[WIP] Fix Pyright static type checking by replacing if-else imports with try-except (#16578)

* rebase and isort

* modify cookiecutter init

* fix cookiecutter auto imports

* fix clean_frameworks_in_init

* fix add_model_to_main_init

* blackify

* replace unnecessary f-strings

* update yolos imports

* fix roberta import bug

* fix yolos missing dependency

* fix add_model_like and cookiecutter bug

* fix repository consistency error

* modify cookiecutter, fix add_new_model_like

* remove stale line

Co-authored-by: Dom Miketa <dmiketa@exscientia.co.uk>

src/transformers/__init__.py

@@ -29,6 +29,7 @@ from typing import TYPE_CHECKING
 # Check the dependencies satisfy the minimal versions required.
 from . import dependency_versions_check
 from .utils import (
+    OptionalDependencyNotAvailable,
     _LazyModule,
     is_flax_available,
     is_scatter_available,
@@ -412,7 +413,16 @@ _import_structure = {
 }
 
 # sentencepiece-backed objects
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_sentencepiece_objects
+
+    _import_structure["utils.dummy_sentencepiece_objects"] = [
+        name for name in dir(dummy_sentencepiece_objects) if not name.startswith("_")
+    ]
+else:
     _import_structure["models.albert"].append("AlbertTokenizer")
     _import_structure["models.barthez"].append("BarthezTokenizer")
     _import_structure["models.bartpho"].append("BartphoTokenizer")
@@ -439,16 +449,19 @@ if is_sentencepiece_available():
     _import_structure["models.xlm_prophetnet"].append("XLMProphetNetTokenizer")
     _import_structure["models.xlm_roberta"].append("XLMRobertaTokenizer")
     _import_structure["models.xlnet"].append("XLNetTokenizer")
-else:
-    from .utils import dummy_sentencepiece_objects
-
-    _import_structure["utils.dummy_sentencepiece_objects"] = [
-        name for name in dir(dummy_sentencepiece_objects) if not name.startswith("_")
-    ]
 
 # tokenizers-backed objects
-if is_tokenizers_available():
+try:
-    # Fast tokenizers
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_tokenizers_objects
+
+    _import_structure["utils.dummy_tokenizers_objects"] = [
+        name for name in dir(dummy_tokenizers_objects) if not name.startswith("_")
+    ]
+else:
+    # Fast tokenizers structure
     _import_structure["models.albert"].append("AlbertTokenizerFast")
     _import_structure["models.bart"].append("BartTokenizerFast")
     _import_structure["models.barthez"].append("BarthezTokenizerFast")
@@ -498,43 +511,55 @@ if is_tokenizers_available():
     _import_structure["models.xlnet"].append("XLNetTokenizerFast")
     _import_structure["tokenization_utils_fast"] = ["PreTrainedTokenizerFast"]
 
-else:
-    from .utils import dummy_tokenizers_objects
 
-    _import_structure["utils.dummy_tokenizers_objects"] = [
+try:
-        name for name in dir(dummy_tokenizers_objects) if not name.startswith("_")
+    if not (is_sentencepiece_available() and is_tokenizers_available()):
-    ]
+        raise OptionalDependencyNotAvailable()
-
+except OptionalDependencyNotAvailable:
-if is_sentencepiece_available() and is_tokenizers_available():
-    _import_structure["convert_slow_tokenizer"] = ["SLOW_TO_FAST_CONVERTERS", "convert_slow_tokenizer"]
-else:
     from .utils import dummy_sentencepiece_and_tokenizers_objects
 
     _import_structure["utils.dummy_sentencepiece_and_tokenizers_objects"] = [
         name for name in dir(dummy_sentencepiece_and_tokenizers_objects) if not name.startswith("_")
     ]
+else:
+    _import_structure["convert_slow_tokenizer"] = ["SLOW_TO_FAST_CONVERTERS", "convert_slow_tokenizer"]
 
 # Speech-specific objects
-if is_speech_available():
+try:
-    _import_structure["models.speech_to_text"].append("Speech2TextFeatureExtractor")
+    if not is_speech_available():
-else:
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
     from .utils import dummy_speech_objects
 
     _import_structure["utils.dummy_speech_objects"] = [
         name for name in dir(dummy_speech_objects) if not name.startswith("_")
     ]
-
-if is_sentencepiece_available() and is_speech_available():
-    _import_structure["models.speech_to_text"].append("Speech2TextProcessor")
 else:
+    _import_structure["models.speech_to_text"].append("Speech2TextFeatureExtractor")
+
+try:
+    if not (is_sentencepiece_available() and is_speech_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
     from .utils import dummy_sentencepiece_and_speech_objects
 
     _import_structure["utils.dummy_sentencepiece_and_speech_objects"] = [
         name for name in dir(dummy_sentencepiece_and_speech_objects) if not name.startswith("_")
     ]
+else:
+    _import_structure["models.speech_to_text"].append("Speech2TextProcessor")
 
 # Vision-specific objects
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_vision_objects
+
+    _import_structure["utils.dummy_vision_objects"] = [
+        name for name in dir(dummy_vision_objects) if not name.startswith("_")
+    ]
+else:
     _import_structure["image_utils"] = ["ImageFeatureExtractionMixin"]
     _import_structure["models.beit"].append("BeitFeatureExtractor")
     _import_structure["models.clip"].append("CLIPFeatureExtractor")
@@ -556,15 +581,18 @@ if is_vision_available():
     _import_structure["models.vilt"].append("ViltProcessor")
     _import_structure["models.vit"].append("ViTFeatureExtractor")
     _import_structure["models.yolos"].append("YolosFeatureExtractor")
-else:
-    from .utils import dummy_vision_objects
-
-    _import_structure["utils.dummy_vision_objects"] = [
-        name for name in dir(dummy_vision_objects) if not name.startswith("_")
-    ]
 
 # Timm-backed objects
-if is_timm_available() and is_vision_available():
+try:
+    if not (is_timm_available() and is_vision_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_timm_objects
+
+    _import_structure["utils.dummy_timm_objects"] = [
+        name for name in dir(dummy_timm_objects) if not name.startswith("_")
+    ]
+else:
     _import_structure["models.detr"].extend(
         [
             "DETR_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -574,14 +602,17 @@ if is_timm_available() and is_vision_available():
             "DetrPreTrainedModel",
         ]
     )
-else:
-    from .utils import dummy_timm_objects
 
-    _import_structure["utils.dummy_timm_objects"] = [
+try:
-        name for name in dir(dummy_timm_objects) if not name.startswith("_")
+    if not is_scatter_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_scatter_objects
+
+    _import_structure["utils.dummy_scatter_objects"] = [
+        name for name in dir(dummy_scatter_objects) if not name.startswith("_")
     ]
-
+else:
-if is_scatter_available():
     _import_structure["models.tapas"].extend(
         [
             "TAPAS_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -593,16 +624,17 @@ if is_scatter_available():
             "load_tf_weights_in_tapas",
         ]
     )
-else:
-    from .utils import dummy_scatter_objects
-
-    _import_structure["utils.dummy_scatter_objects"] = [
-        name for name in dir(dummy_scatter_objects) if not name.startswith("_")
-    ]
 
 
 # PyTorch-backed objects
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_pt_objects
+
+    _import_structure["utils.dummy_pt_objects"] = [name for name in dir(dummy_pt_objects) if not name.startswith("_")]
+else:
     _import_structure["activations"] = []
     _import_structure["benchmark.benchmark"] = ["PyTorchBenchmark"]
     _import_structure["benchmark.benchmark_args"] = ["PyTorchBenchmarkArguments"]
@@ -1723,13 +1755,16 @@ if is_torch_available():
     _import_structure["trainer"] = ["Trainer"]
     _import_structure["trainer_pt_utils"] = ["torch_distributed_zero_first"]
     _import_structure["trainer_seq2seq"] = ["Seq2SeqTrainer"]
-else:
-    from .utils import dummy_pt_objects
-
-    _import_structure["utils.dummy_pt_objects"] = [name for name in dir(dummy_pt_objects) if not name.startswith("_")]
 
 # TensorFlow-backed objects
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_tf_objects
+
+    _import_structure["utils.dummy_tf_objects"] = [name for name in dir(dummy_tf_objects) if not name.startswith("_")]
+else:
     _import_structure["activations_tf"] = []
     _import_structure["benchmark.benchmark_args_tf"] = ["TensorFlowBenchmarkArguments"]
     _import_structure["benchmark.benchmark_tf"] = ["TensorFlowBenchmark"]
@@ -2236,13 +2271,18 @@ if is_tf_available():
     _import_structure["tf_utils"] = []
     _import_structure["trainer_tf"] = ["TFTrainer"]
 
-else:
-    from .utils import dummy_tf_objects
-
-    _import_structure["utils.dummy_tf_objects"] = [name for name in dir(dummy_tf_objects) if not name.startswith("_")]
 
 # FLAX-backed objects
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_flax_objects
+
+    _import_structure["utils.dummy_flax_objects"] = [
+        name for name in dir(dummy_flax_objects) if not name.startswith("_")
+    ]
+else:
     _import_structure["generation_flax_logits_process"] = [
         "FlaxForcedBOSTokenLogitsProcessor",
         "FlaxForcedEOSTokenLogitsProcessor",
@@ -2469,12 +2509,6 @@ if is_flax_available():
             "FlaxXLMRobertaModel",
         ]
     )
-else:
-    from .utils import dummy_flax_objects
-
-    _import_structure["utils.dummy_flax_objects"] = [
-        name for name in dir(dummy_flax_objects) if not name.startswith("_")
-    ]
 
 
 # Direct imports for type-checking
@@ -2816,7 +2850,12 @@ if TYPE_CHECKING:
         logging,
     )
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_sentencepiece_objects import *
+    else:
         from .models.albert import AlbertTokenizer
         from .models.barthez import BarthezTokenizer
         from .models.bartpho import BartphoTokenizer
@@ -2842,10 +2881,14 @@ if TYPE_CHECKING:
         from .models.xlm_prophetnet import XLMProphetNetTokenizer
         from .models.xlm_roberta import XLMRobertaTokenizer
         from .models.xlnet import XLNetTokenizer
-    else:
-        from .utils.dummy_sentencepiece_objects import *
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_tokenizers_objects import *
+    else:
+        # Fast tokenizers imports
         from .models.albert import AlbertTokenizerFast
         from .models.bart import BartTokenizerFast
         from .models.barthez import BarthezTokenizerFast
@@ -2893,25 +2936,36 @@ if TYPE_CHECKING:
         from .models.xlnet import XLNetTokenizerFast
         from .tokenization_utils_fast import PreTrainedTokenizerFast
 
-    else:
+    try:
-        from .utils.dummy_tokenizers_objects import *
+        if not (is_sentencepiece_available() and is_tokenizers_available()):
-
+            raise OptionalDependencyNotAvailable()
-    if is_sentencepiece_available() and is_tokenizers_available():
+    except OptionalDependencyNotAvailable:
-        from .convert_slow_tokenizer import SLOW_TO_FAST_CONVERTERS, convert_slow_tokenizer
-    else:
         from .utils.dummies_sentencepiece_and_tokenizers_objects import *
-
-    if is_speech_available():
-        from .models.speech_to_text import Speech2TextFeatureExtractor
     else:
+        from .convert_slow_tokenizer import SLOW_TO_FAST_CONVERTERS, convert_slow_tokenizer
+
+    try:
+        if not is_speech_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
         from .utils.dummy_speech_objects import *
-
-    if is_speech_available() and is_sentencepiece_available():
-        from .models.speech_to_text import Speech2TextProcessor
     else:
-        from .utils.dummy_sentencepiece_and_speech_objects import *
+        from .models.speech_to_text import Speech2TextFeatureExtractor
 
-    if is_vision_available():
+    try:
+        if not (is_speech_available() and is_sentencepiece_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_sentencepiece_and_speech_objects import *
+    else:
+        from .models.speech_to_text import Speech2TextProcessor
+
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_vision_objects import *
+    else:
         from .image_utils import ImageFeatureExtractionMixin
         from .models.beit import BeitFeatureExtractor
         from .models.clip import CLIPFeatureExtractor, CLIPProcessor
@@ -2930,11 +2984,14 @@ if TYPE_CHECKING:
         from .models.vilt import ViltFeatureExtractor, ViltProcessor
         from .models.vit import ViTFeatureExtractor
         from .models.yolos import YolosFeatureExtractor
-    else:
-        from .utils.dummy_vision_objects import *
 
     # Modeling
-    if is_timm_available() and is_vision_available():
+    try:
+        if not (is_timm_available() and is_vision_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_timm_objects import *
+    else:
         from .models.detr import (
             DETR_PRETRAINED_MODEL_ARCHIVE_LIST,
             DetrForObjectDetection,
@@ -2942,10 +2999,13 @@ if TYPE_CHECKING:
             DetrModel,
             DetrPreTrainedModel,
         )
-    else:
-        from .utils.dummy_timm_objects import *
 
-    if is_scatter_available():
+    try:
+        if not is_scatter_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_scatter_objects import *
+    else:
         from .models.tapas import (
             TAPAS_PRETRAINED_MODEL_ARCHIVE_LIST,
             TapasForMaskedLM,
@@ -2955,10 +3015,13 @@ if TYPE_CHECKING:
             TapasPreTrainedModel,
             load_tf_weights_in_tapas,
         )
-    else:
-        from .utils.dummy_scatter_objects import *
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_pt_objects import *
+    else:
         # Benchmarks
         from .benchmark.benchmark import PyTorchBenchmark
         from .benchmark.benchmark_args import PyTorchBenchmarkArguments
@@ -3005,6 +3068,8 @@ if TYPE_CHECKING:
         )
         from .generation_utils import top_k_top_p_filtering
         from .modeling_utils import PreTrainedModel
+
+        # PyTorch model imports
         from .models.albert import (
             ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             AlbertForMaskedLM,
@@ -3896,12 +3961,16 @@ if TYPE_CHECKING:
         from .trainer import Trainer
         from .trainer_pt_utils import torch_distributed_zero_first
         from .trainer_seq2seq import Seq2SeqTrainer
-    else:
-        from .utils.dummy_pt_objects import *
 
