot"] = [
+        "BLENDERBOT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "BlenderbotForConditionalGeneration",
+        "BlenderbotModel",
+        "BlenderbotPreTrainedModel",
+    ]
+
+
+if is_tf_available():
+    _import_structure["modeling_tf_blenderbot"] = ["TFBlenderbotForConditionalGeneration"]
+
+
+if TYPE_CHECKING:
     from .configuration_blenderbot import BLENDERBOT_PRETRAINED_CONFIG_ARCHIVE_MAP, BlenderbotConfig
     from .tokenization_blenderbot import BlenderbotTokenizer
 
-
     if is_torch_available():
         from .modeling_blenderbot import (
             BLENDERBOT_PRETRAINED_MODEL_ARCHIVE_LIST,
@@ -29,6 +51,23 @@ if is_torch_available():
             BlenderbotPreTrainedModel,
         )
 
-
     if is_tf_available():
         from .modeling_tf_blenderbot import TFBlenderbotForConditionalGeneration
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/blenderbot_small/__init__.py

@@ -15,11 +15,29 @@
 # 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 ...file_utils import is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_blenderbot_small": ["BLENDERBOT_SMALL_PRETRAINED_CONFIG_ARCHIVE_MAP", "BlenderbotSmallConfig"],
+    "tokenization_blenderbot_small": ["BlenderbotSmallTokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_blenderbot_small"] = [
+        "BLENDERBOT_SMALL_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "BlenderbotSmallForConditionalGeneration",
+        "BlenderbotSmallModel",
+        "BlenderbotSmallPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_blenderbot_small import BLENDERBOT_SMALL_PRETRAINED_CONFIG_ARCHIVE_MAP, BlenderbotSmallConfig
     from .tokenization_blenderbot_small import BlenderbotSmallTokenizer
 
-
     if is_torch_available():
         from .modeling_blenderbot_small import (
             BLENDERBOT_SMALL_PRETRAINED_MODEL_ARCHIVE_LIST,
@@ -27,3 +45,21 @@ if is_torch_available():
             BlenderbotSmallModel,
             BlenderbotSmallPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/camembert/__init__.py

@@ -16,9 +16,53 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_sentencepiece_available, is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_camembert import CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, CamembertConfig
 
+from ...file_utils import (
+    _BaseLazyModule,
+    is_sentencepiece_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_camembert": ["CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "CamembertConfig"],
+}
+
+if is_sentencepiece_available():
+    _import_structure["tokenization_camembert"] = ["CamembertTokenizer"]
+
+if is_tokenizers_available():
+    _import_structure["tokenization_camembert_fast"] = ["CamembertTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_camembert"] = [
+        "CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "CamembertForCausalLM",
+        "CamembertForMaskedLM",
+        "CamembertForMultipleChoice",
+        "CamembertForQuestionAnswering",
+        "CamembertForSequenceClassification",
+        "CamembertForTokenClassification",
+        "CamembertModel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_camembert"] = [
+        "TF_CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFCamembertForMaskedLM",
+        "TFCamembertForMultipleChoice",
+        "TFCamembertForQuestionAnswering",
+        "TFCamembertForSequenceClassification",
+        "TFCamembertForTokenClassification",
+        "TFCamembertModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_camembert import CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, CamembertConfig
 
     if is_sentencepiece_available():
         from .tokenization_camembert import CamembertTokenizer
@@ -48,3 +92,21 @@ if is_tf_available():
             TFCamembertForTokenClassification,
             TFCamembertModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/ctrl/__init__.py

@@ -16,11 +16,39 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_ctrl": ["CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP", "CTRLConfig"],
+    "tokenization_ctrl": ["CTRLTokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_ctrl"] = [
+        "CTRL_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "CTRLForSequenceClassification",
+        "CTRLLMHeadModel",
+        "CTRLModel",
+        "CTRLPreTrainedModel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_ctrl"] = [
+        "TF_CTRL_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFCTRLForSequenceClassification",
+        "TFCTRLLMHeadModel",
+        "TFCTRLModel",
+        "TFCTRLPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_ctrl import CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP, CTRLConfig
     from .tokenization_ctrl import CTRLTokenizer
 
-
     if is_torch_available():
         from .modeling_ctrl import (
             CTRL_PRETRAINED_MODEL_ARCHIVE_LIST,
@@ -38,3 +66,21 @@ if is_tf_available():
             TFCTRLModel,
             TFCTRLPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/deberta/__init__.py

@@ -16,11 +16,29 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_deberta": ["DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP", "DebertaConfig"],
+    "tokenization_deberta": ["DebertaTokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_deberta"] = [
+        "DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "DebertaForSequenceClassification",
+        "DebertaModel",
+        "DebertaPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_deberta import DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaConfig
     from .tokenization_deberta import DebertaTokenizer
 
-
     if is_torch_available():
         from .modeling_deberta import (
             DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
@@ -28,3 +46,21 @@ if is_torch_available():
             DebertaModel,
             DebertaPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/distilbert/__init__.py

@@ -16,11 +16,49 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_distilbert": ["DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "DistilBertConfig"],
+    "tokenization_distilbert": ["DistilBertTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_distilbert_fast"] = ["DistilBertTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_distilbert"] = [
+        "DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "DistilBertForMaskedLM",
+        "DistilBertForMultipleChoice",
+        "DistilBertForQuestionAnswering",
+        "DistilBertForSequenceClassification",
+        "DistilBertForTokenClassification",
+        "DistilBertModel",
+        "DistilBertPreTrainedModel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_distilbert"] = [
+        "TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFDistilBertForMaskedLM",
+        "TFDistilBertForMultipleChoice",
+        "TFDistilBertForQuestionAnswering",
+        "TFDistilBertForSequenceClassification",
+        "TFDistilBertForTokenClassification",
+        "TFDistilBertMainLayer",
+        "TFDistilBertModel",
+        "TFDistilBertPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_distilbert import DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DistilBertConfig
     from .tokenization_distilbert import DistilBertTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_distilbert_fast import DistilBertTokenizerFast
 
@@ -48,3 +86,21 @@ if is_tf_available():
             TFDistilBertModel,
             TFDistilBertPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/dpr/__init__.py

