Allow `trust_remote_code` in example scripts (#25248)

* pytorch examples

* pytorch mim no trainer

* cookiecutter

* flax examples

* missed line in pytorch run_glue

* tensorflow examples

* tensorflow run_clip

* tensorflow run_mlm

* tensorflow run_ner

* tensorflow run_clm

* pytorch example from_configs

* pytorch no trainer examples

* Revert "tensorflow run_clip"

This reverts commit 261f86ac1f1c9e05dd3fd0291e1a1f8e573781d5.

* fix: duplicated argument

examples/pytorch/question-answering/run_qa.py

@@ -95,6 +95,16 @@ class ModelArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
 
 
 @dataclass
@@ -336,6 +346,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
@@ -343,6 +354,7 @@ def main():
         use_fast=True,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     model = AutoModelForQuestionAnswering.from_pretrained(
         model_args.model_name_or_path,
@@ -351,6 +363,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
 
     # Tokenizer check: this script requires a fast tokenizer.

examples/pytorch/question-answering/run_qa_no_trainer.py

@@ -273,6 +273,16 @@ def parse_args():
         "--hub_model_id", type=str, help="The name of the repository to keep in sync with the local `output_dir`."
     )
     parser.add_argument("--hub_token", type=str, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--trust_remote_code",
+        type=bool,
+        default=False,
+        help=(
+            "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+            "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+            "execute code present on the Hub on your local machine."
+        ),
+    )
     parser.add_argument(
         "--checkpointing_steps",
         type=str,
@@ -415,17 +425,21 @@ def main():
     # download model & vocab.
 
     if args.config_name:
-        config = AutoConfig.from_pretrained(args.config_name)
+        config = AutoConfig.from_pretrained(args.config_name, trust_remote_code=args.trust_remote_code)
     elif args.model_name_or_path:
-        config = AutoConfig.from_pretrained(args.model_name_or_path)
+        config = AutoConfig.from_pretrained(args.model_name_or_path, trust_remote_code=args.trust_remote_code)
     else:
         config = CONFIG_MAPPING[args.model_type]()
         logger.warning("You are instantiating a new config instance from scratch.")
 
     if args.tokenizer_name:
-        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=True)
+        tokenizer = AutoTokenizer.from_pretrained(
+            args.tokenizer_name, use_fast=True, trust_remote_code=args.trust_remote_code
+        )
     elif args.model_name_or_path:
-        tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=True)
+        tokenizer = AutoTokenizer.from_pretrained(
+            args.model_name_or_path, use_fast=True, trust_remote_code=args.trust_remote_code
+        )
     else:
         raise ValueError(
             "You are instantiating a new tokenizer from scratch. This is not supported by this script."
@@ -437,10 +451,11 @@ def main():
             args.model_name_or_path,
             from_tf=bool(".ckpt" in args.model_name_or_path),
             config=config,
+            trust_remote_code=args.trust_remote_code,
         )
     else:
         logger.info("Training new model from scratch")
-        model = AutoModelForQuestionAnswering.from_config(config)
+        model = AutoModelForQuestionAnswering.from_config(config, trust_remote_code=args.trust_remote_code)
 
     # Preprocessing the datasets.
     # Preprocessing is slighlty different for training and evaluation.

examples/pytorch/question-answering/run_seq2seq_qa.py

@@ -96,6 +96,16 @@ class ModelArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
 
 
 @dataclass
@@ -381,6 +391,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
@@ -388,6 +399,7 @@ def main():
         use_fast=model_args.use_fast_tokenizer,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     model = AutoModelForSeq2SeqLM.from_pretrained(
         model_args.model_name_or_path,
@@ -396,6 +408,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
 
     # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch

examples/pytorch/semantic-segmentation/run_semantic_segmentation.py

@@ -257,6 +257,16 @@ class ModelArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
 
 
 def main():
@@ -393,6 +403,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     model = AutoModelForSemanticSegmentation.from_pretrained(
         model_args.model_name_or_path,
@@ -401,12 +412,14 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     image_processor = AutoImageProcessor.from_pretrained(
         model_args.image_processor_name or model_args.model_name_or_path,
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
 