     # TensorFlow
-    if is_tf_available():
+    try:
-
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        # Import the same objects as dummies to get them in the namespace.
+        # They will raise an import error if the user tries to instantiate / use them.
+        from .utils.dummy_tf_objects import *
+    else:
         from .benchmark.benchmark_args_tf import TensorFlowBenchmarkArguments
 
         # Benchmarks
@@ -3932,6 +4001,8 @@ if TYPE_CHECKING:
             TFLayoutLMPreTrainedModel,
         )
         from .modeling_tf_utils import TFPreTrainedModel, TFSequenceSummary, TFSharedEmbeddings, shape_list
+
+        # TensorFlow model imports
         from .models.albert import (
             TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFAlbertForMaskedLM,
@@ -4310,13 +4381,14 @@ if TYPE_CHECKING:
         # Trainer
         from .trainer_tf import TFTrainer
 
-    else:
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
         # Import the same objects as dummies to get them in the namespace.
         # They will raise an import error if the user tries to instantiate / use them.
-        from .utils.dummy_tf_objects import *
+        from .utils.dummy_flax_objects import *
-
+    else:
-    if is_flax_available():
-
         from .generation_flax_logits_process import (
             FlaxForcedBOSTokenLogitsProcessor,
             FlaxForcedEOSTokenLogitsProcessor,
@@ -4329,6 +4401,8 @@ if TYPE_CHECKING:
             FlaxTopPLogitsWarper,
         )
         from .modeling_flax_utils import FlaxPreTrainedModel
+
+        # Flax model imports
         from .models.albert import (
             FlaxAlbertForMaskedLM,
             FlaxAlbertForMultipleChoice,
@@ -4494,10 +4568,6 @@ if TYPE_CHECKING:
             FlaxXLMRobertaForTokenClassification,
             FlaxXLMRobertaModel,
         )
-    else:
-        # Import the same objects as dummies to get them in the namespace.
-        # They will raise an import error if the user tries to instantiate / use them.
-        from .utils.dummy_flax_objects import *
 
 else:
     import sys

src/transformers/commands/add_new_model_like.py

@@ -766,7 +766,9 @@ def clean_frameworks_in_init(
         return
 
     remove_pattern = "|".join(to_remove)
-    re_conditional_imports = re.compile(rf"^\s*if is_({remove_pattern})_available\(\):\s*$")
+    re_conditional_imports = re.compile(rf"^\s*if not is_({remove_pattern})_available\(\):\s*$")
+    re_try = re.compile(r"\s*try:")
+    re_else = re.compile(r"\s*else:")
     re_is_xxx_available = re.compile(rf"is_({remove_pattern})_available")
 
     with open(init_file, "r", encoding="utf-8") as f:
@@ -776,11 +778,15 @@ def clean_frameworks_in_init(
     new_lines = []
     idx = 0
     while idx < len(lines):
-        # Conditional imports
+        # Conditional imports in try-except-else blocks
-        if re_conditional_imports.search(lines[idx]) is not None:
+        if (re_conditional_imports.search(lines[idx]) is not None) and (re_try.search(lines[idx - 1]) is not None):
+            # Remove the preceding `try:`
+            new_lines.pop()
             idx += 1
-            while is_empty_line(lines[idx]):
+            # Iterate until `else:`
+            while is_empty_line(lines[idx]) or re_else.search(lines[idx]) is None:
                 idx += 1
+            idx += 1
             indent = find_indent(lines[idx])
             while find_indent(lines[idx]) >= indent or is_empty_line(lines[idx]):
                 idx += 1
@@ -790,6 +796,7 @@ def clean_frameworks_in_init(
             for framework in to_remove:
                 line = line.replace(f", is_{framework}_available", "")
                 line = line.replace(f"is_{framework}_available, ", "")
+                line = line.replace(f"is_{framework}_available,", "")
                 line = line.replace(f"is_{framework}_available", "")
 
             if len(line.strip()) > 0:
@@ -836,11 +843,11 @@ def add_model_to_main_init(
     while idx < len(lines):
         if not is_empty_line(lines[idx]) and find_indent(lines[idx]) == 0:
             framework = None
-        elif lines[idx].lstrip().startswith("if is_torch_available"):
+        elif lines[idx].lstrip().startswith("if not is_torch_available"):
             framework = "pt"
-        elif lines[idx].lstrip().startswith("if is_tf_available"):
+        elif lines[idx].lstrip().startswith("if not is_tf_available"):
             framework = "tf"
-        elif lines[idx].lstrip().startswith("if is_flax_available"):
+        elif lines[idx].lstrip().startswith("if not is_flax_available"):
             framework = "flax"
 
         # Skip if we are in a framework not wanted.

src/transformers/models/albert/__init__.py

@@ -19,6 +19,7 @@
 from typing import TYPE_CHECKING
 
 from ...utils import (
+    OptionalDependencyNotAvailable,
     _LazyModule,
     is_flax_available,
     is_sentencepiece_available,
@@ -32,13 +33,28 @@ _import_structure = {
     "configuration_albert": ["ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "AlbertConfig", "AlbertOnnxConfig"],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_albert"] = ["AlbertTokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_albert_fast"] = ["AlbertTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_albert"] = [
         "ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "AlbertForMaskedLM",
@@ -52,7 +68,12 @@ if is_torch_available():
         "load_tf_weights_in_albert",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_albert"] = [
         "TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFAlbertForMaskedLM",
@@ -66,7 +87,12 @@ if is_tf_available():
         "TFAlbertPreTrainedModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_albert"] = [
         "FlaxAlbertForMaskedLM",
         "FlaxAlbertForMultipleChoice",
@@ -81,13 +107,28 @@ if is_flax_available():
 if TYPE_CHECKING:
     from .configuration_albert import ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, AlbertConfig, AlbertOnnxConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_albert import AlbertTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_albert_fast import AlbertTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_albert import (
             ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             AlbertForMaskedLM,
@@ -101,7 +142,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_albert,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_albert import (
             TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFAlbertForMaskedLM,
@@ -115,7 +161,12 @@ if TYPE_CHECKING:
             TFAlbertPreTrainedModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_albert import (
             FlaxAlbertForMaskedLM,
             FlaxAlbertForMultipleChoice,

src/transformers/models/auto/__init__.py

@@ -18,7 +18,13 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -29,7 +35,12 @@ _import_structure = {
     "tokenization_auto": ["TOKENIZER_MAPPING", "AutoTokenizer"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_auto"] = [
         "MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING",
         "MODEL_FOR_AUDIO_XVECTOR_MAPPING",
@@ -81,7 +92,12 @@ if is_torch_available():
         "AutoModelWithLMHead",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_auto"] = [
         "TF_MODEL_FOR_CAUSAL_LM_MAPPING",
         "TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING",
@@ -115,7 +131,12 @@ if is_tf_available():
         "TFAutoModelWithLMHead",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_auto"] = [
         "FLAX_MODEL_FOR_CAUSAL_LM_MAPPING",
         "FLAX_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING",
@@ -151,7 +172,12 @@ if TYPE_CHECKING:
     from .processing_auto import PROCESSOR_MAPPING, AutoProcessor
     from .tokenization_auto import TOKENIZER_MAPPING, AutoTokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_auto import (
             MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING,
             MODEL_FOR_AUDIO_XVECTOR_MAPPING,
@@ -203,7 +229,12 @@ if TYPE_CHECKING:
             AutoModelWithLMHead,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_auto import (
             TF_MODEL_FOR_CAUSAL_LM_MAPPING,
             TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
@@ -237,7 +268,12 @@ if TYPE_CHECKING:
             TFAutoModelWithLMHead,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_auto import (
             FLAX_MODEL_FOR_CAUSAL_LM_MAPPING,
             FLAX_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,

src/transformers/models/bart/__init__.py

@@ -17,7 +17,14 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -25,10 +32,20 @@ _import_structure = {
     "tokenization_bart": ["BartTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_bart_fast"] = ["BartTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_bart"] = [
         "BART_PRETRAINED_MODEL_ARCHIVE_LIST",
         "BartForCausalLM",
@@ -40,10 +57,20 @@ if is_torch_available():
         "PretrainedBartModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_bart"] = ["TFBartForConditionalGeneration", "TFBartModel", "TFBartPretrainedModel"]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_bart"] = [
         "FlaxBartDecoderPreTrainedModel",
         "FlaxBartForCausalLM",
@@ -58,10 +85,20 @@ if TYPE_CHECKING:
     from .configuration_bart import BART_PRETRAINED_CONFIG_ARCHIVE_MAP, BartConfig, BartOnnxConfig
     from .tokenization_bart import BartTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_bart_fast import BartTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_bart import (
             BART_PRETRAINED_MODEL_ARCHIVE_LIST,
             BartForCausalLM,
@@ -73,10 +110,20 @@ if TYPE_CHECKING:
             PretrainedBartModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_bart import TFBartForConditionalGeneration, TFBartModel, TFBartPretrainedModel
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_bart import (
             FlaxBartDecoderPreTrainedModel,
             FlaxBartForCausalLM,

src/transformers/models/barthez/__init__.py

@@ -18,24 +18,44 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_sentencepiece_available, is_tokenizers_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_sentencepiece_available, is_tokenizers_available
 
 
 _import_structure = {}
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_barthez"] = ["BarthezTokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_barthez_fast"] = ["BarthezTokenizerFast"]
 
 
 if TYPE_CHECKING:
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_barthez import BarthezTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_barthez_fast import BarthezTokenizerFast
 
 else:

src/transformers/models/bartpho/__init__.py

@@ -18,16 +18,26 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_sentencepiece_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_sentencepiece_available
 
 
 _import_structure = {}
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_bartpho"] = ["BartphoTokenizer"]
 
 if TYPE_CHECKING:
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_bartpho import BartphoTokenizer
 
 else:

src/transformers/models/beit/__init__.py

@@ -18,17 +18,33 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_torch_available, is_vision_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_torch_available,
+    is_vision_available,
+)
 
 
 _import_structure = {
     "configuration_beit": ["BEIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "BeitConfig", "BeitOnnxConfig"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_beit"] = ["BeitFeatureExtractor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_beit"] = [
         "BEIT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "BeitForImageClassification",
@@ -39,7 +55,12 @@ if is_torch_available():
     ]
 
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_beit"] = [
         "FlaxBeitForImageClassification",
         "FlaxBeitForMaskedImageModeling",
@@ -50,10 +71,20 @@ if is_flax_available():
 if TYPE_CHECKING:
     from .configuration_beit import BEIT_PRETRAINED_CONFIG_ARCHIVE_MAP, BeitConfig, BeitOnnxConfig
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_beit import BeitFeatureExtractor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_beit import (
             BEIT_PRETRAINED_MODEL_ARCHIVE_LIST,
             BeitForImageClassification,
@@ -63,7 +94,12 @@ if TYPE_CHECKING:
             BeitPreTrainedModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_beit import (
             FlaxBeitForImageClassification,
             FlaxBeitForMaskedImageModeling,

src/transformers/models/bert/__init__.py

@@ -18,7 +18,14 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -26,10 +33,20 @@ _import_structure = {
     "tokenization_bert": ["BasicTokenizer", "BertTokenizer", "WordpieceTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_bert_fast"] = ["BertTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_bert"] = [
         "BERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "BertForMaskedLM",
@@ -46,7 +63,12 @@ if is_torch_available():
         "load_tf_weights_in_bert",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_bert"] = [
         "TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFBertEmbeddings",
@@ -63,7 +85,12 @@ if is_tf_available():
         "TFBertPreTrainedModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_bert"] = [
         "FlaxBertForCausalLM",
         "FlaxBertForMaskedLM",
@@ -81,10 +108,20 @@ if TYPE_CHECKING:
     from .configuration_bert import BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, BertConfig, BertOnnxConfig
     from .tokenization_bert import BasicTokenizer, BertTokenizer, WordpieceTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_bert_fast import BertTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_bert import (
             BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             BertForMaskedLM,
@@ -101,7 +138,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_bert,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_bert import (
             TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFBertEmbeddings,
@@ -118,7 +160,12 @@ if TYPE_CHECKING:
             TFBertPreTrainedModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_bert import (
             FlaxBertForCausalLM,
             FlaxBertForMaskedLM,

src/transformers/models/bert_generation/__init__.py

@@ -18,17 +18,27 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_sentencepiece_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_sentencepiece_available, is_torch_available
 