@@ -16,7 +16,57 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_dpr": ["DPR_PRETRAINED_CONFIG_ARCHIVE_MAP", "DPRConfig"],
+    "tokenization_dpr": [
+        "DPRContextEncoderTokenizer",
+        "DPRQuestionEncoderTokenizer",
+        "DPRReaderOutput",
+        "DPRReaderTokenizer",
+    ],
+}
+
+
+if is_tokenizers_available():
+    _import_structure["tokenization_dpr_fast"] = [
+        "DPRContextEncoderTokenizerFast",
+        "DPRQuestionEncoderTokenizerFast",
+        "DPRReaderTokenizerFast",
+    ]
+
+if is_torch_available():
+    _import_structure["modeling_dpr"] = [
+        "DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "DPRContextEncoder",
+        "DPRPretrainedContextEncoder",
+        "DPRPretrainedQuestionEncoder",
+        "DPRPretrainedReader",
+        "DPRQuestionEncoder",
+        "DPRReader",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_dpr"] = [
+        "TF_DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TF_DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TF_DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFDPRContextEncoder",
+        "TFDPRPretrainedContextEncoder",
+        "TFDPRPretrainedQuestionEncoder",
+        "TFDPRPretrainedReader",
+        "TFDPRQuestionEncoder",
+        "TFDPRReader",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_dpr import DPR_PRETRAINED_CONFIG_ARCHIVE_MAP, DPRConfig
     from .tokenization_dpr import (
         DPRContextEncoderTokenizer,
@@ -25,7 +75,6 @@ from .tokenization_dpr import (
         DPRReaderTokenizer,
     )
 
-
     if is_tokenizers_available():
         from .tokenization_dpr_fast import (
             DPRContextEncoderTokenizerFast,
@@ -58,3 +107,21 @@ if is_tf_available():
             TFDPRQuestionEncoder,
             TFDPRReader,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/electra/__init__.py

@@ -16,11 +16,51 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_electra": ["ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP", "ElectraConfig"],
+    "tokenization_electra": ["ElectraTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_electra_fast"] = ["ElectraTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_electra"] = [
+        "ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "ElectraForMaskedLM",
+        "ElectraForMultipleChoice",
+        "ElectraForPreTraining",
+        "ElectraForQuestionAnswering",
+        "ElectraForSequenceClassification",
+        "ElectraForTokenClassification",
+        "ElectraModel",
+        "ElectraPreTrainedModel",
+        "load_tf_weights_in_electra",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_electra"] = [
+        "TF_ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFElectraForMaskedLM",
+        "TFElectraForMultipleChoice",
+        "TFElectraForPreTraining",
+        "TFElectraForQuestionAnswering",
+        "TFElectraForSequenceClassification",
+        "TFElectraForTokenClassification",
+        "TFElectraModel",
+        "TFElectraPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_electra import ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP, ElectraConfig
     from .tokenization_electra import ElectraTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_electra_fast import ElectraTokenizerFast
 
@@ -50,3 +90,21 @@ if is_tf_available():
             TFElectraModel,
             TFElectraPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/encoder_decoder/__init__.py

@@ -16,9 +16,39 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_encoder_decoder import EncoderDecoderConfig
 
+from ...file_utils import _BaseLazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_encoder_decoder": ["EncoderDecoderConfig"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_encoder_decoder"] = ["EncoderDecoderModel"]
+
+
+if TYPE_CHECKING:
+    from .configuration_encoder_decoder import EncoderDecoderConfig
 
     if is_torch_available():
         from .modeling_encoder_decoder import EncoderDecoderModel
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/flaubert/__init__.py

@@ -16,11 +16,44 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_flaubert": ["FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "FlaubertConfig"],
+    "tokenization_flaubert": ["FlaubertTokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_flaubert"] = [
+        "FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "FlaubertForMultipleChoice",
+        "FlaubertForQuestionAnswering",
+        "FlaubertForQuestionAnsweringSimple",
+        "FlaubertForSequenceClassification",
+        "FlaubertForTokenClassification",
+        "FlaubertModel",
+        "FlaubertWithLMHeadModel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_flaubert"] = [
+        "TF_FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFFlaubertForMultipleChoice",
+        "TFFlaubertForQuestionAnsweringSimple",
+        "TFFlaubertForSequenceClassification",
+        "TFFlaubertForTokenClassification",
+        "TFFlaubertModel",
+        "TFFlaubertWithLMHeadModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_flaubert import FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, FlaubertConfig
     from .tokenization_flaubert import FlaubertTokenizer
 
-
     if is_torch_available():
         from .modeling_flaubert import (
             FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
@@ -43,3 +76,21 @@ if is_tf_available():
             TFFlaubertModel,
             TFFlaubertWithLMHeadModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/fsmt/__init__.py

@@ -16,10 +16,41 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_fsmt": ["FSMT_PRETRAINED_CONFIG_ARCHIVE_MAP", "FSMTConfig"],
+    "tokenization_fsmt": ["FSMTTokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_fsmt"] = ["FSMTForConditionalGeneration", "FSMTModel", "PretrainedFSMTModel"]
+
+
+if TYPE_CHECKING:
     from .configuration_fsmt import FSMT_PRETRAINED_CONFIG_ARCHIVE_MAP, FSMTConfig
     from .tokenization_fsmt import FSMTTokenizer
 
-
     if is_torch_available():
         from .modeling_fsmt import FSMTForConditionalGeneration, FSMTModel, PretrainedFSMTModel
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/funnel/__init__.py

@@ -16,11 +16,51 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_funnel": ["FUNNEL_PRETRAINED_CONFIG_ARCHIVE_MAP", "FunnelConfig"],
+    "tokenization_funnel": ["FunnelTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_funnel_fast"] = ["FunnelTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_funnel"] = [
+        "FUNNEL_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "FunnelBaseModel",
+        "FunnelForMaskedLM",
+        "FunnelForMultipleChoice",
+        "FunnelForPreTraining",
+        "FunnelForQuestionAnswering",
+        "FunnelForSequenceClassification",
+        "FunnelForTokenClassification",
+        "FunnelModel",
+        "load_tf_weights_in_funnel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_funnel"] = [
+        "TF_FUNNEL_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFFunnelBaseModel",
+        "TFFunnelForMaskedLM",
+        "TFFunnelForMultipleChoice",
+        "TFFunnelForPreTraining",
+        "TFFunnelForQuestionAnswering",
+        "TFFunnelForSequenceClassification",
+        "TFFunnelForTokenClassification",
+        "TFFunnelModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_funnel import FUNNEL_PRETRAINED_CONFIG_ARCHIVE_MAP, FunnelConfig
     from .tokenization_funnel import FunnelTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_funnel_fast import FunnelTokenizerFast
 