     # Define torchvision transforms to be applied to each image + target.

examples/pytorch/semantic-segmentation/run_semantic_segmentation_no_trainer.py

@@ -273,6 +273,16 @@ def parse_args():
         "--hub_model_id", type=str, help="The name of the repository to keep in sync with the local `output_dir`."
     )
     parser.add_argument("--hub_token", type=str, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--trust_remote_code",
+        type=bool,
+        default=False,
+        help=(
+            "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+            "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+            "execute code present on the Hub on your local machine."
+        ),
+    )
     parser.add_argument(
         "--checkpointing_steps",
         type=str,
@@ -400,9 +410,15 @@ def main():
     label2id = {v: k for k, v in id2label.items()}
 
     # Load pretrained model and image processor
-    config = AutoConfig.from_pretrained(args.model_name_or_path, id2label=id2label, label2id=label2id)
-    image_processor = AutoImageProcessor.from_pretrained(args.model_name_or_path)
-    model = AutoModelForSemanticSegmentation.from_pretrained(args.model_name_or_path, config=config)
+    config = AutoConfig.from_pretrained(
+        args.model_name_or_path, id2label=id2label, label2id=label2id, trust_remote_code=args.trust_remote_code
+    )
+    image_processor = AutoImageProcessor.from_pretrained(
+        args.model_name_or_path, trust_remote_code=args.trust_remote_code
+    )
+    model = AutoModelForSemanticSegmentation.from_pretrained(
+        args.model_name_or_path, config=config, trust_remote_code=args.trust_remote_code
+    )
 
     # Preprocessing the datasets
     # Define torchvision transforms to be applied to each image + target.

examples/pytorch/speech-recognition/run_speech_recognition_ctc.py

@@ -244,6 +244,16 @@ class DataTrainingArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
     unk_token: str = field(
         default="[UNK]",
         metadata={"help": "The unk token for the tokenizer"},
@@ -505,6 +515,7 @@ def main():
         model_args.model_name_or_path,
         cache_dir=model_args.cache_dir,
         token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
     )
 
     # 4. Next, if no tokenizer file is defined,
@@ -561,12 +572,14 @@ def main():
     tokenizer = AutoTokenizer.from_pretrained(
         tokenizer_name_or_path,
         token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
         **tokenizer_kwargs,
     )
     feature_extractor = AutoFeatureExtractor.from_pretrained(
         model_args.model_name_or_path,
         cache_dir=model_args.cache_dir,
         token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
     )
 
     # adapt config
@@ -595,6 +608,7 @@ def main():
         cache_dir=model_args.cache_dir,
         config=config,
         token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
     )
 
     # freeze encoder

examples/pytorch/speech-recognition/run_speech_recognition_ctc_adapter.py

@@ -247,6 +247,16 @@ class DataTrainingArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
     unk_token: str = field(
         default="[UNK]",
         metadata={"help": "The unk token for the tokenizer"},
@@ -501,6 +511,7 @@ def main():
         model_args.model_name_or_path,
         cache_dir=model_args.cache_dir,
         token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
     )
 
     # 4. Next, if no tokenizer file is defined,
@@ -517,6 +528,7 @@ def main():
         tokenizer = AutoTokenizer.from_pretrained(
             tokenizer_name_or_path,
             token=data_args.token,
+            trust_remote_code=data_args.trust_remote_code,
         )
         vocab_dict = tokenizer.vocab.copy()
         if tokenizer.target_lang is None:
@@ -584,12 +596,14 @@ def main():
     tokenizer = AutoTokenizer.from_pretrained(
         tokenizer_name_or_path,
         token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
         **tokenizer_kwargs,
     )
     feature_extractor = AutoFeatureExtractor.from_pretrained(
         model_args.model_name_or_path,
         cache_dir=model_args.cache_dir,
         token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
     )
 