 
 _import_structure = {
     "configuration_bert_generation": ["BertGenerationConfig"],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_bert_generation"] = ["BertGenerationTokenizer"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_bert_generation"] = [
         "BertGenerationDecoder",
         "BertGenerationEncoder",
@@ -40,10 +50,20 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_bert_generation import BertGenerationConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_bert_generation import BertGenerationTokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_bert_generation import (
             BertGenerationDecoder,
             BertGenerationEncoder,

src/transformers/models/big_bird/__init__.py

@@ -18,6 +18,7 @@
 from typing import TYPE_CHECKING
 
 from ...utils import (
+    OptionalDependencyNotAvailable,
     _LazyModule,
     is_flax_available,
     is_sentencepiece_available,
@@ -31,13 +32,28 @@ _import_structure = {
     "configuration_big_bird": ["BIG_BIRD_PRETRAINED_CONFIG_ARCHIVE_MAP", "BigBirdConfig", "BigBirdOnnxConfig"],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_big_bird"] = ["BigBirdTokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_big_bird_fast"] = ["BigBirdTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_big_bird"] = [
         "BIG_BIRD_PRETRAINED_MODEL_ARCHIVE_LIST",
         "BigBirdForCausalLM",
@@ -53,7 +69,12 @@ if is_torch_available():
         "load_tf_weights_in_big_bird",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_big_bird"] = [
         "FlaxBigBirdForCausalLM",
         "FlaxBigBirdForMaskedLM",
@@ -69,13 +90,28 @@ if is_flax_available():
 if TYPE_CHECKING:
     from .configuration_big_bird import BIG_BIRD_PRETRAINED_CONFIG_ARCHIVE_MAP, BigBirdConfig, BigBirdOnnxConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_big_bird import BigBirdTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_big_bird_fast import BigBirdTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_big_bird import (
             BIG_BIRD_PRETRAINED_MODEL_ARCHIVE_LIST,
             BigBirdForCausalLM,
@@ -91,7 +127,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_big_bird,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_big_bird import (
             FlaxBigBirdForCausalLM,
             FlaxBigBirdForMaskedLM,

src/transformers/models/bigbird_pegasus/__init__.py

@@ -17,7 +17,7 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
@@ -28,7 +28,12 @@ _import_structure = {
     ],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_bigbird_pegasus"] = [
         "BIGBIRD_PEGASUS_PRETRAINED_MODEL_ARCHIVE_LIST",
         "BigBirdPegasusForCausalLM",
@@ -47,7 +52,12 @@ if TYPE_CHECKING:
         BigBirdPegasusOnnxConfig,
     )
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_bigbird_pegasus import (
             BIGBIRD_PEGASUS_PRETRAINED_MODEL_ARCHIVE_LIST,
             BigBirdPegasusForCausalLM,

src/transformers/models/blenderbot/__init__.py

@@ -18,7 +18,14 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -30,10 +37,20 @@ _import_structure = {
     "tokenization_blenderbot": ["BlenderbotTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_blenderbot_fast"] = ["BlenderbotTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_blenderbot"] = [
         "BLENDERBOT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "BlenderbotForCausalLM",
@@ -43,7 +60,12 @@ if is_torch_available():
     ]
 
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_blenderbot"] = [
         "TFBlenderbotForConditionalGeneration",
         "TFBlenderbotModel",
@@ -51,7 +73,12 @@ if is_tf_available():
     ]
 
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_blenderbot"] = [
         "FlaxBlenderbotForConditionalGeneration",
         "FlaxBlenderbotModel",
@@ -67,10 +94,20 @@ if TYPE_CHECKING:
     )
     from .tokenization_blenderbot import BlenderbotTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_blenderbot_fast import BlenderbotTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_blenderbot import (
             BLENDERBOT_PRETRAINED_MODEL_ARCHIVE_LIST,
             BlenderbotForCausalLM,
@@ -79,14 +116,24 @@ if TYPE_CHECKING:
             BlenderbotPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_blenderbot import (
             TFBlenderbotForConditionalGeneration,
             TFBlenderbotModel,
             TFBlenderbotPreTrainedModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_blenderbot import (
             FlaxBlenderbotForConditionalGeneration,
             FlaxBlenderbotModel,

src/transformers/models/blenderbot_small/__init__.py

@@ -17,7 +17,14 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -29,10 +36,20 @@ _import_structure = {
     "tokenization_blenderbot_small": ["BlenderbotSmallTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_blenderbot_small_fast"] = ["BlenderbotSmallTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_blenderbot_small"] = [
         "BLENDERBOT_SMALL_PRETRAINED_MODEL_ARCHIVE_LIST",
         "BlenderbotSmallForCausalLM",
@@ -41,14 +58,24 @@ if is_torch_available():
         "BlenderbotSmallPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_blenderbot_small"] = [
         "TFBlenderbotSmallForConditionalGeneration",
         "TFBlenderbotSmallModel",
         "TFBlenderbotSmallPreTrainedModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_blenderbot_small"] = [
         "FlaxBlenderbotSmallForConditionalGeneration",
         "FlaxBlenderbotSmallModel",
@@ -63,10 +90,20 @@ if TYPE_CHECKING:
     )
     from .tokenization_blenderbot_small import BlenderbotSmallTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_blenderbot_small_fast import BlenderbotSmallTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_blenderbot_small import (
             BLENDERBOT_SMALL_PRETRAINED_MODEL_ARCHIVE_LIST,
             BlenderbotSmallForCausalLM,
@@ -75,14 +112,24 @@ if TYPE_CHECKING:
             BlenderbotSmallPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_blenderbot_small import (
             TFBlenderbotSmallForConditionalGeneration,
             TFBlenderbotSmallModel,
             TFBlenderbotSmallPreTrainedModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_blenderbot_small import (
             FlaxBlenderbotSmallForConditionalGeneration,
             FlaxBlenderbotSmallModel,

src/transformers/models/camembert/__init__.py

@@ -19,6 +19,7 @@
 from typing import TYPE_CHECKING
 
 from ...utils import (
+    OptionalDependencyNotAvailable,
     _LazyModule,
     is_sentencepiece_available,
     is_tf_available,
@@ -31,13 +32,28 @@ _import_structure = {
     "configuration_camembert": ["CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "CamembertConfig", "CamembertOnnxConfig"],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_camembert"] = ["CamembertTokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_camembert_fast"] = ["CamembertTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_camembert"] = [
         "CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "CamembertForCausalLM",
@@ -49,7 +65,12 @@ if is_torch_available():
         "CamembertModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_camembert"] = [
         "TF_CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFCamembertForCausalLM",
@@ -65,13 +86,28 @@ if is_tf_available():
 if TYPE_CHECKING:
     from .configuration_camembert import CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, CamembertConfig, CamembertOnnxConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_camembert import CamembertTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_camembert_fast import CamembertTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_camembert import (
             CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             CamembertForCausalLM,
@@ -83,7 +119,12 @@ if TYPE_CHECKING:
             CamembertModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_camembert import (
             TF_CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFCamembertForCausalLM,

src/transformers/models/canine/__init__.py

@@ -17,7 +17,7 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tokenizers_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
 
 
 _import_structure = {
@@ -25,7 +25,12 @@ _import_structure = {
     "tokenization_canine": ["CanineTokenizer"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_canine"] = [
         "CANINE_PRETRAINED_MODEL_ARCHIVE_LIST",
         "CanineForMultipleChoice",
@@ -43,7 +48,12 @@ if TYPE_CHECKING:
     from .configuration_canine import CANINE_PRETRAINED_CONFIG_ARCHIVE_MAP, CanineConfig
     from .tokenization_canine import CanineTokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_canine import (
             CANINE_PRETRAINED_MODEL_ARCHIVE_LIST,
             CanineForMultipleChoice,

src/transformers/models/clip/__init__.py

@@ -18,6 +18,7 @@
 from typing import TYPE_CHECKING
 
 from ...utils import (
+    OptionalDependencyNotAvailable,
     _LazyModule,
     is_flax_available,
     is_tf_available,
@@ -32,14 +33,29 @@ _import_structure = {
     "tokenization_clip": ["CLIPTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_clip_fast"] = ["CLIPTokenizerFast"]
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_clip"] = ["CLIPFeatureExtractor"]
     _import_structure["processing_clip"] = ["CLIPProcessor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_clip"] = [
         "CLIP_PRETRAINED_MODEL_ARCHIVE_LIST",
         "CLIPModel",
@@ -48,7 +64,12 @@ if is_torch_available():
         "CLIPVisionModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_clip"] = [
         "TF_CLIP_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFCLIPModel",
@@ -57,7 +78,12 @@ if is_tf_available():
         "TFCLIPVisionModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_clip"] = [
         "FlaxCLIPModel",
         "FlaxCLIPPreTrainedModel",
@@ -72,14 +98,29 @@ if TYPE_CHECKING:
     from .configuration_clip import CLIP_PRETRAINED_CONFIG_ARCHIVE_MAP, CLIPConfig, CLIPTextConfig, CLIPVisionConfig
     from .tokenization_clip import CLIPTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_clip_fast import CLIPTokenizerFast
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_clip import CLIPFeatureExtractor
         from .processing_clip import CLIPProcessor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_clip import (
             CLIP_PRETRAINED_MODEL_ARCHIVE_LIST,
             CLIPModel,
@@ -88,7 +129,12 @@ if TYPE_CHECKING:
             CLIPVisionModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_clip import (
             TF_CLIP_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFCLIPModel,
@@ -97,7 +143,12 @@ if TYPE_CHECKING:
             TFCLIPVisionModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_clip import (
             FlaxCLIPModel,
             FlaxCLIPPreTrainedModel,

src/transformers/models/convbert/__init__.py

@@ -17,7 +17,13 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -25,10 +31,20 @@ _import_structure = {
     "tokenization_convbert": ["ConvBertTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_convbert_fast"] = ["ConvBertTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_convbert"] = [
         "CONVBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "ConvBertForMaskedLM",
@@ -43,7 +59,12 @@ if is_torch_available():
     ]
 
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_convbert"] = [
         "TF_CONVBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFConvBertForMaskedLM",
@@ -61,10 +82,20 @@ if TYPE_CHECKING:
     from .configuration_convbert import CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, ConvBertConfig, ConvBertOnnxConfig
     from .tokenization_convbert import ConvBertTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_convbert_fast import ConvBertTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_convbert import (
             CONVBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             ConvBertForMaskedLM,
@@ -78,7 +109,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_convbert,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_convbert import (
             TF_CONVBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFConvBertForMaskedLM,

src/transformers/models/convnext/__init__.py

@@ -18,17 +18,33 @@
 from typing import TYPE_CHECKING
 
 # rely on isort to merge the imports
-from ...utils import _LazyModule, is_tf_available, is_torch_available, is_vision_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tf_available,
+    is_torch_available,
+    is_vision_available,
+)
 
 
 _import_structure = {
     "configuration_convnext": ["CONVNEXT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ConvNextConfig"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_convnext"] = ["ConvNextFeatureExtractor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_convnext"] = [
         "CONVNEXT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "ConvNextForImageClassification",
@@ -36,7 +52,12 @@ if is_torch_available():
         "ConvNextPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_convnext"] = [
         "TFConvNextForImageClassification",
         "TFConvNextModel",
@@ -46,10 +67,20 @@ if is_tf_available():
 if TYPE_CHECKING:
     from .configuration_convnext import CONVNEXT_PRETRAINED_CONFIG_ARCHIVE_MAP, ConvNextConfig
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_convnext import ConvNextFeatureExtractor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_convnext import (
             CONVNEXT_PRETRAINED_MODEL_ARCHIVE_LIST,
             ConvNextForImageClassification,
@@ -57,7 +88,12 @@ if TYPE_CHECKING:
             ConvNextPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_convnext import TFConvNextForImageClassification, TFConvNextModel, TFConvNextPreTrainedModel
 
 

src/transformers/models/cpm/__init__.py

@@ -18,23 +18,43 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_sentencepiece_available, is_tokenizers_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_sentencepiece_available, is_tokenizers_available
 
 
 _import_structure = {}
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_cpm"] = ["CpmTokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_cpm_fast"] = ["CpmTokenizerFast"]
 
 
 if TYPE_CHECKING:
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_cpm import CpmTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_cpm_fast import CpmTokenizerFast
 
 else:

src/transformers/models/ctrl/__init__.py

@@ -18,7 +18,7 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tf_available, is_torch_available
 
 
 _import_structure = {
@@ -26,7 +26,12 @@ _import_structure = {
     "tokenization_ctrl": ["CTRLTokenizer"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_ctrl"] = [
         "CTRL_PRETRAINED_MODEL_ARCHIVE_LIST",
         "CTRLForSequenceClassification",
@@ -35,7 +40,12 @@ if is_torch_available():
         "CTRLPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_ctrl"] = [
         "TF_CTRL_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFCTRLForSequenceClassification",
@@ -49,7 +59,12 @@ if TYPE_CHECKING:
     from .configuration_ctrl import CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP, CTRLConfig
     from .tokenization_ctrl import CTRLTokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_ctrl import (
             CTRL_PRETRAINED_MODEL_ARCHIVE_LIST,
             CTRLForSequenceClassification,
@@ -58,7 +73,12 @@ if TYPE_CHECKING:
             CTRLPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_ctrl import (
             TF_CTRL_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFCTRLForSequenceClassification,

src/transformers/models/data2vec/__init__.py

@@ -18,9 +18,7 @@
 
 from typing import TYPE_CHECKING
 
-from transformers.utils.import_utils import is_tf_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tf_available, is_torch_available
-
-from ...utils import _LazyModule, is_torch_available
 