@@ -50,3 +90,21 @@ if is_tf_available():
             TFFunnelForTokenClassification,
             TFFunnelModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/gpt2/__init__.py

@@ -16,11 +16,46 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_gpt2": ["GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP", "GPT2Config"],
+    "tokenization_gpt2": ["GPT2Tokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_gpt2_fast"] = ["GPT2TokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_gpt2"] = [
+        "GPT2_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "GPT2DoubleHeadsModel",
+        "GPT2ForSequenceClassification",
+        "GPT2LMHeadModel",
+        "GPT2Model",
+        "GPT2PreTrainedModel",
+        "load_tf_weights_in_gpt2",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_gpt2"] = [
+        "TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFGPT2DoubleHeadsModel",
+        "TFGPT2ForSequenceClassification",
+        "TFGPT2LMHeadModel",
+        "TFGPT2MainLayer",
+        "TFGPT2Model",
+        "TFGPT2PreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config
     from .tokenization_gpt2 import GPT2Tokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_gpt2_fast import GPT2TokenizerFast
 
@@ -45,3 +80,21 @@ if is_tf_available():
             TFGPT2Model,
             TFGPT2PreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/herbert/__init__.py

@@ -16,9 +16,39 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tokenizers_available
+from typing import TYPE_CHECKING
-from .tokenization_herbert import HerbertTokenizer
 
+from ...file_utils import _BaseLazyModule, is_tokenizers_available
+
+
+_import_structure = {
+    "tokenization_herbert": ["HerbertTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_herbert_fast"] = ["HerbertTokenizerFast"]
+
+
+if TYPE_CHECKING:
+    from .tokenization_herbert import HerbertTokenizer
 
     if is_tokenizers_available():
         from .tokenization_herbert_fast import HerbertTokenizerFast
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/layoutlm/__init__.py

@@ -16,11 +16,32 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_layoutlm": ["LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP", "LayoutLMConfig"],
+    "tokenization_layoutlm": ["LayoutLMTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_layoutlm_fast"] = ["LayoutLMTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_layoutlm"] = [
+        "LAYOUTLM_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "LayoutLMForMaskedLM",
+        "LayoutLMForTokenClassification",
+        "LayoutLMModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_layoutlm import LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP, LayoutLMConfig
     from .tokenization_layoutlm import LayoutLMTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_layoutlm_fast import LayoutLMTokenizerFast
 
@@ -31,3 +52,21 @@ if is_torch_available():
             LayoutLMForTokenClassification,
             LayoutLMModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/led/__init__.py

@@ -15,11 +15,38 @@
 # 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 ...file_utils import is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_led": ["LED_PRETRAINED_CONFIG_ARCHIVE_MAP", "LEDConfig"],
+    "tokenization_led": ["LEDTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_led_fast"] = ["LEDTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_led"] = [
+        "LED_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "LEDForConditionalGeneration",
+        "LEDForQuestionAnswering",
+        "LEDForSequenceClassification",
+        "LEDModel",
+        "LEDPreTrainedModel",
+    ]
+
+
+if is_tf_available():
+    _import_structure["modeling_tf_led"] = ["TFLEDForConditionalGeneration", "TFLEDModel", "TFLEDPreTrainedModel"]
+
+
+if TYPE_CHECKING:
     from .configuration_led import LED_PRETRAINED_CONFIG_ARCHIVE_MAP, LEDConfig
     from .tokenization_led import LEDTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_led_fast import LEDTokenizerFast
 
@@ -33,6 +60,23 @@ if is_torch_available():
             LEDPreTrainedModel,
         )
 
-
     if is_tf_available():
         from .modeling_tf_led import TFLEDForConditionalGeneration, TFLEDModel, TFLEDPreTrainedModel
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/longformer/__init__.py

@@ -16,11 +16,48 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_longformer": ["LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "LongformerConfig"],
+    "tokenization_longformer": ["LongformerTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_longformer_fast"] = ["LongformerTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_longformer"] = [
+        "LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "LongformerForMaskedLM",
+        "LongformerForMultipleChoice",
+        "LongformerForQuestionAnswering",
+        "LongformerForSequenceClassification",
+        "LongformerForTokenClassification",
+        "LongformerModel",
+        "LongformerSelfAttention",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_longformer"] = [
+        "TF_LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFLongformerForMaskedLM",
+        "TFLongformerForMultipleChoice",
+        "TFLongformerForQuestionAnswering",
+        "TFLongformerForSequenceClassification",
+        "TFLongformerForTokenClassification",
+        "TFLongformerModel",
+        "TFLongformerSelfAttention",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_longformer import LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, LongformerConfig
     from .tokenization_longformer import LongformerTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_longformer_fast import LongformerTokenizerFast
 
@@ -47,3 +84,21 @@ if is_tf_available():
             TFLongformerModel,
             TFLongformerSelfAttention,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/lxmert/__init__.py

@@ -16,11 +16,45 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_lxmert": ["LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "LxmertConfig"],
+    "tokenization_lxmert": ["LxmertTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_lxmert_fast"] = ["LxmertTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_lxmert"] = [
+        "LxmertEncoder",
+        "LxmertForPreTraining",
+        "LxmertForQuestionAnswering",
+        "LxmertModel",
+        "LxmertPreTrainedModel",
+        "LxmertVisualFeatureEncoder",
+        "LxmertXLayer",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_lxmert"] = [
+        "TF_LXMERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFLxmertForPreTraining",
+        "TFLxmertMainLayer",
+        "TFLxmertModel",
+        "TFLxmertPreTrainedModel",
+        "TFLxmertVisualFeatureEncoder",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_lxmert import LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP, LxmertConfig
     from .tokenization_lxmert import LxmertTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_lxmert_fast import LxmertTokenizerFast
 
@@ -44,3 +78,21 @@ if is_tf_available():
             TFLxmertPreTrainedModel,
             TFLxmertVisualFeatureEncoder,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/marian/__init__.py

@@ -15,9 +15,38 @@
 # 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 ...file_utils import is_sentencepiece_available, is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_marian import MARIAN_PRETRAINED_CONFIG_ARCHIVE_MAP, MarianConfig
 