     # adapt config
@@ -615,6 +629,7 @@ def main():
         cache_dir=model_args.cache_dir,
         config=config,
         token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
         ignore_mismatched_sizes=True,
     )
 

examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py

@@ -101,6 +101,16 @@ class ModelArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
     freeze_feature_encoder: bool = field(
         default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
     )
@@ -384,6 +394,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
 
     config.update({"forced_decoder_ids": model_args.forced_decoder_ids, "suppress_tokens": model_args.suppress_tokens})
@@ -397,6 +408,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
@@ -404,6 +416,7 @@ def main():
         use_fast=model_args.use_fast_tokenizer,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     model = AutoModelForSpeechSeq2Seq.from_pretrained(
         model_args.model_name_or_path,
@@ -411,6 +424,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
 
     if model.config.decoder_start_token_id is None:

examples/pytorch/summarization/run_summarization.py

@@ -115,6 +115,16 @@ class ModelArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
     resize_position_embeddings: Optional[bool] = field(
         default=None,
         metadata={
@@ -431,6 +441,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
@@ -438,6 +449,7 @@ def main():
         use_fast=model_args.use_fast_tokenizer,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     model = AutoModelForSeq2SeqLM.from_pretrained(
         model_args.model_name_or_path,
@@ -446,6 +458,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
 
     # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch

examples/pytorch/summarization/run_summarization_no_trainer.py

@@ -266,6 +266,16 @@ def parse_args():
         "--hub_model_id", type=str, help="The name of the repository to keep in sync with the local `output_dir`."
     )
     parser.add_argument("--hub_token", type=str, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--trust_remote_code",
+        type=bool,
+        default=False,
+        help=(
+            "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+            "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+            "execute code present on the Hub on your local machine."
+        ),
+    )
     parser.add_argument(
         "--checkpointing_steps",
         type=str,
@@ -406,17 +416,21 @@ def main():
     # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
     # download model & vocab.
     if args.config_name:
-        config = AutoConfig.from_pretrained(args.config_name)
+        config = AutoConfig.from_pretrained(args.config_name, trust_remote_code=args.trust_remote_code)
     elif args.model_name_or_path:
-        config = AutoConfig.from_pretrained(args.model_name_or_path)
+        config = AutoConfig.from_pretrained(args.model_name_or_path, trust_remote_code=args.trust_remote_code)
     else:
         config = CONFIG_MAPPING[args.model_type]()
         logger.warning("You are instantiating a new config instance from scratch.")
 
     if args.tokenizer_name:
-        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
+        tokenizer = AutoTokenizer.from_pretrained(
+            args.tokenizer_name, use_fast=not args.use_slow_tokenizer, trust_remote_code=args.trust_remote_code
+        )
     elif args.model_name_or_path:
-        tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
+        tokenizer = AutoTokenizer.from_pretrained(
+            args.model_name_or_path, use_fast=not args.use_slow_tokenizer, trust_remote_code=args.trust_remote_code
+        )
     else:
         raise ValueError(
             "You are instantiating a new tokenizer from scratch. This is not supported by this script."
@@ -428,10 +442,11 @@ def main():
             args.model_name_or_path,
             from_tf=bool(".ckpt" in args.model_name_or_path),
             config=config,
+            trust_remote_code=args.trust_remote_code,
         )
     else:
         logger.info("Training new model from scratch")
-        model = AutoModelForSeq2SeqLM.from_config(config)
+        model = AutoModelForSeq2SeqLM.from_config(config, trust_remote_code=args.trust_remote_code)
 
     # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
     # on a small vocab and want a smaller embedding size, remove this test.