 
 _import_structure = {
@@ -40,7 +38,12 @@ _import_structure = {
     ],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_data2vec_audio"] = [
         "DATA2VEC_AUDIO_PRETRAINED_MODEL_ARCHIVE_LIST",
         "Data2VecAudioForAudioFrameClassification",
@@ -90,7 +93,12 @@ if TYPE_CHECKING:
         Data2VecVisionOnnxConfig,
     )
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_data2vec_audio import (
             DATA2VEC_AUDIO_PRETRAINED_MODEL_ARCHIVE_LIST,
             Data2VecAudioForAudioFrameClassification,

src/transformers/models/deberta/__init__.py

@@ -18,7 +18,13 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -26,10 +32,20 @@ _import_structure = {
     "tokenization_deberta": ["DebertaTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_deberta_fast"] = ["DebertaTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_deberta"] = [
         "DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
         "DebertaForMaskedLM",
@@ -40,7 +56,12 @@ if is_torch_available():
         "DebertaPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_deberta"] = [
         "TF_DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFDebertaForMaskedLM",
@@ -56,10 +77,20 @@ if TYPE_CHECKING:
     from .configuration_deberta import DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaConfig
     from .tokenization_deberta import DebertaTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_deberta_fast import DebertaTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_deberta import (
             DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
             DebertaForMaskedLM,
@@ -70,7 +101,12 @@ if TYPE_CHECKING:
             DebertaPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_deberta import (
             TF_DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFDebertaForMaskedLM,

src/transformers/models/deberta_v2/__init__.py

@@ -18,7 +18,13 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -26,10 +32,20 @@ _import_structure = {
     "tokenization_deberta_v2": ["DebertaV2Tokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_deberta_v2_fast"] = ["DebertaV2TokenizerFast"]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_deberta_v2"] = [
         "TF_DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFDebertaV2ForMaskedLM",
@@ -40,7 +56,12 @@ if is_tf_available():
         "TFDebertaV2PreTrainedModel",
     ]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_deberta_v2"] = [
         "DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST",
         "DebertaV2ForMaskedLM",
@@ -56,10 +77,20 @@ if TYPE_CHECKING:
     from .configuration_deberta_v2 import DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaV2Config
     from .tokenization_deberta_v2 import DebertaV2Tokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_deberta_v2_fast import DebertaV2TokenizerFast
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_deberta_v2 import (
             TF_DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFDebertaV2ForMaskedLM,
@@ -70,7 +101,12 @@ if TYPE_CHECKING:
             TFDebertaV2PreTrainedModel,
         )
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_deberta_v2 import (
             DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST,
             DebertaV2ForMaskedLM,

src/transformers/models/decision_transformer/__init__.py

@@ -18,7 +18,7 @@
 from typing import TYPE_CHECKING
 
 # rely on isort to merge the imports
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
@@ -28,7 +28,12 @@ _import_structure = {
     ],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_decision_transformer"] = [
         "DECISION_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "DecisionTransformerGPT2Model",
@@ -44,7 +49,12 @@ if TYPE_CHECKING:
         DecisionTransformerConfig,
     )
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_decision_transformer import (
             DECISION_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             DecisionTransformerGPT2Model,

src/transformers/models/deit/__init__.py

@@ -17,17 +17,27 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available, is_vision_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
 
 
 _import_structure = {
     "configuration_deit": ["DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "DeiTConfig", "DeiTOnnxConfig"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_deit"] = ["DeiTFeatureExtractor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_deit"] = [
         "DEIT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "DeiTForImageClassification",
@@ -41,10 +51,20 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_deit import DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP, DeiTConfig, DeiTOnnxConfig
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_deit import DeiTFeatureExtractor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_deit import (
             DEIT_PRETRAINED_MODEL_ARCHIVE_LIST,
             DeiTForImageClassification,

src/transformers/models/detr/__init__.py

@@ -18,17 +18,27 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_timm_available, is_vision_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_timm_available, is_vision_available
 
 
 _import_structure = {
     "configuration_detr": ["DETR_PRETRAINED_CONFIG_ARCHIVE_MAP", "DetrConfig"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_detr"] = ["DetrFeatureExtractor"]
 
-if is_timm_available():
+try:
+    if not is_timm_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_detr"] = [
         "DETR_PRETRAINED_MODEL_ARCHIVE_LIST",
         "DetrForObjectDetection",
@@ -41,10 +51,20 @@ if is_timm_available():
 if TYPE_CHECKING:
     from .configuration_detr import DETR_PRETRAINED_CONFIG_ARCHIVE_MAP, DetrConfig
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_detr import DetrFeatureExtractor
 
-    if is_timm_available():
+    try:
+        if not is_timm_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_detr import (
             DETR_PRETRAINED_MODEL_ARCHIVE_LIST,
             DetrForObjectDetection,

src/transformers/models/distilbert/__init__.py

@@ -18,7 +18,14 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -30,10 +37,20 @@ _import_structure = {
     "tokenization_distilbert": ["DistilBertTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_distilbert_fast"] = ["DistilBertTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_distilbert"] = [
         "DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "DistilBertForMaskedLM",
@@ -45,7 +62,12 @@ if is_torch_available():
         "DistilBertPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_distilbert"] = [
         "TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFDistilBertForMaskedLM",
@@ -58,7 +80,12 @@ if is_tf_available():
         "TFDistilBertPreTrainedModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_distilbert"] = [
         "FlaxDistilBertForMaskedLM",
         "FlaxDistilBertForMultipleChoice",
@@ -78,10 +105,20 @@ if TYPE_CHECKING:
     )
     from .tokenization_distilbert import DistilBertTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_distilbert_fast import DistilBertTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_distilbert import (
             DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             DistilBertForMaskedLM,
@@ -93,7 +130,12 @@ if TYPE_CHECKING:
             DistilBertPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_distilbert import (
             TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFDistilBertForMaskedLM,
@@ -106,7 +148,12 @@ if TYPE_CHECKING:
             TFDistilBertPreTrainedModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_distilbert import (
             FlaxDistilBertForMaskedLM,
             FlaxDistilBertForMultipleChoice,

src/transformers/models/dpr/__init__.py

@@ -18,7 +18,13 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -32,14 +38,24 @@ _import_structure = {
 }
 
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_dpr_fast"] = [
         "DPRContextEncoderTokenizerFast",
         "DPRQuestionEncoderTokenizerFast",
         "DPRReaderTokenizerFast",
     ]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_dpr"] = [
         "DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -53,7 +69,12 @@ if is_torch_available():
         "DPRReader",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_dpr"] = [
         "TF_DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TF_DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -76,14 +97,24 @@ if TYPE_CHECKING:
         DPRReaderTokenizer,
     )
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_dpr_fast import (
             DPRContextEncoderTokenizerFast,
             DPRQuestionEncoderTokenizerFast,
             DPRReaderTokenizerFast,
         )
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_dpr import (
             DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
             DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
@@ -97,7 +128,12 @@ if TYPE_CHECKING:
             DPRReader,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_dpr import (
             TF_DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
             TF_DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,

src/transformers/models/dpt/__init__.py

@@ -18,16 +18,27 @@
 from typing import TYPE_CHECKING
 
 from ...file_utils import _LazyModule, is_tokenizers_available, is_torch_available, is_vision_available
+from ...utils import OptionalDependencyNotAvailable
 
 
 _import_structure = {
     "configuration_dpt": ["DPT_PRETRAINED_CONFIG_ARCHIVE_MAP", "DPTConfig"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_dpt"] = ["DPTFeatureExtractor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_dpt"] = [
         "DPT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "DPTForDepthEstimation",
@@ -40,10 +51,20 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_dpt import DPT_PRETRAINED_CONFIG_ARCHIVE_MAP, DPTConfig
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_dpt import DPTFeatureExtractor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_dpt import (
             DPT_PRETRAINED_MODEL_ARCHIVE_LIST,
             DPTForDepthEstimation,

src/transformers/models/electra/__init__.py

@@ -18,7 +18,14 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -26,10 +33,20 @@ _import_structure = {
     "tokenization_electra": ["ElectraTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_electra_fast"] = ["ElectraTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_electra"] = [
         "ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST",
         "ElectraForCausalLM",
@@ -44,7 +61,12 @@ if is_torch_available():
         "load_tf_weights_in_electra",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_electra"] = [
         "TF_ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFElectraForMaskedLM",
@@ -57,7 +79,12 @@ if is_tf_available():
         "TFElectraPreTrainedModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_electra"] = [
         "FlaxElectraForCausalLM",
         "FlaxElectraForMaskedLM",
@@ -75,10 +102,20 @@ if TYPE_CHECKING:
     from .configuration_electra import ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP, ElectraConfig, ElectraOnnxConfig
     from .tokenization_electra import ElectraTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_electra_fast import ElectraTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_electra import (
             ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST,
             ElectraForCausalLM,
@@ -93,7 +130,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_electra,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_electra import (
             TF_ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFElectraForMaskedLM,
@@ -106,7 +148,12 @@ if TYPE_CHECKING:
             TFElectraPreTrainedModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_electra import (
             FlaxElectraForCausalLM,
             FlaxElectraForMaskedLM,

src/transformers/models/encoder_decoder/__init__.py

@@ -18,32 +18,68 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
     "configuration_encoder_decoder": ["EncoderDecoderConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_encoder_decoder"] = ["EncoderDecoderModel"]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_encoder_decoder"] = ["TFEncoderDecoderModel"]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_encoder_decoder"] = ["FlaxEncoderDecoderModel"]
 
 if TYPE_CHECKING:
     from .configuration_encoder_decoder import EncoderDecoderConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_encoder_decoder import EncoderDecoderModel
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_encoder_decoder import TFEncoderDecoderModel
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_encoder_decoder import FlaxEncoderDecoderModel
 
 else:

src/transformers/models/flaubert/__init__.py

@@ -18,7 +18,7 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tf_available, is_torch_available
 
 
 _import_structure = {
@@ -26,7 +26,12 @@ _import_structure = {
     "tokenization_flaubert": ["FlaubertTokenizer"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flaubert"] = [
         "FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "FlaubertForMultipleChoice",
@@ -38,7 +43,12 @@ if is_torch_available():
         "FlaubertWithLMHeadModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_flaubert"] = [
         "TF_FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFFlaubertForMultipleChoice",
@@ -55,7 +65,12 @@ if TYPE_CHECKING:
     from .configuration_flaubert import FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, FlaubertConfig, FlaubertOnnxConfig
     from .tokenization_flaubert import FlaubertTokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flaubert import (
             FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             FlaubertForMultipleChoice,
@@ -67,7 +82,12 @@ if TYPE_CHECKING:
             FlaubertWithLMHeadModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_flaubert import (
             TF_FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFFlaubertForMultipleChoice,

src/transformers/models/fnet/__init__.py

@@ -17,22 +17,41 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from transformers import is_sentencepiece_available
+from ...utils import (
-
+    OptionalDependencyNotAvailable,
-from ...utils import _LazyModule, is_tokenizers_available, is_torch_available
+    _LazyModule,
+    is_sentencepiece_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
     "configuration_fnet": ["FNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "FNetConfig"],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_fnet"] = ["FNetTokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_fnet_fast"] = ["FNetTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_fnet"] = [
         "FNET_PRETRAINED_MODEL_ARCHIVE_LIST",
         "FNetForMaskedLM",
@@ -51,13 +70,28 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_fnet import FNET_PRETRAINED_CONFIG_ARCHIVE_MAP, FNetConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_fnet import FNetTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_fnet_fast import FNetTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_fnet import (
             FNET_PRETRAINED_MODEL_ARCHIVE_LIST,
             FNetForMaskedLM,

src/transformers/models/fsmt/__init__.py

@@ -18,7 +18,7 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
@@ -26,7 +26,12 @@ _import_structure = {
     "tokenization_fsmt": ["FSMTTokenizer"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_fsmt"] = ["FSMTForConditionalGeneration", "FSMTModel", "PretrainedFSMTModel"]
 
 
@@ -34,7 +39,12 @@ if TYPE_CHECKING:
     from .configuration_fsmt import FSMT_PRETRAINED_CONFIG_ARCHIVE_MAP, FSMTConfig
     from .tokenization_fsmt import FSMTTokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_fsmt import FSMTForConditionalGeneration, FSMTModel, PretrainedFSMTModel
 
 else:

src/transformers/models/funnel/__init__.py

@@ -18,7 +18,13 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -27,10 +33,20 @@ _import_structure = {
     "tokenization_funnel": ["FunnelTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_funnel_fast"] = ["FunnelTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_funnel"] = [
         "FUNNEL_PRETRAINED_MODEL_ARCHIVE_LIST",
         "FunnelBaseModel",
@@ -45,7 +61,12 @@ if is_torch_available():
         "load_tf_weights_in_funnel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_funnel"] = [
         "TF_FUNNEL_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFFunnelBaseModel",
@@ -64,10 +85,20 @@ if TYPE_CHECKING:
     from .configuration_funnel import FUNNEL_PRETRAINED_CONFIG_ARCHIVE_MAP, FunnelConfig
     from .tokenization_funnel import FunnelTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_funnel_fast import FunnelTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_funnel import (
             FUNNEL_PRETRAINED_MODEL_ARCHIVE_LIST,
             FunnelBaseModel,
@@ -82,7 +113,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_funnel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_funnel import (
             TF_FUNNEL_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFFunnelBaseModel,

src/transformers/models/glpn/__init__.py

@@ -18,17 +18,27 @@
 from typing import TYPE_CHECKING
 
 # rely on isort to merge the imports
-from ...utils import _LazyModule, is_torch_available, is_vision_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
 
 
 _import_structure = {
     "configuration_glpn": ["GLPN_PRETRAINED_CONFIG_ARCHIVE_MAP", "GLPNConfig"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_glpn"] = ["GLPNFeatureExtractor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_glpn"] = [
         "GLPN_PRETRAINED_MODEL_ARCHIVE_LIST",
         "GLPNForDepthEstimation",
@@ -41,10 +51,20 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_glpn import GLPN_PRETRAINED_CONFIG_ARCHIVE_MAP, GLPNConfig
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_glpn import GLPNFeatureExtractor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_glpn import (
             GLPN_PRETRAINED_MODEL_ARCHIVE_LIST,
             GLPNForDepthEstimation,

src/transformers/models/gpt2/__init__.py

@@ -18,7 +18,14 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -26,10 +33,20 @@ _import_structure = {
     "tokenization_gpt2": ["GPT2Tokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_gpt2_fast"] = ["GPT2TokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_gpt2"] = [
         "GPT2_PRETRAINED_MODEL_ARCHIVE_LIST",
         "GPT2DoubleHeadsModel",
@@ -41,7 +58,12 @@ if is_torch_available():
         "load_tf_weights_in_gpt2",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_gpt2"] = [
         "TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFGPT2DoubleHeadsModel",
@@ -52,17 +74,32 @@ if is_tf_available():
         "TFGPT2PreTrainedModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_gpt2"] = ["FlaxGPT2LMHeadModel", "FlaxGPT2Model", "FlaxGPT2PreTrainedModel"]
 
 if TYPE_CHECKING:
     from .configuration_gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config, GPT2OnnxConfig
     from .tokenization_gpt2 import GPT2Tokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_gpt2_fast import GPT2TokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_gpt2 import (
             GPT2_PRETRAINED_MODEL_ARCHIVE_LIST,
             GPT2DoubleHeadsModel,
@@ -74,7 +111,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_gpt2,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_gpt2 import (
             TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFGPT2DoubleHeadsModel,
@@ -85,7 +127,12 @@ if TYPE_CHECKING:
             TFGPT2PreTrainedModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_gpt2 import FlaxGPT2LMHeadModel, FlaxGPT2Model, FlaxGPT2PreTrainedModel
 
 else:

src/transformers/models/gpt_neo/__init__.py

@@ -17,14 +17,19 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_flax_available, is_torch_available
 