+from ...file_utils import (
+    _BaseLazyModule,
+    is_sentencepiece_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_marian": ["MARIAN_PRETRAINED_CONFIG_ARCHIVE_MAP", "MarianConfig"],
+}
+
+if is_sentencepiece_available():
+    _import_structure["tokenization_marian"] = ["MarianTokenizer"]
+
+if is_torch_available():
+    _import_structure["modeling_marian"] = [
+        "MARIAN_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "MarianModel",
+        "MarianMTModel",
+        "MarianPreTrainedModel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_marian"] = ["TFMarianMTModel"]
+
+
+if TYPE_CHECKING:
+    from .configuration_marian import MARIAN_PRETRAINED_CONFIG_ARCHIVE_MAP, MarianConfig
 
     if is_sentencepiece_available():
         from .tokenization_marian import MarianTokenizer
@@ -32,3 +61,21 @@ if is_torch_available():
 
     if is_tf_available():
         from .modeling_tf_marian import TFMarianMTModel
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/mbart/__init__.py

@@ -15,9 +15,43 @@
 # 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 ...file_utils import is_sentencepiece_available, is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_mbart import MBART_PRETRAINED_CONFIG_ARCHIVE_MAP, MBartConfig
 
+from ...file_utils import (
+    _BaseLazyModule,
+    is_sentencepiece_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_mbart": ["MBART_PRETRAINED_CONFIG_ARCHIVE_MAP", "MBartConfig"],
+}
+
+if is_sentencepiece_available():
+    _import_structure["tokenization_mbart"] = ["MBartTokenizer"]
+
+if is_tokenizers_available():
+    _import_structure["tokenization_mbart_fast"] = ["MBartTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_mbart"] = [
+        "MBART_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "MBartForConditionalGeneration",
+        "MBartForQuestionAnswering",
+        "MBartForSequenceClassification",
+        "MBartModel",
+        "MBartPreTrainedModel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_mbart"] = ["TFMBartForConditionalGeneration"]
+
+
+if TYPE_CHECKING:
+    from .configuration_mbart import MBART_PRETRAINED_CONFIG_ARCHIVE_MAP, MBartConfig
 
     if is_sentencepiece_available():
         from .tokenization_mbart import MBartTokenizer
@@ -37,3 +71,21 @@ if is_torch_available():
 
     if is_tf_available():
         from .modeling_tf_mbart import TFMBartForConditionalGeneration
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/mmbt/__init__.py

@@ -16,9 +16,39 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_mmbt import MMBTConfig
 
+from ...file_utils import _BaseLazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_mmbt": ["MMBTConfig"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_mmbt"] = ["MMBTForClassification", "MMBTModel", "ModalEmbeddings"]
+
+
+if TYPE_CHECKING:
+    from .configuration_mmbt import MMBTConfig
 
     if is_torch_available():
         from .modeling_mmbt import MMBTForClassification, MMBTModel, ModalEmbeddings
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/mobilebert/__init__.py

@@ -16,11 +16,55 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_mobilebert": ["MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "MobileBertConfig"],
+    "tokenization_mobilebert": ["MobileBertTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_mobilebert_fast"] = ["MobileBertTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_mobilebert"] = [
+        "MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "MobileBertForMaskedLM",
+        "MobileBertForMultipleChoice",
+        "MobileBertForNextSentencePrediction",
+        "MobileBertForPreTraining",
+        "MobileBertForQuestionAnswering",
+        "MobileBertForSequenceClassification",
+        "MobileBertForTokenClassification",
+        "MobileBertLayer",
+        "MobileBertModel",
+        "MobileBertPreTrainedModel",
+        "load_tf_weights_in_mobilebert",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_mobilebert"] = [
+        "TF_MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFMobileBertForMaskedLM",
+        "TFMobileBertForMultipleChoice",
+        "TFMobileBertForNextSentencePrediction",
+        "TFMobileBertForPreTraining",
+        "TFMobileBertForQuestionAnswering",
+        "TFMobileBertForSequenceClassification",
+        "TFMobileBertForTokenClassification",
+        "TFMobileBertMainLayer",
+        "TFMobileBertModel",
+        "TFMobileBertPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_mobilebert import MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, MobileBertConfig
     from .tokenization_mobilebert import MobileBertTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_mobilebert_fast import MobileBertTokenizerFast
 
@@ -54,3 +98,21 @@ if is_tf_available():
             TFMobileBertModel,
             TFMobileBertPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/mpnet/__init__.py

@@ -16,11 +16,57 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import (
+    _BaseLazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_mpnet": ["MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "MPNetConfig"],
+    "tokenization_mpnet": ["MPNetTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_mpnet_fast"] = ["MPNetTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_mpnet"] = [
+        "MPNET_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "MPNetForMaskedLM",
+        "MPNetForMultipleChoice",
+        "MPNetForQuestionAnswering",
+        "MPNetForSequenceClassification",
+        "MPNetForTokenClassification",
+        "MPNetLayer",
+        "MPNetModel",
+        "MPNetPreTrainedModel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_mpnet"] = [
+        "TF_MPNET_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFMPNetEmbeddings",
+        "TFMPNetForMaskedLM",
+        "TFMPNetForMultipleChoice",
+        "TFMPNetForQuestionAnswering",
+        "TFMPNetForSequenceClassification",
+        "TFMPNetForTokenClassification",
+        "TFMPNetMainLayer",
+        "TFMPNetModel",
+        "TFMPNetPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_mpnet import MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP, MPNetConfig
     from .tokenization_mpnet import MPNetTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_mpnet_fast import MPNetTokenizerFast
 
@@ -50,3 +96,21 @@ if is_tf_available():
             TFMPNetModel,
             TFMPNetPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/mt5/__init__.py

@@ -16,9 +16,40 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_sentencepiece_available, is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_mt5 import MT5Config
 
+from ...file_utils import (
+    _BaseLazyModule,
+    is_sentencepiece_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+if is_sentencepiece_available():
+    from ..t5.tokenization_t5 import T5Tokenizer
+
+    MT5Tokenizer = T5Tokenizer
+
+if is_tokenizers_available():
+    from ..t5.tokenization_t5_fast import T5TokenizerFast
+
+    MT5TokenizerFast = T5TokenizerFast
+
+_import_structure = {
+    "configuration_mt5": ["MT5Config"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_mt5"] = ["MT5EncoderModel", "MT5ForConditionalGeneration", "MT5Model"]
+
+if is_tf_available():
+    _import_structure["modeling_tf_mt5"] = ["TFMT5EncoderModel", "TFMT5ForConditionalGeneration", "TFMT5Model"]
+
+
+if TYPE_CHECKING:
+    from .configuration_mt5 import MT5Config
 
     if is_sentencepiece_available():
         from ..t5.tokenization_t5 import T5Tokenizer
@@ -35,3 +66,29 @@ if is_torch_available():
 
     if is_tf_available():
         from .modeling_tf_mt5 import TFMT5EncoderModel, TFMT5ForConditionalGeneration, TFMT5Model
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+        def __getattr__(self, name):
+            if name == "MT5Tokenizer":
+                return MT5Tokenizer
+            elif name == name == "MT5TokenizerFast":
+                return MT5TokenizerFast
+            else:
+                return super().__getattr__(name)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/openai/__init__.py