examples/pytorch/text-classification/run_classification.py

@@ -243,6 +243,16 @@ class ModelArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
     ignore_mismatched_sizes: bool = field(
         default=False,
         metadata={"help": "Will enable to load a pretrained model whose head dimensions are different."},
@@ -482,6 +492,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
 
     if is_regression:
@@ -500,6 +511,7 @@ def main():
         use_fast=model_args.use_fast_tokenizer,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     model = AutoModelForSequenceClassification.from_pretrained(
         model_args.model_name_or_path,
@@ -508,6 +520,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
         ignore_mismatched_sizes=model_args.ignore_mismatched_sizes,
     )
 

examples/pytorch/text-classification/run_glue.py

@@ -204,6 +204,16 @@ class ModelArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
     ignore_mismatched_sizes: bool = field(
         default=False,
         metadata={"help": "Will enable to load a pretrained model whose head dimensions are different."},
@@ -375,6 +385,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
@@ -382,6 +393,7 @@ def main():
         use_fast=model_args.use_fast_tokenizer,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     model = AutoModelForSequenceClassification.from_pretrained(
         model_args.model_name_or_path,
@@ -390,6 +402,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
         ignore_mismatched_sizes=model_args.ignore_mismatched_sizes,
     )
 

examples/pytorch/text-classification/run_glue_no_trainer.py

@@ -156,6 +156,16 @@ def parse_args():
         "--hub_model_id", type=str, help="The name of the repository to keep in sync with the local `output_dir`."
     )
     parser.add_argument("--hub_token", type=str, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--trust_remote_code",
+        type=bool,
+        default=False,
+        help=(
+            "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+            "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+            "execute code present on the Hub on your local machine."
+        ),
+    )
     parser.add_argument(
         "--checkpointing_steps",
         type=str,
@@ -309,13 +319,21 @@ def main():
     #
     # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
     # download model & vocab.
-    config = AutoConfig.from_pretrained(args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name)
-    tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
+    config = AutoConfig.from_pretrained(
+        args.model_name_or_path,
+        num_labels=num_labels,
+        finetuning_task=args.task_name,
+        trust_remote_code=args.trust_remote_code,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        args.model_name_or_path, use_fast=not args.use_slow_tokenizer, trust_remote_code=args.trust_remote_code
+    )
     model = AutoModelForSequenceClassification.from_pretrained(
         args.model_name_or_path,
         from_tf=bool(".ckpt" in args.model_name_or_path),
         config=config,
         ignore_mismatched_sizes=args.ignore_mismatched_sizes,
+        trust_remote_code=args.trust_remote_code,
     )
 
     # Preprocessing the datasets

examples/pytorch/text-classification/run_xnli.py

@@ -168,6 +168,16 @@ class ModelArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
     ignore_mismatched_sizes: bool = field(
         default=False,
         metadata={"help": "Will enable to load a pretrained model whose head dimensions are different."},
@@ -292,6 +302,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
@@ -300,6 +311,7 @@ def main():
         use_fast=model_args.use_fast_tokenizer,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     model = AutoModelForSequenceClassification.from_pretrained(
         model_args.model_name_or_path,
@@ -308,6 +320,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
         ignore_mismatched_sizes=model_args.ignore_mismatched_sizes,
     )
 

examples/pytorch/token-classification/run_ner.py

@@ -95,6 +95,16 @@ class ModelArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
     ignore_mismatched_sizes: bool = field(
         default=False,
         metadata={"help": "Will enable to load a pretrained model whose head dimensions are different."},
@@ -362,6 +372,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
 
     tokenizer_name_or_path = model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path
@@ -372,6 +383,7 @@ def main():
             use_fast=True,
             revision=model_args.model_revision,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
             add_prefix_space=True,
         )
     else:
@@ -381,6 +393,7 @@ def main():
             use_fast=True,
             revision=model_args.model_revision,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
 
     model = AutoModelForTokenClassification.from_pretrained(
@@ -390,6 +403,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
         ignore_mismatched_sizes=model_args.ignore_mismatched_sizes,
     )
 