 
 _import_structure = {
     "configuration_gpt_neo": ["GPT_NEO_PRETRAINED_CONFIG_ARCHIVE_MAP", "GPTNeoConfig", "GPTNeoOnnxConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_gpt_neo"] = [
         "GPT_NEO_PRETRAINED_MODEL_ARCHIVE_LIST",
         "GPTNeoForCausalLM",
@@ -34,7 +39,12 @@ if is_torch_available():
         "load_tf_weights_in_gpt_neo",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_gpt_neo"] = [
         "FlaxGPTNeoForCausalLM",
         "FlaxGPTNeoModel",
@@ -45,7 +55,12 @@ if is_flax_available():
 if TYPE_CHECKING:
     from .configuration_gpt_neo import GPT_NEO_PRETRAINED_CONFIG_ARCHIVE_MAP, GPTNeoConfig, GPTNeoOnnxConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_gpt_neo import (
             GPT_NEO_PRETRAINED_MODEL_ARCHIVE_LIST,
             GPTNeoForCausalLM,
@@ -55,7 +70,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_gpt_neo,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_gpt_neo import FlaxGPTNeoForCausalLM, FlaxGPTNeoModel, FlaxGPTNeoPreTrainedModel
 
 

src/transformers/models/gptj/__init__.py

@@ -17,14 +17,25 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
     "configuration_gptj": ["GPTJ_PRETRAINED_CONFIG_ARCHIVE_MAP", "GPTJConfig", "GPTJOnnxConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_gptj"] = [
         "GPTJ_PRETRAINED_MODEL_ARCHIVE_LIST",
         "GPTJForCausalLM",
@@ -34,7 +45,12 @@ if is_torch_available():
         "GPTJPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_gptj"] = [
         "TFGPTJForCausalLM",
         "TFGPTJForQuestionAnswering",
@@ -43,7 +59,12 @@ if is_tf_available():
         "TFGPTJPreTrainedModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_gptj"] = [
         "FlaxGPTJForCausalLM",
         "FlaxGPTJModel",
@@ -54,7 +75,12 @@ if is_flax_available():
 if TYPE_CHECKING:
     from .configuration_gptj import GPTJ_PRETRAINED_CONFIG_ARCHIVE_MAP, GPTJConfig, GPTJOnnxConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_gptj import (
             GPTJ_PRETRAINED_MODEL_ARCHIVE_LIST,
             GPTJForCausalLM,
@@ -64,7 +90,12 @@ if TYPE_CHECKING:
             GPTJPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_gptj import (
             TFGPTJForCausalLM,
             TFGPTJForQuestionAnswering,
@@ -73,7 +104,12 @@ if TYPE_CHECKING:
             TFGPTJPreTrainedModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_gptj import FlaxGPTJForCausalLM, FlaxGPTJModel, FlaxGPTJPreTrainedModel
 
 else:

src/transformers/models/herbert/__init__.py

@@ -18,21 +18,31 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tokenizers_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available
 
 
 _import_structure = {
     "tokenization_herbert": ["HerbertTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_herbert_fast"] = ["HerbertTokenizerFast"]
 
 
 if TYPE_CHECKING:
     from .tokenization_herbert import HerbertTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_herbert_fast import HerbertTokenizerFast
 
 else:

src/transformers/models/hubert/__init__.py

@@ -17,7 +17,7 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tf_available, is_torch_available
 
 
 _import_structure = {
@@ -25,7 +25,12 @@ _import_structure = {
     "configuration_hubert": ["HUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "HubertConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_hubert"] = [
         "HUBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "HubertForCTC",
@@ -35,7 +40,12 @@ if is_torch_available():
     ]
 
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_hubert"] = [
         "TF_HUBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFHubertForCTC",
@@ -47,7 +57,12 @@ if TYPE_CHECKING:
     from ..wav2vec2.feature_extraction_wav2vec2 import Wav2Vec2FeatureExtractor
     from .configuration_hubert import HUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, HubertConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_hubert import (
             HUBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             HubertForCTC,
@@ -56,7 +71,12 @@ if TYPE_CHECKING:
             HubertPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_hubert import (
             TF_HUBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFHubertForCTC,

src/transformers/models/ibert/__init__.py

@@ -18,14 +18,19 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
     "configuration_ibert": ["IBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "IBertConfig", "IBertOnnxConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_ibert"] = [
         "IBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "IBertForMaskedLM",
@@ -40,7 +45,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_ibert import IBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, IBertConfig, IBertOnnxConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_ibert import (
             IBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             IBertForMaskedLM,

src/transformers/models/imagegpt/__init__.py

@@ -18,17 +18,27 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available, is_vision_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
 
 
 _import_structure = {
     "configuration_imagegpt": ["IMAGEGPT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ImageGPTConfig"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_imagegpt"] = ["ImageGPTFeatureExtractor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_imagegpt"] = [
         "IMAGEGPT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "ImageGPTForCausalImageModeling",
@@ -42,10 +52,20 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_imagegpt import IMAGEGPT_PRETRAINED_CONFIG_ARCHIVE_MAP, ImageGPTConfig
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_imagegpt import ImageGPTFeatureExtractor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_imagegpt import (
             IMAGEGPT_PRETRAINED_MODEL_ARCHIVE_LIST,
             ImageGPTForCausalImageModeling,

src/transformers/models/layoutlm/__init__.py

@@ -18,7 +18,13 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 from .configuration_layoutlm import LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP, LayoutLMConfig
 from .tokenization_layoutlm import LayoutLMTokenizer
 
@@ -28,10 +34,20 @@ _import_structure = {
     "tokenization_layoutlm": ["LayoutLMTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_layoutlm_fast"] = ["LayoutLMTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_layoutlm"] = [
         "LAYOUTLM_PRETRAINED_MODEL_ARCHIVE_LIST",
         "LayoutLMForMaskedLM",
@@ -41,7 +57,12 @@ if is_torch_available():
         "LayoutLMPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_layoutlm"] = [
         "TF_LAYOUTLM_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFLayoutLMForMaskedLM",
@@ -57,10 +78,20 @@ if TYPE_CHECKING:
     from .configuration_layoutlm import LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP, LayoutLMConfig, LayoutLMOnnxConfig
     from .tokenization_layoutlm import LayoutLMTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_layoutlm_fast import LayoutLMTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_layoutlm import (
             LAYOUTLM_PRETRAINED_MODEL_ARCHIVE_LIST,
             LayoutLMForMaskedLM,
@@ -69,7 +100,12 @@ if TYPE_CHECKING:
             LayoutLMModel,
             LayoutLMPreTrainedModel,
         )
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_layoutlm import (
             TF_LAYOUTLM_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFLayoutLMForMaskedLM,

src/transformers/models/layoutlmv2/__init__.py

@@ -18,7 +18,13 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tokenizers_available, is_torch_available, is_vision_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tokenizers_available,
+    is_torch_available,
+    is_vision_available,
+)
 
 
 _import_structure = {
@@ -26,14 +32,29 @@ _import_structure = {
     "tokenization_layoutlmv2": ["LayoutLMv2Tokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_layoutlmv2_fast"] = ["LayoutLMv2TokenizerFast"]
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_layoutlmv2"] = ["LayoutLMv2FeatureExtractor"]
     _import_structure["processing_layoutlmv2"] = ["LayoutLMv2Processor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_layoutlmv2"] = [
         "LAYOUTLMV2_PRETRAINED_MODEL_ARCHIVE_LIST",
         "LayoutLMv2ForQuestionAnswering",
@@ -48,14 +69,29 @@ if TYPE_CHECKING:
     from .configuration_layoutlmv2 import LAYOUTLMV2_PRETRAINED_CONFIG_ARCHIVE_MAP, LayoutLMv2Config
     from .tokenization_layoutlmv2 import LayoutLMv2Tokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_layoutlmv2_fast import LayoutLMv2TokenizerFast
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_layoutlmv2 import LayoutLMv2FeatureExtractor
         from .processing_layoutlmv2 import LayoutLMv2Processor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_layoutlmv2 import (
             LAYOUTLMV2_PRETRAINED_MODEL_ARCHIVE_LIST,
             LayoutLMv2ForQuestionAnswering,

src/transformers/models/layoutxlm/__init__.py

@@ -19,6 +19,7 @@
 from typing import TYPE_CHECKING
 
 from ...utils import (
+    OptionalDependencyNotAvailable,
     _LazyModule,
     is_sentencepiece_available,
     is_tokenizers_available,
@@ -29,23 +30,53 @@ from ...utils import (
 
 _import_structure = {}
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_layoutxlm"] = ["LayoutXLMTokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_layoutxlm_fast"] = ["LayoutXLMTokenizerFast"]
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["processing_layoutxlm"] = ["LayoutXLMProcessor"]
 
 if TYPE_CHECKING:
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_layoutxlm import LayoutXLMTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_layoutxlm_fast import LayoutXLMTokenizerFast
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .processing_layoutlmv2 import LayoutXLMProcessor
 
 else:

src/transformers/models/led/__init__.py

@@ -17,7 +17,13 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -25,10 +31,20 @@ _import_structure = {
     "tokenization_led": ["LEDTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_led_fast"] = ["LEDTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_led"] = [
         "LED_PRETRAINED_MODEL_ARCHIVE_LIST",
         "LEDForConditionalGeneration",
@@ -39,7 +55,12 @@ if is_torch_available():
     ]
 
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_led"] = ["TFLEDForConditionalGeneration", "TFLEDModel", "TFLEDPreTrainedModel"]
 
 
@@ -47,10 +68,20 @@ if TYPE_CHECKING:
     from .configuration_led import LED_PRETRAINED_CONFIG_ARCHIVE_MAP, LEDConfig
     from .tokenization_led import LEDTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_led_fast import LEDTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_led import (
             LED_PRETRAINED_MODEL_ARCHIVE_LIST,
             LEDForConditionalGeneration,
@@ -60,7 +91,12 @@ if TYPE_CHECKING:
             LEDPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_led import TFLEDForConditionalGeneration, TFLEDModel, TFLEDPreTrainedModel
 
 else:

src/transformers/models/longformer/__init__.py

@@ -18,7 +18,13 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -30,10 +36,20 @@ _import_structure = {
     "tokenization_longformer": ["LongformerTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_longformer_fast"] = ["LongformerTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_longformer"] = [
         "LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "LongformerForMaskedLM",
@@ -46,7 +62,12 @@ if is_torch_available():
         "LongformerSelfAttention",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_longformer"] = [
         "TF_LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFLongformerForMaskedLM",
@@ -68,10 +89,20 @@ if TYPE_CHECKING:
     )
     from .tokenization_longformer import LongformerTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_longformer_fast import LongformerTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_longformer import (
             LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             LongformerForMaskedLM,
@@ -84,7 +115,12 @@ if TYPE_CHECKING:
             LongformerSelfAttention,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_longformer import (
             TF_LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFLongformerForMaskedLM,

src/transformers/models/luke/__init__.py

@@ -18,7 +18,7 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
@@ -26,7 +26,12 @@ _import_structure = {
     "tokenization_luke": ["LukeTokenizer"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_luke"] = [
         "LUKE_PRETRAINED_MODEL_ARCHIVE_LIST",
         "LukeForEntityClassification",
@@ -42,7 +47,12 @@ if TYPE_CHECKING:
     from .configuration_luke import LUKE_PRETRAINED_CONFIG_ARCHIVE_MAP, LukeConfig
     from .tokenization_luke import LukeTokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_luke import (
             LUKE_PRETRAINED_MODEL_ARCHIVE_LIST,
             LukeForEntityClassification,

src/transformers/models/lxmert/__init__.py

@@ -18,7 +18,13 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -26,10 +32,20 @@ _import_structure = {
     "tokenization_lxmert": ["LxmertTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_lxmert_fast"] = ["LxmertTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_lxmert"] = [
         "LxmertEncoder",
         "LxmertForPreTraining",
@@ -40,7 +56,12 @@ if is_torch_available():
         "LxmertXLayer",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_lxmert"] = [
         "TF_LXMERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFLxmertForPreTraining",
@@ -55,10 +76,20 @@ if TYPE_CHECKING:
     from .configuration_lxmert import LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP, LxmertConfig
     from .tokenization_lxmert import LxmertTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_lxmert_fast import LxmertTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_lxmert import (
             LxmertEncoder,
             LxmertForPreTraining,
@@ -69,7 +100,12 @@ if TYPE_CHECKING:
             LxmertXLayer,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_lxmert import (
             TF_LXMERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFLxmertForPreTraining,

src/transformers/models/m2m_100/__init__.py

@@ -17,7 +17,7 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tokenizers_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
 