@@ -16,11 +16,46 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_openai": ["OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP", "OpenAIGPTConfig"],
+    "tokenization_openai": ["OpenAIGPTTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_openai_fast"] = ["OpenAIGPTTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_openai"] = [
+        "OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "OpenAIGPTDoubleHeadsModel",
+        "OpenAIGPTForSequenceClassification",
+        "OpenAIGPTLMHeadModel",
+        "OpenAIGPTModel",
+        "OpenAIGPTPreTrainedModel",
+        "load_tf_weights_in_openai_gpt",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_openai"] = [
+        "TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFOpenAIGPTDoubleHeadsModel",
+        "TFOpenAIGPTForSequenceClassification",
+        "TFOpenAIGPTLMHeadModel",
+        "TFOpenAIGPTMainLayer",
+        "TFOpenAIGPTModel",
+        "TFOpenAIGPTPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_openai import OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP, OpenAIGPTConfig
     from .tokenization_openai import OpenAIGPTTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_openai_fast import OpenAIGPTTokenizerFast
 
@@ -45,3 +80,21 @@ if is_tf_available():
             TFOpenAIGPTModel,
             TFOpenAIGPTPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/pegasus/__init__.py

@@ -15,9 +15,41 @@
 # 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 ...file_utils import is_sentencepiece_available, is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_pegasus import PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP, PegasusConfig
 
+from ...file_utils import (
+    _BaseLazyModule,
+    is_sentencepiece_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_pegasus": ["PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP", "PegasusConfig"],
+}
+
+if is_sentencepiece_available():
+    _import_structure["tokenization_pegasus"] = ["PegasusTokenizer"]
+
+if is_tokenizers_available():
+    _import_structure["tokenization_pegasus_fast"] = ["PegasusTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_pegasus"] = [
+        "PEGASUS_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "PegasusForConditionalGeneration",
+        "PegasusModel",
+        "PegasusPreTrainedModel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_pegasus"] = ["TFPegasusForConditionalGeneration"]
+
+
+if TYPE_CHECKING:
+    from .configuration_pegasus import PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP, PegasusConfig
 
     if is_sentencepiece_available():
         from .tokenization_pegasus import PegasusTokenizer
@@ -35,3 +67,21 @@ if is_torch_available():
 
     if is_tf_available():
         from .modeling_tf_pegasus import TFPegasusForConditionalGeneration
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/phobert/__init__.py

@@ -16,4 +16,33 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule
+
+
+_import_structure = {
+    "tokenization_phobert": ["PhobertTokenizer"],
+}
+
+
+if TYPE_CHECKING:
     from .tokenization_phobert import PhobertTokenizer
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/prophetnet/__init__.py

@@ -16,11 +16,32 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_prophetnet": ["PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "ProphetNetConfig"],
+    "tokenization_prophetnet": ["ProphetNetTokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_prophetnet"] = [
+        "PROPHETNET_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "ProphetNetDecoder",
+        "ProphetNetEncoder",
+        "ProphetNetForCausalLM",
+        "ProphetNetForConditionalGeneration",
+        "ProphetNetModel",
+        "ProphetNetPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_prophetnet import PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP, ProphetNetConfig
     from .tokenization_prophetnet import ProphetNetTokenizer
 
-
     if is_torch_available():
         from .modeling_prophetnet import (
             PROPHETNET_PRETRAINED_MODEL_ARCHIVE_LIST,
@@ -31,3 +52,21 @@ if is_torch_available():
             ProphetNetModel,
             ProphetNetPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/rag/__init__.py

@@ -16,11 +16,43 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_rag": ["RagConfig"],
+    "retrieval_rag": ["RagRetriever"],
+    "tokenization_rag": ["RagTokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_rag"] = ["RagModel", "RagSequenceForGeneration", "RagTokenForGeneration"]
+
+
+if TYPE_CHECKING:
     from .configuration_rag import RagConfig
     from .retrieval_rag import RagRetriever
     from .tokenization_rag import RagTokenizer
 
-
     if is_torch_available():
         from .modeling_rag import RagModel, RagSequenceForGeneration, RagTokenForGeneration
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/reformer/__init__.py

@@ -16,9 +16,36 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_sentencepiece_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_reformer import REFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, ReformerConfig
 
+from ...file_utils import _BaseLazyModule, is_sentencepiece_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_reformer": ["REFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "ReformerConfig"],
+}
+
+if is_sentencepiece_available():
+    _import_structure["tokenization_reformer"] = ["ReformerTokenizer"]
+
+if is_tokenizers_available():
+    _import_structure["tokenization_reformer_fast"] = ["ReformerTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_reformer"] = [
+        "REFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "ReformerAttention",
+        "ReformerForMaskedLM",
+        "ReformerForQuestionAnswering",
+        "ReformerForSequenceClassification",
+        "ReformerLayer",
+        "ReformerModel",
+        "ReformerModelWithLMHead",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_reformer import REFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, ReformerConfig
 
     if is_sentencepiece_available():
         from .tokenization_reformer import ReformerTokenizer
@@ -37,3 +64,21 @@ if is_torch_available():
             ReformerModel,
             ReformerModelWithLMHead,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/retribert/__init__.py

@@ -16,13 +16,55 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_retribert": ["RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "RetriBertConfig"],
+    "tokenization_retribert": ["RetriBertTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_retribert_fast"] = ["RetriBertTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_retribert"] = [
+        "RETRIBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "RetriBertModel",
+        "RetriBertPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, RetriBertConfig
     from .tokenization_retribert import RetriBertTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_retribert_fast import RetriBertTokenizerFast
 
     if is_torch_available():
-    from .modeling_retribert import RETRIBERT_PRETRAINED_MODEL_ARCHIVE_LIST, RetriBertModel, RetriBertPreTrainedModel
+        from .modeling_retribert import (
+            RETRIBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            RetriBertModel,
+            RetriBertPreTrainedModel,
+        )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/roberta/__init__.py