 
 _import_structure = {
@@ -26,7 +26,12 @@ _import_structure = {
 }
 
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_m2m_100"] = [
         "M2M_100_PRETRAINED_MODEL_ARCHIVE_LIST",
         "M2M100ForConditionalGeneration",
@@ -39,7 +44,12 @@ if TYPE_CHECKING:
     from .configuration_m2m_100 import M2M_100_PRETRAINED_CONFIG_ARCHIVE_MAP, M2M100Config, M2M100OnnxConfig
     from .tokenization_m2m_100 import M2M100Tokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_m2m_100 import (
             M2M_100_PRETRAINED_MODEL_ARCHIVE_LIST,
             M2M100ForConditionalGeneration,

src/transformers/models/marian/__init__.py

@@ -18,6 +18,7 @@
 from typing import TYPE_CHECKING
 
 from ...utils import (
+    OptionalDependencyNotAvailable,
     _LazyModule,
     is_flax_available,
     is_sentencepiece_available,
@@ -31,10 +32,20 @@ _import_structure = {
     "configuration_marian": ["MARIAN_PRETRAINED_CONFIG_ARCHIVE_MAP", "MarianConfig", "MarianOnnxConfig"],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_marian"] = ["MarianTokenizer"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_marian"] = [
         "MARIAN_PRETRAINED_MODEL_ARCHIVE_LIST",
         "MarianForCausalLM",
@@ -43,19 +54,39 @@ if is_torch_available():
         "MarianPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_marian"] = ["TFMarianModel", "TFMarianMTModel", "TFMarianPreTrainedModel"]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_marian"] = ["FlaxMarianModel", "FlaxMarianMTModel", "FlaxMarianPreTrainedModel"]
 
 if TYPE_CHECKING:
     from .configuration_marian import MARIAN_PRETRAINED_CONFIG_ARCHIVE_MAP, MarianConfig, MarianOnnxConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_marian import MarianTokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_marian import (
             MARIAN_PRETRAINED_MODEL_ARCHIVE_LIST,
             MarianForCausalLM,
@@ -64,10 +95,20 @@ if TYPE_CHECKING:
             MarianPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_marian import TFMarianModel, TFMarianMTModel, TFMarianPreTrainedModel
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_marian import FlaxMarianModel, FlaxMarianMTModel, FlaxMarianPreTrainedModel
 
 else:

src/transformers/models/maskformer/__init__.py

@@ -17,18 +17,28 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available, is_vision_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
 
 
 _import_structure = {
     "configuration_maskformer": ["MASKFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "MaskFormerConfig"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_maskformer"] = ["MaskFormerFeatureExtractor"]
 
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_maskformer"] = [
         "MASKFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "MaskFormerForInstanceSegmentation",
@@ -39,9 +49,19 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_maskformer import MASKFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, MaskFormerConfig
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_maskformer import MaskFormerFeatureExtractor
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_maskformer import (
             MASKFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             MaskFormerForInstanceSegmentation,

src/transformers/models/mbart/__init__.py

@@ -18,6 +18,7 @@
 from typing import TYPE_CHECKING
 
 from ...utils import (
+    OptionalDependencyNotAvailable,
     _LazyModule,
     is_flax_available,
     is_sentencepiece_available,
@@ -31,13 +32,28 @@ _import_structure = {
     "configuration_mbart": ["MBART_PRETRAINED_CONFIG_ARCHIVE_MAP", "MBartConfig", "MBartOnnxConfig"],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_mbart"] = ["MBartTokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_mbart_fast"] = ["MBartTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_mbart"] = [
         "MBART_PRETRAINED_MODEL_ARCHIVE_LIST",
         "MBartForCausalLM",
@@ -48,14 +64,24 @@ if is_torch_available():
         "MBartPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_mbart"] = [
         "TFMBartForConditionalGeneration",
         "TFMBartModel",
         "TFMBartPreTrainedModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_mbart"] = [
         "FlaxMBartForConditionalGeneration",
         "FlaxMBartForQuestionAnswering",
@@ -68,13 +94,28 @@ if is_flax_available():
 if TYPE_CHECKING:
     from .configuration_mbart import MBART_PRETRAINED_CONFIG_ARCHIVE_MAP, MBartConfig, MBartOnnxConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_mbart import MBartTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_mbart_fast import MBartTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_mbart import (
             MBART_PRETRAINED_MODEL_ARCHIVE_LIST,
             MBartForCausalLM,
@@ -85,10 +126,20 @@ if TYPE_CHECKING:
             MBartPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_mbart import TFMBartForConditionalGeneration, TFMBartModel, TFMBartPreTrainedModel
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_mbart import (
             FlaxMBartForConditionalGeneration,
             FlaxMBartForQuestionAnswering,

src/transformers/models/mbart50/__init__.py

@@ -17,23 +17,43 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_sentencepiece_available, is_tokenizers_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_sentencepiece_available, is_tokenizers_available
 
 
 _import_structure = {}
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_mbart50"] = ["MBart50Tokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_mbart50_fast"] = ["MBart50TokenizerFast"]
 
 
 if TYPE_CHECKING:
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_mbart50 import MBart50Tokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_mbart50_fast import MBart50TokenizerFast
 
 else:

src/transformers/models/megatron_bert/__init__.py

@@ -17,14 +17,19 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
     "configuration_megatron_bert": ["MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "MegatronBertConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_megatron_bert"] = [
         "MEGATRON_BERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "MegatronBertForCausalLM",
@@ -42,7 +47,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_megatron_bert import MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, MegatronBertConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_megatron_bert import (
             MEGATRON_BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             MegatronBertForCausalLM,

src/transformers/models/mluke/__init__.py

@@ -18,17 +18,27 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_sentencepiece_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_sentencepiece_available
 
 
 _import_structure = {}
 
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_mluke"] = ["MLukeTokenizer"]
 
 if TYPE_CHECKING:
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_mluke import MLukeTokenizer
 
 

src/transformers/models/mmbt/__init__.py

@@ -18,21 +18,31 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
     "configuration_mmbt": ["MMBTConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_mmbt"] = ["MMBTForClassification", "MMBTModel", "ModalEmbeddings"]
 
 
 if TYPE_CHECKING:
     from .configuration_mmbt import MMBTConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_mmbt import MMBTForClassification, MMBTModel, ModalEmbeddings
 
 else:

src/transformers/models/mobilebert/__init__.py

@@ -18,7 +18,13 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -30,10 +36,20 @@ _import_structure = {
     "tokenization_mobilebert": ["MobileBertTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_mobilebert_fast"] = ["MobileBertTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_mobilebert"] = [
         "MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "MobileBertForMaskedLM",
@@ -49,7 +65,12 @@ if is_torch_available():
         "load_tf_weights_in_mobilebert",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_mobilebert"] = [
         "TF_MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFMobileBertForMaskedLM",
@@ -73,10 +94,20 @@ if TYPE_CHECKING:
     )
     from .tokenization_mobilebert import MobileBertTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_mobilebert_fast import MobileBertTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_mobilebert import (
             MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             MobileBertForMaskedLM,
@@ -92,7 +123,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_mobilebert,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_mobilebert import (
             TF_MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFMobileBertForMaskedLM,

src/transformers/models/mpnet/__init__.py

@@ -18,7 +18,14 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -26,10 +33,20 @@ _import_structure = {
     "tokenization_mpnet": ["MPNetTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_mpnet_fast"] = ["MPNetTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_mpnet"] = [
         "MPNET_PRETRAINED_MODEL_ARCHIVE_LIST",
         "MPNetForMaskedLM",
@@ -42,7 +59,12 @@ if is_torch_available():
         "MPNetPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_mpnet"] = [
         "TF_MPNET_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFMPNetEmbeddings",
@@ -61,10 +83,20 @@ if TYPE_CHECKING:
     from .configuration_mpnet import MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP, MPNetConfig
     from .tokenization_mpnet import MPNetTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_mpnet_fast import MPNetTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_mpnet import (
             MPNET_PRETRAINED_MODEL_ARCHIVE_LIST,
             MPNetForMaskedLM,
@@ -77,7 +109,12 @@ if TYPE_CHECKING:
             MPNetPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_mpnet import (
             TF_MPNET_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFMPNetEmbeddings,

src/transformers/models/mt5/__init__.py

@@ -19,6 +19,7 @@
 from typing import TYPE_CHECKING
 
 from ...utils import (
+    OptionalDependencyNotAvailable,
     _LazyModule,
     is_flax_available,
     is_sentencepiece_available,
@@ -46,26 +47,56 @@ _import_structure = {
     "configuration_mt5": ["MT5Config"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_mt5"] = ["MT5EncoderModel", "MT5ForConditionalGeneration", "MT5Model"]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_mt5"] = ["TFMT5EncoderModel", "TFMT5ForConditionalGeneration", "TFMT5Model"]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_mt5"] = ["FlaxMT5ForConditionalGeneration", "FlaxMT5Model"]
 
 
 if TYPE_CHECKING:
     from .configuration_mt5 import MT5Config
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_mt5 import MT5EncoderModel, MT5ForConditionalGeneration, MT5Model
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_mt5 import TFMT5EncoderModel, TFMT5ForConditionalGeneration, TFMT5Model
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_mt5 import FlaxMT5ForConditionalGeneration, FlaxMT5Model
 
 else:

src/transformers/models/nystromformer/__init__.py

@@ -18,14 +18,19 @@
 from typing import TYPE_CHECKING
 
 # rely on isort to merge the imports
-from ...utils import _LazyModule, is_tokenizers_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
 
 
 _import_structure = {
     "configuration_nystromformer": ["NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "NystromformerConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_nystromformer"] = [
         "NYSTROMFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "NystromformerForMaskedLM",
@@ -42,7 +47,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_nystromformer import NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, NystromformerConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_nystromformer import (
             NYSTROMFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             NystromformerForMaskedLM,

src/transformers/models/openai/__init__.py

@@ -18,7 +18,13 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -26,10 +32,20 @@ _import_structure = {
     "tokenization_openai": ["OpenAIGPTTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_openai_fast"] = ["OpenAIGPTTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_openai"] = [
         "OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "OpenAIGPTDoubleHeadsModel",
@@ -40,7 +56,12 @@ if is_torch_available():
         "load_tf_weights_in_openai_gpt",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_openai"] = [
         "TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFOpenAIGPTDoubleHeadsModel",
@@ -56,10 +77,20 @@ if TYPE_CHECKING:
     from .configuration_openai import OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP, OpenAIGPTConfig
     from .tokenization_openai import OpenAIGPTTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_openai_fast import OpenAIGPTTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_openai import (
             OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST,
             OpenAIGPTDoubleHeadsModel,
@@ -70,7 +101,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_openai_gpt,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_openai import (
             TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFOpenAIGPTDoubleHeadsModel,

src/transformers/models/pegasus/__init__.py

@@ -18,6 +18,7 @@
 from typing import TYPE_CHECKING
 
 from ...utils import (
+    OptionalDependencyNotAvailable,
     _LazyModule,
     is_flax_available,
     is_sentencepiece_available,
@@ -31,13 +32,28 @@ _import_structure = {
     "configuration_pegasus": ["PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP", "PegasusConfig"],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_pegasus"] = ["PegasusTokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_pegasus_fast"] = ["PegasusTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_pegasus"] = [
         "PEGASUS_PRETRAINED_MODEL_ARCHIVE_LIST",
         "PegasusForCausalLM",
@@ -46,14 +62,24 @@ if is_torch_available():
         "PegasusPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_pegasus"] = [
         "TFPegasusForConditionalGeneration",
         "TFPegasusModel",
         "TFPegasusPreTrainedModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_pegasus"] = [
         "FlaxPegasusForConditionalGeneration",
         "FlaxPegasusModel",
@@ -64,13 +90,28 @@ if is_flax_available():
 if TYPE_CHECKING:
     from .configuration_pegasus import PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP, PegasusConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_pegasus import PegasusTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_pegasus_fast import PegasusTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_pegasus import (
             PEGASUS_PRETRAINED_MODEL_ARCHIVE_LIST,
             PegasusForCausalLM,
@@ -79,10 +120,20 @@ if TYPE_CHECKING:
             PegasusPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_pegasus import TFPegasusForConditionalGeneration, TFPegasusModel, TFPegasusPreTrainedModel
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_pegasus import (
             FlaxPegasusForConditionalGeneration,
             FlaxPegasusModel,

src/transformers/models/perceiver/__init__.py

@@ -17,7 +17,13 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tokenizers_available, is_torch_available, is_vision_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tokenizers_available,
+    is_torch_available,
+    is_vision_available,
+)
 
 
 _import_structure = {
@@ -25,10 +31,20 @@ _import_structure = {
     "tokenization_perceiver": ["PerceiverTokenizer"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_perceiver"] = ["PerceiverFeatureExtractor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_perceiver"] = [
         "PERCEIVER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "PerceiverForImageClassificationConvProcessing",
@@ -48,10 +64,20 @@ if TYPE_CHECKING:
     from .configuration_perceiver import PERCEIVER_PRETRAINED_CONFIG_ARCHIVE_MAP, PerceiverConfig
     from .tokenization_perceiver import PerceiverTokenizer
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_perceiver import PerceiverFeatureExtractor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_perceiver import (
             PERCEIVER_PRETRAINED_MODEL_ARCHIVE_LIST,
             PerceiverForImageClassificationConvProcessing,

src/transformers/models/plbart/__init__.py

@@ -17,17 +17,33 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_sentencepiece_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_sentencepiece_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
     "configuration_plbart": ["PLBART_PRETRAINED_CONFIG_ARCHIVE_MAP", "PLBartConfig"],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_plbart"] = ["PLBartTokenizer"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_plbart"] = [
         "PLBART_PRETRAINED_MODEL_ARCHIVE_LIST",
         "PLBartForCausalLM",
@@ -41,10 +57,20 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_plbart import PLBART_PRETRAINED_CONFIG_ARCHIVE_MAP, PLBartConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_plbart import PLBartTokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_plbart import (
             PLBART_PRETRAINED_MODEL_ARCHIVE_LIST,
             PLBartForCausalLM,

src/transformers/models/poolformer/__init__.py

@@ -18,17 +18,27 @@
 from typing import TYPE_CHECKING
 
 # rely on isort to merge the imports
-from ...utils import _LazyModule, is_torch_available, is_vision_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
 
 
 _import_structure = {
     "configuration_poolformer": ["POOLFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "PoolFormerConfig"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_poolformer"] = ["PoolFormerFeatureExtractor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_poolformer"] = [
         "POOLFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "PoolFormerForImageClassification",
@@ -40,10 +50,20 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_poolformer import POOLFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, PoolFormerConfig
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_poolformer import PoolFormerFeatureExtractor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_poolformer import (
             POOLFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             PoolFormerForImageClassification,

src/transformers/models/prophetnet/__init__.py

@@ -18,7 +18,7 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
@@ -26,7 +26,12 @@ _import_structure = {
     "tokenization_prophetnet": ["ProphetNetTokenizer"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_prophetnet"] = [
         "PROPHETNET_PRETRAINED_MODEL_ARCHIVE_LIST",
         "ProphetNetDecoder",
@@ -42,7 +47,12 @@ if TYPE_CHECKING:
     from .configuration_prophetnet import PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP, ProphetNetConfig
     from .tokenization_prophetnet import ProphetNetTokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_prophetnet import (
             PROPHETNET_PRETRAINED_MODEL_ARCHIVE_LIST,
             ProphetNetDecoder,

src/transformers/models/qdqbert/__init__.py

@@ -17,14 +17,19 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
     "configuration_qdqbert": ["QDQBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "QDQBertConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_qdqbert"] = [
         "QDQBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "QDQBertForMaskedLM",
@@ -44,7 +49,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_qdqbert import QDQBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, QDQBertConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_qdqbert import (
             QDQBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             QDQBertForMaskedLM,

src/transformers/models/rag/__init__.py

@@ -18,7 +18,7 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tf_available, is_torch_available
 
 
 _import_structure = {
@@ -27,7 +27,12 @@ _import_structure = {
     "tokenization_rag": ["RagTokenizer"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_rag"] = [
         "RagModel",
         "RagPreTrainedModel",
@@ -35,7 +40,12 @@ if is_torch_available():
         "RagTokenForGeneration",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_rag"] = [
         "TFRagModel",
         "TFRagPreTrainedModel",
@@ -49,10 +59,20 @@ if TYPE_CHECKING:
     from .retrieval_rag import RagRetriever
     from .tokenization_rag import RagTokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_rag import RagModel, RagPreTrainedModel, RagSequenceForGeneration, RagTokenForGeneration
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_rag import (
             TFRagModel,
             TFRagPreTrainedModel,

src/transformers/models/realm/__init__.py

@@ -17,7 +17,7 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tokenizers_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
 
 
 _import_structure = {
@@ -25,10 +25,20 @@ _import_structure = {
     "tokenization_realm": ["RealmTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_realm_fast"] = ["RealmTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_realm"] = [
         "REALM_PRETRAINED_MODEL_ARCHIVE_LIST",
         "RealmEmbedder",
@@ -46,10 +56,20 @@ if TYPE_CHECKING:
     from .configuration_realm import REALM_PRETRAINED_CONFIG_ARCHIVE_MAP, RealmConfig
     from .tokenization_realm import RealmTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_realm import RealmTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_realm import (
             REALM_PRETRAINED_MODEL_ARCHIVE_LIST,
             RealmEmbedder,

src/transformers/models/reformer/__init__.py

@@ -18,20 +18,41 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_sentencepiece_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_sentencepiece_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
     "configuration_reformer": ["REFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "ReformerConfig"],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_reformer"] = ["ReformerTokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_reformer_fast"] = ["ReformerTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_reformer"] = [
         "REFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "ReformerAttention",
@@ -48,13 +69,28 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_reformer import REFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, ReformerConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_reformer import ReformerTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_reformer_fast import ReformerTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_reformer import (
             REFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             ReformerAttention,

src/transformers/models/regnet/__init__.py

@@ -19,13 +19,19 @@ from typing import TYPE_CHECKING
 
 # rely on isort to merge the imports
 from ...file_utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable
 
 
 _import_structure = {
     "configuration_regnet": ["REGNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "RegNetConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_regnet"] = [
         "REGNET_PRETRAINED_MODEL_ARCHIVE_LIST",
         "RegNetForImageClassification",
@@ -37,7 +43,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_regnet import REGNET_PRETRAINED_CONFIG_ARCHIVE_MAP, RegNetConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_regnet import (
             REGNET_PRETRAINED_MODEL_ARCHIVE_LIST,
             RegNetForImageClassification,

src/transformers/models/rembert/__init__.py

@@ -19,6 +19,7 @@
 from typing import TYPE_CHECKING
 
 from ...utils import (
+    OptionalDependencyNotAvailable,
     _LazyModule,
     is_sentencepiece_available,
     is_tf_available,
@@ -31,13 +32,28 @@ _import_structure = {
     "configuration_rembert": ["REMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "RemBertConfig"],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_rembert"] = ["RemBertTokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_rembert_fast"] = ["RemBertTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_rembert"] = [
         "REMBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "RemBertForCausalLM",
@@ -53,7 +69,12 @@ if is_torch_available():
     ]
 
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_rembert"] = [
         "TF_REMBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFRemBertForCausalLM",
@@ -71,13 +92,28 @@ if is_tf_available():
 if TYPE_CHECKING:
     from .configuration_rembert import REMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, RemBertConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_rembert import RemBertTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_rembert_fast import RemBertTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_rembert import (
             REMBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             RemBertForCausalLM,
@@ -92,7 +128,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_rembert,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_rembert import (
             TF_REMBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFRemBertForCausalLM,

src/transformers/models/resnet/__init__.py

@@ -18,14 +18,19 @@
 from typing import TYPE_CHECKING
 
 # rely on isort to merge the imports
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
     "configuration_resnet": ["RESNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "ResNetConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_resnet"] = [
         "RESNET_PRETRAINED_MODEL_ARCHIVE_LIST",
         "ResNetForImageClassification",
@@ -37,7 +42,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_resnet import RESNET_PRETRAINED_CONFIG_ARCHIVE_MAP, ResNetConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_resnet import (
             RESNET_PRETRAINED_MODEL_ARCHIVE_LIST,
             ResNetForImageClassification,

src/transformers/models/retribert/__init__.py

@@ -18,7 +18,7 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tokenizers_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
 
 
 _import_structure = {
@@ -26,10 +26,20 @@ _import_structure = {
     "tokenization_retribert": ["RetriBertTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_retribert_fast"] = ["RetriBertTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_retribert"] = [
         "RETRIBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "RetriBertModel",
@@ -41,10 +51,20 @@ if TYPE_CHECKING:
     from .configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, RetriBertConfig
     from .tokenization_retribert import RetriBertTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_retribert_fast import RetriBertTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_retribert import (
             RETRIBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             RetriBertModel,

src/transformers/models/roberta/__init__.py

@@ -18,7 +18,14 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -26,10 +33,20 @@ _import_structure = {
     "tokenization_roberta": ["RobertaTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_roberta_fast"] = ["RobertaTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_roberta"] = [
         "ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
         "RobertaForCausalLM",
@@ -42,7 +59,12 @@ if is_torch_available():
         "RobertaPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_roberta"] = [
         "TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFRobertaForCausalLM",
@@ -56,7 +78,12 @@ if is_tf_available():
         "TFRobertaPreTrainedModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_roberta"] = [
         "FlaxRobertaForCausalLM",
         "FlaxRobertaForMaskedLM",
@@ -73,10 +100,20 @@ if TYPE_CHECKING:
     from .configuration_roberta import ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, RobertaConfig, RobertaOnnxConfig
     from .tokenization_roberta import RobertaTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_roberta_fast import RobertaTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_roberta import (
             ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
             RobertaForCausalLM,
@@ -89,7 +126,12 @@ if TYPE_CHECKING:
             RobertaPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_roberta import (
             TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFRobertaForCausalLM,
@@ -103,7 +145,12 @@ if TYPE_CHECKING:
             TFRobertaPreTrainedModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_roberta import (
             FlaxRobertaForCausalLM,
             FlaxRobertaForMaskedLM,

src/transformers/models/roformer/__init__.py

@@ -17,7 +17,14 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_tokenizers_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -25,10 +32,20 @@ _import_structure = {
     "tokenization_roformer": ["RoFormerTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_roformer_fast"] = ["RoFormerTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_roformer"] = [
         "ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "RoFormerForCausalLM",
@@ -44,7 +61,12 @@ if is_torch_available():
     ]
 
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_roformer"] = [
         "TF_ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFRoFormerForCausalLM",
@@ -59,7 +81,12 @@ if is_tf_available():
     ]
 
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_roformer"] = [
         "FLAX_ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "FlaxRoFormerForMaskedLM",
@@ -76,10 +103,20 @@ if TYPE_CHECKING:
     from .configuration_roformer import ROFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, RoFormerConfig, RoFormerOnnxConfig
     from .tokenization_roformer import RoFormerTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_roformer_fast import RoFormerTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_roformer import (
             ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             RoFormerForCausalLM,
@@ -94,7 +131,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_roformer,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_roformer import (
             TF_ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFRoFormerForCausalLM,
@@ -108,7 +150,12 @@ if TYPE_CHECKING:
             TFRoFormerPreTrainedModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_roformer import (
             FLAX_ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             FlaxRoFormerForMaskedLM,

src/transformers/models/segformer/__init__.py

@@ -17,17 +17,27 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available, is_vision_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
 
 
 _import_structure = {
     "configuration_segformer": ["SEGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "SegformerConfig"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_segformer"] = ["SegformerFeatureExtractor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_segformer"] = [
         "SEGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "SegformerDecodeHead",
@@ -42,10 +52,20 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_segformer import SEGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, SegformerConfig
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_segformer import SegformerFeatureExtractor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_segformer import (
             SEGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             SegformerDecodeHead,

src/transformers/models/sew/__init__.py

@@ -17,14 +17,19 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
     "configuration_sew": ["SEW_PRETRAINED_CONFIG_ARCHIVE_MAP", "SEWConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_sew"] = [
         "SEW_PRETRAINED_MODEL_ARCHIVE_LIST",
         "SEWForCTC",
@@ -36,7 +41,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_sew import SEW_PRETRAINED_CONFIG_ARCHIVE_MAP, SEWConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_sew import (
             SEW_PRETRAINED_MODEL_ARCHIVE_LIST,
             SEWForCTC,

src/transformers/models/sew_d/__init__.py

@@ -17,14 +17,19 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
     "configuration_sew_d": ["SEW_D_PRETRAINED_CONFIG_ARCHIVE_MAP", "SEWDConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_sew_d"] = [
         "SEW_D_PRETRAINED_MODEL_ARCHIVE_LIST",
         "SEWDForCTC",
@@ -36,7 +41,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_sew_d import SEW_D_PRETRAINED_CONFIG_ARCHIVE_MAP, SEWDConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_sew_d import (
             SEW_D_PRETRAINED_MODEL_ARCHIVE_LIST,
             SEWDForCTC,

src/transformers/models/speech_encoder_decoder/__init__.py

@@ -18,26 +18,46 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_flax_available, is_torch_available
 
 
 _import_structure = {
     "configuration_speech_encoder_decoder": ["SpeechEncoderDecoderConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_speech_encoder_decoder"] = ["SpeechEncoderDecoderModel"]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_speech_encoder_decoder"] = ["FlaxSpeechEncoderDecoderModel"]
 
 if TYPE_CHECKING:
     from .configuration_speech_encoder_decoder import SpeechEncoderDecoderConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_speech_encoder_decoder import SpeechEncoderDecoderModel
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_speech_encoder_decoder import FlaxSpeechEncoderDecoderModel
 
 else:

src/transformers/models/speech_to_text/__init__.py

@@ -17,7 +17,14 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_sentencepiece_available, is_speech_available, is_tf_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_sentencepiece_available,
+    is_speech_available,
+    is_tf_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -27,16 +34,31 @@ _import_structure = {
     ],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_speech_to_text"] = ["Speech2TextTokenizer"]
 
-if is_speech_available():
+try:
+    if not is_speech_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_speech_to_text"] = ["Speech2TextFeatureExtractor"]
 
     if is_sentencepiece_available():
         _import_structure["processing_speech_to_text"] = ["Speech2TextProcessor"]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_speech_to_text"] = [
         "TF_SPEECH_TO_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFSpeech2TextForConditionalGeneration",
@@ -44,7 +66,12 @@ if is_tf_available():
         "TFSpeech2TextPreTrainedModel",
     ]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_speech_to_text"] = [
         "SPEECH_TO_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "Speech2TextForConditionalGeneration",
@@ -56,16 +83,31 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_speech_to_text import SPEECH_TO_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP, Speech2TextConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_speech_to_text import Speech2TextTokenizer
 
-    if is_speech_available():
+    try:
+        if not is_speech_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_speech_to_text import Speech2TextFeatureExtractor
 
         if is_sentencepiece_available():
             from .processing_speech_to_text import Speech2TextProcessor
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_speech_to_text import (
             TF_SPEECH_TO_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFSpeech2TextForConditionalGeneration,
@@ -73,7 +115,12 @@ if TYPE_CHECKING:
             TFSpeech2TextPreTrainedModel,
         )
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_speech_to_text import (
             SPEECH_TO_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST,
             Speech2TextForConditionalGeneration,

src/transformers/models/speech_to_text_2/__init__.py

@@ -17,7 +17,13 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_sentencepiece_available, is_speech_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_sentencepiece_available,
+    is_speech_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -30,7 +36,12 @@ _import_structure = {
 }
 
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_speech_to_text_2"] = [
         "SPEECH_TO_TEXT_2_PRETRAINED_MODEL_ARCHIVE_LIST",
         "Speech2Text2ForCausalLM",
@@ -43,7 +54,12 @@ if TYPE_CHECKING:
     from .processing_speech_to_text_2 import Speech2Text2Processor
     from .tokenization_speech_to_text_2 import Speech2Text2Tokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_speech_to_text_2 import (
             SPEECH_TO_TEXT_2_PRETRAINED_MODEL_ARCHIVE_LIST,
             Speech2Text2ForCausalLM,

src/transformers/models/splinter/__init__.py

@@ -17,7 +17,7 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tokenizers_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
 