@@ -16,11 +16,58 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import (
+    _BaseLazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_roberta": ["ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP", "RobertaConfig"],
+    "tokenization_roberta": ["RobertaTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_roberta_fast"] = ["RobertaTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_roberta"] = [
+        "ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "RobertaForCausalLM",
+        "RobertaForMaskedLM",
+        "RobertaForMultipleChoice",
+        "RobertaForQuestionAnswering",
+        "RobertaForSequenceClassification",
+        "RobertaForTokenClassification",
+        "RobertaModel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_roberta"] = [
+        "TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFRobertaForMaskedLM",
+        "TFRobertaForMultipleChoice",
+        "TFRobertaForQuestionAnswering",
+        "TFRobertaForSequenceClassification",
+        "TFRobertaForTokenClassification",
+        "TFRobertaMainLayer",
+        "TFRobertaModel",
+        "TFRobertaPreTrainedModel",
+    ]
+
+if is_flax_available():
+    _import_structure["modeling_flax_roberta"] = ["FlaxRobertaModel"]
+
+
+if TYPE_CHECKING:
     from .configuration_roberta import ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, RobertaConfig
     from .tokenization_roberta import RobertaTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_roberta_fast import RobertaTokenizerFast
 
@@ -51,3 +98,21 @@ if is_tf_available():
 
     if is_flax_available():
         from .modeling_flax_roberta import FlaxRobertaModel
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/squeezebert/__init__.py

@@ -16,11 +16,37 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_squeezebert": ["SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "SqueezeBertConfig"],
+    "tokenization_squeezebert": ["SqueezeBertTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_squeezebert_fast"] = ["SqueezeBertTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_squeezebert"] = [
+        "SQUEEZEBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "SqueezeBertForMaskedLM",
+        "SqueezeBertForMultipleChoice",
+        "SqueezeBertForQuestionAnswering",
+        "SqueezeBertForSequenceClassification",
+        "SqueezeBertForTokenClassification",
+        "SqueezeBertModel",
+        "SqueezeBertModule",
+        "SqueezeBertPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_squeezebert import SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, SqueezeBertConfig
     from .tokenization_squeezebert import SqueezeBertTokenizer
 
-
     if is_tokenizers_available():
         from .tokenization_squeezebert_fast import SqueezeBertTokenizerFast
 
@@ -36,3 +62,21 @@ if is_torch_available():
             SqueezeBertModule,
             SqueezeBertPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/t5/__init__.py

@@ -16,9 +16,49 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_sentencepiece_available, is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config
 
+from ...file_utils import (
+    _BaseLazyModule,
+    is_sentencepiece_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_t5": ["T5_PRETRAINED_CONFIG_ARCHIVE_MAP", "T5Config"],
+}
+
+if is_sentencepiece_available():
+    _import_structure["tokenization_t5"] = ["T5Tokenizer"]
+
+if is_tokenizers_available():
+    _import_structure["tokenization_t5_fast"] = ["T5TokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_t5"] = [
+        "T5_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "T5EncoderModel",
+        "T5ForConditionalGeneration",
+        "T5Model",
+        "T5PreTrainedModel",
+        "load_tf_weights_in_t5",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_t5"] = [
+        "TF_T5_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFT5EncoderModel",
+        "TFT5ForConditionalGeneration",
+        "TFT5Model",
+        "TFT5PreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config
 
     if is_sentencepiece_available():
         from .tokenization_t5 import T5Tokenizer
@@ -44,3 +84,21 @@ if is_tf_available():
             TFT5Model,
             TFT5PreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/tapas/__init__.py

@@ -16,11 +16,30 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_tapas": ["TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP", "TapasConfig"],
+    "tokenization_tapas": ["TapasTokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_tapas"] = [
+        "TAPAS_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TapasForMaskedLM",
+        "TapasForQuestionAnswering",
+        "TapasForSequenceClassification",
+        "TapasModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_tapas import TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP, TapasConfig
     from .tokenization_tapas import TapasTokenizer
 
-
     if is_torch_available():
         from .modeling_tapas import (
             TAPAS_PRETRAINED_MODEL_ARCHIVE_LIST,
@@ -29,3 +48,21 @@ if is_torch_available():
             TapasForSequenceClassification,
             TapasModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/transfo_xl/__init__.py

@@ -16,11 +16,43 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_transfo_xl": ["TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP", "TransfoXLConfig"],
+    "tokenization_transfo_xl": ["TransfoXLCorpus", "TransfoXLTokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_transfo_xl"] = [
+        "TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "AdaptiveEmbedding",
+        "TransfoXLForSequenceClassification",
+        "TransfoXLLMHeadModel",
+        "TransfoXLModel",
+        "TransfoXLPreTrainedModel",
+        "load_tf_weights_in_transfo_xl",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_transfo_xl"] = [
+        "TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFAdaptiveEmbedding",
+        "TFTransfoXLForSequenceClassification",
+        "TFTransfoXLLMHeadModel",
+        "TFTransfoXLMainLayer",
+        "TFTransfoXLModel",
+        "TFTransfoXLPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_transfo_xl import TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP, TransfoXLConfig
     from .tokenization_transfo_xl import TransfoXLCorpus, TransfoXLTokenizer
 
-
     if is_torch_available():
         from .modeling_transfo_xl import (
             TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST,
@@ -42,3 +74,21 @@ if is_tf_available():
             TFTransfoXLModel,
             TFTransfoXLPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/xlm/__init__.py

@@ -16,11 +16,47 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_tf_available, is_torch_available
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule, is_tf_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_xlm": ["XLM_PRETRAINED_CONFIG_ARCHIVE_MAP", "XLMConfig"],
+    "tokenization_xlm": ["XLMTokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_xlm"] = [
+        "XLM_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "XLMForMultipleChoice",
+        "XLMForQuestionAnswering",
+        "XLMForQuestionAnsweringSimple",
+        "XLMForSequenceClassification",
+        "XLMForTokenClassification",
+        "XLMModel",
+        "XLMPreTrainedModel",
+        "XLMWithLMHeadModel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_xlm"] = [
+        "TF_XLM_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFXLMForMultipleChoice",
+        "TFXLMForQuestionAnsweringSimple",
+        "TFXLMForSequenceClassification",
+        "TFXLMForTokenClassification",
+        "TFXLMMainLayer",
+        "TFXLMModel",
+        "TFXLMPreTrainedModel",
+        "TFXLMWithLMHeadModel",
+    ]
+
+
+if TYPE_CHECKING:
     from .configuration_xlm import XLM_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMConfig
     from .tokenization_xlm import XLMTokenizer
 