 
 _import_structure = {
@@ -25,10 +25,20 @@ _import_structure = {
     "tokenization_splinter": ["SplinterTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_splinter_fast"] = ["SplinterTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_splinter"] = [
         "SPLINTER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "SplinterForQuestionAnswering",
@@ -42,10 +52,20 @@ if TYPE_CHECKING:
     from .configuration_splinter import SPLINTER_PRETRAINED_CONFIG_ARCHIVE_MAP, SplinterConfig
     from .tokenization_splinter import SplinterTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_splinter_fast import SplinterTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_splinter import (
             SPLINTER_PRETRAINED_MODEL_ARCHIVE_LIST,
             SplinterForQuestionAnswering,

src/transformers/models/squeezebert/__init__.py

@@ -18,7 +18,7 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tokenizers_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
 
 
 _import_structure = {
@@ -26,10 +26,20 @@ _import_structure = {
     "tokenization_squeezebert": ["SqueezeBertTokenizer"],
 }
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_squeezebert_fast"] = ["SqueezeBertTokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_squeezebert"] = [
         "SQUEEZEBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "SqueezeBertForMaskedLM",
@@ -47,10 +57,20 @@ if TYPE_CHECKING:
     from .configuration_squeezebert import SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, SqueezeBertConfig
     from .tokenization_squeezebert import SqueezeBertTokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_squeezebert_fast import SqueezeBertTokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_squeezebert import (
             SQUEEZEBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             SqueezeBertForMaskedLM,

src/transformers/models/swin/__init__.py

@@ -18,7 +18,7 @@
 from typing import TYPE_CHECKING
 
 # rely on isort to merge the imports
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
@@ -26,7 +26,12 @@ _import_structure = {
 }
 
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_swin"] = [
         "SWIN_PRETRAINED_MODEL_ARCHIVE_LIST",
         "SwinForImageClassification",
@@ -39,7 +44,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_swin import SWIN_PRETRAINED_CONFIG_ARCHIVE_MAP, SwinConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_swin import (
             SWIN_PRETRAINED_MODEL_ARCHIVE_LIST,
             SwinForImageClassification,

src/transformers/models/t5/__init__.py

@@ -19,6 +19,7 @@
 from typing import TYPE_CHECKING
 
 from ...utils import (
+    OptionalDependencyNotAvailable,
     _LazyModule,
     is_flax_available,
     is_sentencepiece_available,
@@ -32,13 +33,28 @@ _import_structure = {
     "configuration_t5": ["T5_PRETRAINED_CONFIG_ARCHIVE_MAP", "T5Config", "T5OnnxConfig"],
 }
 
-if is_sentencepiece_available():
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_t5"] = ["T5Tokenizer"]
 
-if is_tokenizers_available():
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["tokenization_t5_fast"] = ["T5TokenizerFast"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_t5"] = [
         "T5_PRETRAINED_MODEL_ARCHIVE_LIST",
         "T5EncoderModel",
@@ -48,7 +64,12 @@ if is_torch_available():
         "load_tf_weights_in_t5",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_t5"] = [
         "TF_T5_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFT5EncoderModel",
@@ -57,7 +78,12 @@ if is_tf_available():
         "TFT5PreTrainedModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_t5"] = [
         "FlaxT5ForConditionalGeneration",
         "FlaxT5Model",
@@ -68,13 +94,28 @@ if is_flax_available():
 if TYPE_CHECKING:
     from .configuration_t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config, T5OnnxConfig
 
-    if is_sentencepiece_available():
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_t5 import T5Tokenizer
 
-    if is_tokenizers_available():
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .tokenization_t5_fast import T5TokenizerFast
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_t5 import (
             T5_PRETRAINED_MODEL_ARCHIVE_LIST,
             T5EncoderModel,
@@ -84,7 +125,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_t5,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_t5 import (
             TF_T5_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFT5EncoderModel,
@@ -93,7 +139,12 @@ if TYPE_CHECKING:
             TFT5PreTrainedModel,
         )
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_t5 import FlaxT5ForConditionalGeneration, FlaxT5Model, FlaxT5PreTrainedModel
 
 

src/transformers/models/tapas/__init__.py

@@ -18,7 +18,7 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tf_available, is_torch_available
 
 
 _import_structure = {
@@ -26,7 +26,12 @@ _import_structure = {
     "tokenization_tapas": ["TapasTokenizer"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tapas"] = [
         "TAPAS_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TapasForMaskedLM",
@@ -36,7 +41,12 @@ if is_torch_available():
         "TapasPreTrainedModel",
         "load_tf_weights_in_tapas",
     ]
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_tapas"] = [
         "TF_TAPAS_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFTapasForMaskedLM",
@@ -51,7 +61,12 @@ if TYPE_CHECKING:
     from .configuration_tapas import TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP, TapasConfig
     from .tokenization_tapas import TapasTokenizer
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tapas import (
             TAPAS_PRETRAINED_MODEL_ARCHIVE_LIST,
             TapasForMaskedLM,
@@ -62,7 +77,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_tapas,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_tapas import (
             TF_TAPAS_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFTapasForMaskedLM,

src/transformers/models/transfo_xl/__init__.py

@@ -18,7 +18,7 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_tf_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tf_available, is_torch_available
 
 
 _import_structure = {
@@ -26,7 +26,12 @@ _import_structure = {
     "tokenization_transfo_xl": ["TransfoXLCorpus", "TransfoXLTokenizer"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_transfo_xl"] = [
         "TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST",
         "AdaptiveEmbedding",
@@ -37,7 +42,12 @@ if is_torch_available():
         "load_tf_weights_in_transfo_xl",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_transfo_xl"] = [
         "TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TFAdaptiveEmbedding",
@@ -53,7 +63,12 @@ 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():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_transfo_xl import (
             TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST,
             AdaptiveEmbedding,
@@ -64,7 +79,12 @@ if TYPE_CHECKING:
             load_tf_weights_in_transfo_xl,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_transfo_xl import (
             TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFAdaptiveEmbedding,

src/transformers/models/trocr/__init__.py

@@ -17,7 +17,13 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_sentencepiece_available, is_speech_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_sentencepiece_available,
+    is_speech_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -29,7 +35,12 @@ _import_structure = {
 }
 
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_trocr"] = [
         "TROCR_PRETRAINED_MODEL_ARCHIVE_LIST",
         "TrOCRForCausalLM",
@@ -41,7 +52,12 @@ if TYPE_CHECKING:
     from .configuration_trocr import TROCR_PRETRAINED_CONFIG_ARCHIVE_MAP, TrOCRConfig
     from .processing_trocr import TrOCRProcessor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_trocr import TROCR_PRETRAINED_MODEL_ARCHIVE_LIST, TrOCRForCausalLM, TrOCRPreTrainedModel
 
 else:

src/transformers/models/unispeech/__init__.py

@@ -17,14 +17,25 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
     "configuration_unispeech": ["UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP", "UniSpeechConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_unispeech"] = [
         "UNISPEECH_PRETRAINED_MODEL_ARCHIVE_LIST",
         "UniSpeechForCTC",
@@ -37,7 +48,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_unispeech import UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP, UniSpeechConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_unispeech import (
             UNISPEECH_PRETRAINED_MODEL_ARCHIVE_LIST,
             UniSpeechForCTC,

src/transformers/models/unispeech_sat/__init__.py

@@ -17,14 +17,25 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
     "configuration_unispeech_sat": ["UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP", "UniSpeechSatConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_unispeech_sat"] = [
         "UNISPEECH_SAT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "UniSpeechSatForAudioFrameClassification",
@@ -39,7 +50,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_unispeech_sat import UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP, UniSpeechSatConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_unispeech_sat import (
             UNISPEECH_SAT_PRETRAINED_MODEL_ARCHIVE_LIST,
             UniSpeechSatForAudioFrameClassification,

src/transformers/models/van/__init__.py

@@ -18,7 +18,7 @@
 from typing import TYPE_CHECKING
 
 # rely on isort to merge the imports
-from ...utils import _LazyModule, is_torch_available, is_vision_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
 
 
 _import_structure = {
@@ -26,7 +26,12 @@ _import_structure = {
 }
 
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_van"] = [
         "VAN_PRETRAINED_MODEL_ARCHIVE_LIST",
         "VanForImageClassification",
@@ -37,7 +42,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_van import VAN_PRETRAINED_CONFIG_ARCHIVE_MAP, VanConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_van import (
             VAN_PRETRAINED_MODEL_ARCHIVE_LIST,
             VanForImageClassification,

src/transformers/models/vilt/__init__.py

@@ -18,18 +18,28 @@
 from typing import TYPE_CHECKING
 
 # rely on isort to merge the imports
-from ...utils import _LazyModule, is_torch_available, is_vision_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
 
 
 _import_structure = {
     "configuration_vilt": ["VILT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ViltConfig"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_vilt"] = ["ViltFeatureExtractor"]
     _import_structure["processing_vilt"] = ["ViltProcessor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_vilt"] = [
         "VILT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "ViltForImageAndTextRetrieval",
@@ -45,11 +55,21 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_vilt import VILT_PRETRAINED_CONFIG_ARCHIVE_MAP, ViltConfig
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_vilt import ViltFeatureExtractor
         from .processing_vilt import ViltProcessor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_vilt import (
             VILT_PRETRAINED_MODEL_ARCHIVE_LIST,
             ViltForImageAndTextRetrieval,

src/transformers/models/vision_encoder_decoder/__init__.py

@@ -18,32 +18,68 @@
 
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
     "configuration_vision_encoder_decoder": ["VisionEncoderDecoderConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_vision_encoder_decoder"] = ["VisionEncoderDecoderModel"]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_vision_encoder_decoder"] = ["TFVisionEncoderDecoderModel"]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_vision_encoder_decoder"] = ["FlaxVisionEncoderDecoderModel"]
 
 if TYPE_CHECKING:
     from .configuration_vision_encoder_decoder import VisionEncoderDecoderConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_vision_encoder_decoder import VisionEncoderDecoderModel
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_vision_encoder_decoder import TFVisionEncoderDecoderModel
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_vision_encoder_decoder import FlaxVisionEncoderDecoderModel
 
 else:

src/transformers/models/vision_text_dual_encoder/__init__.py

@@ -18,7 +18,7 @@
 from typing import TYPE_CHECKING
 
 # rely on isort to merge the imports
-from ...utils import _LazyModule, is_flax_available, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_flax_available, is_torch_available
 
 
 _import_structure = {
@@ -27,11 +27,21 @@ _import_structure = {
 }
 
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_vision_text_dual_encoder"] = ["VisionTextDualEncoderModel"]
 
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_vision_text_dual_encoder"] = ["FlaxVisionTextDualEncoderModel"]
 
 
@@ -39,10 +49,20 @@ if TYPE_CHECKING:
     from .configuration_vision_text_dual_encoder import VisionTextDualEncoderConfig
     from .processing_visiotn_text_dual_encoder import VisionTextDualEncoderProcessor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_vision_text_dual_encoder import VisionTextDualEncoderModel
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_vision_text_dual_encoder import FlaxVisionTextDualEncoderModel
 
 

src/transformers/models/visual_bert/__init__.py

@@ -17,14 +17,19 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
 
 
 _import_structure = {
     "configuration_visual_bert": ["VISUAL_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "VisualBertConfig"],
 }
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_visual_bert"] = [
         "VISUAL_BERT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "VisualBertForMultipleChoice",
@@ -41,7 +46,12 @@ if is_torch_available():
 if TYPE_CHECKING:
     from .configuration_visual_bert import VISUAL_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, VisualBertConfig
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_visual_bert import (
             VISUAL_BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             VisualBertForMultipleChoice,

src/transformers/models/vit/__init__.py

@@ -17,17 +17,34 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import _LazyModule, is_flax_available, is_tf_available, is_torch_available, is_vision_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_torch_available,
+    is_vision_available,
+)
 
 
 _import_structure = {
     "configuration_vit": ["VIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ViTConfig", "ViTOnnxConfig"],
 }
 
-if is_vision_available():
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["feature_extraction_vit"] = ["ViTFeatureExtractor"]
 
-if is_torch_available():
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_vit"] = [
         "VIT_PRETRAINED_MODEL_ARCHIVE_LIST",
         "ViTForImageClassification",
@@ -36,14 +53,24 @@ if is_torch_available():
         "ViTPreTrainedModel",
     ]
 
-if is_tf_available():
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_tf_vit"] = [
         "TFViTForImageClassification",
         "TFViTModel",
         "TFViTPreTrainedModel",
     ]
 
-if is_flax_available():
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
     _import_structure["modeling_flax_vit"] = [
         "FlaxViTForImageClassification",
         "FlaxViTModel",
@@ -53,10 +80,20 @@ if is_flax_available():
 if TYPE_CHECKING:
     from .configuration_vit import VIT_PRETRAINED_CONFIG_ARCHIVE_MAP, ViTConfig, ViTOnnxConfig
 
-    if is_vision_available():
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .feature_extraction_vit import ViTFeatureExtractor
 
-    if is_torch_available():
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_vit import (
             VIT_PRETRAINED_MODEL_ARCHIVE_LIST,
             ViTForImageClassification,
@@ -65,10 +102,20 @@ if TYPE_CHECKING:
             ViTPreTrainedModel,
         )
 
-    if is_tf_available():
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_tf_vit import TFViTForImageClassification, TFViTModel, TFViTPreTrainedModel
 
-    if is_flax_available():
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
         from .modeling_flax_vit import FlaxViTForImageClassification, FlaxViTModel, FlaxViTPreTrainedModel