-
     if is_torch_available():
         from .modeling_xlm import (
             XLM_PRETRAINED_MODEL_ARCHIVE_LIST,
@@ -46,3 +82,21 @@ if is_tf_available():
             TFXLMPreTrainedModel,
             TFXLMWithLMHeadModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/xlm_roberta/__init__.py

@@ -16,9 +16,53 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_sentencepiece_available, is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_xlm_roberta import XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMRobertaConfig
 
+from ...file_utils import (
+    _BaseLazyModule,
+    is_sentencepiece_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_xlm_roberta": ["XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP", "XLMRobertaConfig"],
+}
+
+if is_sentencepiece_available():
+    _import_structure["tokenization_xlm_roberta"] = ["XLMRobertaTokenizer"]
+
+if is_tokenizers_available():
+    _import_structure["tokenization_xlm_roberta_fast"] = ["XLMRobertaTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_xlm_roberta"] = [
+        "XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "XLMRobertaForCausalLM",
+        "XLMRobertaForMaskedLM",
+        "XLMRobertaForMultipleChoice",
+        "XLMRobertaForQuestionAnswering",
+        "XLMRobertaForSequenceClassification",
+        "XLMRobertaForTokenClassification",
+        "XLMRobertaModel",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_xlm_roberta"] = [
+        "TF_XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFXLMRobertaForMaskedLM",
+        "TFXLMRobertaForMultipleChoice",
+        "TFXLMRobertaForQuestionAnswering",
+        "TFXLMRobertaForSequenceClassification",
+        "TFXLMRobertaForTokenClassification",
+        "TFXLMRobertaModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_xlm_roberta import XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMRobertaConfig
 
     if is_sentencepiece_available():
         from .tokenization_xlm_roberta import XLMRobertaTokenizer
@@ -48,3 +92,21 @@ if is_tf_available():
             TFXLMRobertaForTokenClassification,
             TFXLMRobertaModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/xlnet/__init__.py

@@ -16,9 +16,57 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...file_utils import is_sentencepiece_available, is_tf_available, is_tokenizers_available, is_torch_available
+from typing import TYPE_CHECKING
-from .configuration_xlnet import XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP, XLNetConfig
 
+from ...file_utils import (
+    _BaseLazyModule,
+    is_sentencepiece_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_xlnet": ["XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "XLNetConfig"],
+}
+
+if is_sentencepiece_available():
+    _import_structure["tokenization_xlnet"] = ["XLNetTokenizer"]
+
+if is_tokenizers_available():
+    _import_structure["tokenization_xlnet_fast"] = ["XLNetTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_xlnet"] = [
+        "XLNET_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "XLNetForMultipleChoice",
+        "XLNetForQuestionAnswering",
+        "XLNetForQuestionAnsweringSimple",
+        "XLNetForSequenceClassification",
+        "XLNetForTokenClassification",
+        "XLNetLMHeadModel",
+        "XLNetModel",
+        "XLNetPreTrainedModel",
+        "load_tf_weights_in_xlnet",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_xlnet"] = [
+        "TF_XLNET_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFXLNetForMultipleChoice",
+        "TFXLNetForQuestionAnsweringSimple",
+        "TFXLNetForSequenceClassification",
+        "TFXLNetForTokenClassification",
+        "TFXLNetLMHeadModel",
+        "TFXLNetMainLayer",
+        "TFXLNetModel",
+        "TFXLNetPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_xlnet import XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP, XLNetConfig
 
     if is_sentencepiece_available():
         from .tokenization_xlnet import XLNetTokenizer
@@ -52,3 +100,21 @@ if is_tf_available():
             TFXLNetModel,
             TFXLNetPreTrainedModel,
         )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/tokenization_utils_base.py

@@ -25,7 +25,7 @@ import warnings
 from collections import OrderedDict, UserDict
 from dataclasses import dataclass, field
 from enum import Enum
-from typing import Any, Dict, List, NamedTuple, Optional, Sequence, Tuple, Union
+from typing import TYPE_CHECKING, Any, Dict, List, NamedTuple, Optional, Sequence, Tuple, Union
 
 import numpy as np
 
@@ -45,21 +45,34 @@ from .file_utils import (
 from .utils import logging
 
 
-if is_tf_available():
+if TYPE_CHECKING:
-    import tensorflow as tf
-
     if is_torch_available():
         import torch
-
+    if is_tf_available():
+        import tensorflow as tf
     if is_flax_available():
-    import jax.numpy as jnp
+        import jax.numpy as jnp  # noqa: F401
 
 
 def _is_numpy(x):
     return isinstance(x, np.ndarray)
 
 
+def _is_torch(x):
+    import torch
+
+    return isinstance(x, torch.Tensor)
+
+
+def _is_tensorflow(x):
+    import tensorflow as tf
+
+    return isinstance(x, tf.Tensor)
+
+
 def _is_jax(x):
+    import jax.numpy as jnp  # noqa: F811
+
     return isinstance(x, jnp.ndarray)
 
 
@@ -196,9 +209,9 @@ def to_py_obj(obj):
         return {k: to_py_obj(v) for k, v in obj.items()}
     elif isinstance(obj, (list, tuple)):
         return [to_py_obj(o) for o in obj]
-    elif is_tf_available() and isinstance(obj, tf.Tensor):
+    elif is_tf_available() and _is_tensorflow(obj):
         return obj.numpy().tolist()
-    elif is_torch_available() and isinstance(obj, torch.Tensor):
+    elif is_torch_available() and _is_torch(obj):
         return obj.detach().cpu().tolist()
     elif isinstance(obj, np.ndarray):
         return obj.tolist()
@@ -714,16 +727,22 @@ class BatchEncoding(UserDict):
                 raise ImportError(
                     "Unable to convert output to TensorFlow tensors format, TensorFlow is not installed."
                 )
+            import tensorflow as tf
+
             as_tensor = tf.constant
             is_tensor = tf.is_tensor
         elif tensor_type == TensorType.PYTORCH:
             if not is_torch_available():
                 raise ImportError("Unable to convert output to PyTorch tensors format, PyTorch is not installed.")
+            import torch
+
             as_tensor = torch.tensor
             is_tensor = torch.is_tensor
         elif tensor_type == TensorType.JAX:
             if not is_flax_available():
                 raise ImportError("Unable to convert output to JAX tensors format, JAX is not installed.")
+            import jax.numpy as jnp  # noqa: F811
+
             as_tensor = jnp.array
             is_tensor = _is_jax
         else:
@@ -2684,9 +2703,9 @@ class PreTrainedTokenizerBase(SpecialTokensMixin):
                 first_element = encoded_inputs["input_ids"][index][0]
         # At this state, if `first_element` is still a list/tuple, it's an empty one so there is nothing to do.
         if not isinstance(first_element, (int, list, tuple)):
-            if is_tf_available() and isinstance(first_element, tf.Tensor):
+            if is_tf_available() and _is_tensorflow(first_element):
                 return_tensors = "tf" if return_tensors is None else return_tensors
-            elif is_torch_available() and isinstance(first_element, torch.Tensor):
+            elif is_torch_available() and _is_torch(first_element):
                 return_tensors = "pt" if return_tensors is None else return_tensors
             elif isinstance(first_element, np.ndarray):
                 return_tensors = "np" if return_tensors is None else return_tensors

templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/__init__.py

@@ -15,14 +15,78 @@
 # 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
 
 {%- if cookiecutter.generate_tensorflow_and_pytorch == "PyTorch & TensorFlow" %}
-from ...file_utils import is_tf_available, is_torch_available, is_tokenizers_available
+from ...file_utils import _BaseLazyModule, is_tf_available, is_torch_available, is_tokenizers_available
 {%- elif cookiecutter.generate_tensorflow_and_pytorch == "PyTorch" %}
-from ...file_utils import is_torch_available, is_tokenizers_available
+from ...file_utils import _BaseLazyModule, is_torch_available, is_tokenizers_available
 {%- elif cookiecutter.generate_tensorflow_and_pytorch == "TensorFlow" %}
-from ...file_utils import is_tf_available, is_tokenizers_available
+from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available
 {% endif %}
+_import_structure = {
+    "configuration_{{cookiecutter.lowercase_modelname}}": ["{{cookiecutter.uppercase_modelname}}_PRETRAINED_CONFIG_ARCHIVE_MAP", "{{cookiecutter.camelcase_modelname}}Config"],
+    "tokenization_{{cookiecutter.lowercase_modelname}}": ["{{cookiecutter.camelcase_modelname}}Tokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_{{cookiecutter.lowercase_modelname}}_fast"] = ["{{cookiecutter.camelcase_modelname}}TokenizerFast"]
+
+{%- if (cookiecutter.generate_tensorflow_and_pytorch == "PyTorch & TensorFlow" or cookiecutter.generate_tensorflow_and_pytorch == "PyTorch") %}
+{% if cookiecutter.is_encoder_decoder_model == "False" %}
+if is_torch_available():
+    _import_structure["modeling_{{cookiecutter.lowercase_modelname}}"] = [
+        "{{cookiecutter.uppercase_modelname}}_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "{{cookiecutter.camelcase_modelname}}ForMaskedLM",
+        "{{cookiecutter.camelcase_modelname}}ForCausalLM",
+        "{{cookiecutter.camelcase_modelname}}ForMultipleChoice",
+        "{{cookiecutter.camelcase_modelname}}ForQuestionAnswering",
+        "{{cookiecutter.camelcase_modelname}}ForSequenceClassification",
+        "{{cookiecutter.camelcase_modelname}}ForTokenClassification",
+        "{{cookiecutter.camelcase_modelname}}Layer",
+        "{{cookiecutter.camelcase_modelname}}Model",
+        "{{cookiecutter.camelcase_modelname}}PreTrainedModel",
+        "load_tf_weights_in_{{cookiecutter.lowercase_modelname}}",
+    ]
+{% else %}
+if is_torch_available():
+    _import_structure["modeling_{{cookiecutter.lowercase_modelname}}"] = [
+        "{{cookiecutter.uppercase_modelname}}_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "{{cookiecutter.camelcase_modelname}}ForConditionalGeneration",
+        "{{cookiecutter.camelcase_modelname}}ForQuestionAnswering",
+        "{{cookiecutter.camelcase_modelname}}ForSequenceClassification",
+        "{{cookiecutter.camelcase_modelname}}Model",
+        "{{cookiecutter.camelcase_modelname}}PreTrainedModel",
+    ]
+{% endif %}
+{% endif %}
+{%- if (cookiecutter.generate_tensorflow_and_pytorch == "PyTorch & TensorFlow" or cookiecutter.generate_tensorflow_and_pytorch == "TensorFlow") %}
+{% if cookiecutter.is_encoder_decoder_model == "False" %}
+if is_tf_available():
+    _import_structure["modeling_tf_{{cookiecutter.lowercase_modelname}}"] = [
+        "TF_{{cookiecutter.uppercase_modelname}}_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TF{{cookiecutter.camelcase_modelname}}ForMaskedLM",
+        "TF{{cookiecutter.camelcase_modelname}}ForCausalLM",
+        "TF{{cookiecutter.camelcase_modelname}}ForMultipleChoice",
+        "TF{{cookiecutter.camelcase_modelname}}ForQuestionAnswering",
+        "TF{{cookiecutter.camelcase_modelname}}ForSequenceClassification",
+        "TF{{cookiecutter.camelcase_modelname}}ForTokenClassification",
+        "TF{{cookiecutter.camelcase_modelname}}Layer",
+        "TF{{cookiecutter.camelcase_modelname}}Model",
+        "TF{{cookiecutter.camelcase_modelname}}PreTrainedModel",
+    ]
+{% else %}
+if is_tf_available():
+    _import_structure["modeling_tf_{{cookiecutter.lowercase_modelname}}"] = [
+        "TF{{cookiecutter.camelcase_modelname}}ForConditionalGeneration",
+        "TF{{cookiecutter.camelcase_modelname}}Model",
+        "TF{{cookiecutter.camelcase_modelname}}PreTrainedModel",
+    ]
+{% endif %}
+{% endif %}
+
+
+if TYPE_CHECKING:
     from .configuration_{{cookiecutter.lowercase_modelname}} import {{cookiecutter.uppercase_modelname}}_PRETRAINED_CONFIG_ARCHIVE_MAP, {{cookiecutter.camelcase_modelname}}Config
     from .tokenization_{{cookiecutter.lowercase_modelname}} import {{cookiecutter.camelcase_modelname}}Tokenizer
 
@@ -81,3 +145,20 @@ if is_tf_available():
         )
 {% endif %}
 {% endif %}
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)