examples/pytorch/token-classification/run_ner_no_trainer.py

@@ -210,6 +210,16 @@ def parse_args():
         "--hub_model_id", type=str, help="The name of the repository to keep in sync with the local `output_dir`."
     )
     parser.add_argument("--hub_token", type=str, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--trust_remote_code",
+        type=bool,
+        default=False,
+        help=(
+            "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+            "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+            "execute code present on the Hub on your local machine."
+        ),
+    )
     parser.add_argument(
         "--checkpointing_steps",
         type=str,
@@ -388,9 +398,13 @@ def main():
     # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
     # download model & vocab.
     if args.config_name:
-        config = AutoConfig.from_pretrained(args.config_name, num_labels=num_labels)
+        config = AutoConfig.from_pretrained(
+            args.config_name, num_labels=num_labels, trust_remote_code=args.trust_remote_code
+        )
     elif args.model_name_or_path:
-        config = AutoConfig.from_pretrained(args.model_name_or_path, num_labels=num_labels)
+        config = AutoConfig.from_pretrained(
+            args.model_name_or_path, num_labels=num_labels, trust_remote_code=args.trust_remote_code
+        )
     else:
         config = CONFIG_MAPPING[args.model_type]()
         logger.warning("You are instantiating a new config instance from scratch.")
@@ -403,9 +417,13 @@ def main():
         )
 
     if config.model_type in {"bloom", "gpt2", "roberta"}:
-        tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path, use_fast=True, add_prefix_space=True)
+        tokenizer = AutoTokenizer.from_pretrained(
+            tokenizer_name_or_path, use_fast=True, add_prefix_space=True, trust_remote_code=args.trust_remote_code
+        )
     else:
-        tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path, use_fast=True)
+        tokenizer = AutoTokenizer.from_pretrained(
+            tokenizer_name_or_path, use_fast=True, trust_remote_code=args.trust_remote_code
+        )
 
     if args.model_name_or_path:
         model = AutoModelForTokenClassification.from_pretrained(
@@ -413,10 +431,11 @@ def main():
             from_tf=bool(".ckpt" in args.model_name_or_path),
             config=config,
             ignore_mismatched_sizes=args.ignore_mismatched_sizes,
+            trust_remote_code=args.trust_remote_code,
         )
     else:
         logger.info("Training new model from scratch")
-        model = AutoModelForTokenClassification.from_config(config)
+        model = AutoModelForTokenClassification.from_config(config, trust_remote_code=args.trust_remote_code)
 
     # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
     # on a small vocab and want a smaller embedding size, remove this test.

examples/pytorch/translation/run_translation.py

@@ -105,6 +105,16 @@ class ModelArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
 
 
 @dataclass
@@ -380,6 +390,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
@@ -387,6 +398,7 @@ def main():
         use_fast=model_args.use_fast_tokenizer,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     model = AutoModelForSeq2SeqLM.from_pretrained(
         model_args.model_name_or_path,
@@ -395,6 +407,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
 
     # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch

examples/pytorch/translation/run_translation_no_trainer.py

@@ -257,6 +257,16 @@ def parse_args():
         "--hub_model_id", type=str, help="The name of the repository to keep in sync with the local `output_dir`."
     )
     parser.add_argument("--hub_token", type=str, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--trust_remote_code",
+        type=bool,
+        default=False,
+        help=(
+            "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+            "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+            "execute code present on the Hub on your local machine."
+        ),
+    )
     parser.add_argument(
         "--checkpointing_steps",
         type=str,
@@ -386,17 +396,21 @@ def main():
     # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
     # download model & vocab.
     if args.config_name:
-        config = AutoConfig.from_pretrained(args.config_name)
+        config = AutoConfig.from_pretrained(args.config_name, trust_remote_code=args.trust_remote_code)
     elif args.model_name_or_path:
-        config = AutoConfig.from_pretrained(args.model_name_or_path)
+        config = AutoConfig.from_pretrained(args.model_name_or_path, trust_remote_code=args.trust_remote_code)
     else:
         config = CONFIG_MAPPING[args.model_type]()
         logger.warning("You are instantiating a new config instance from scratch.")
 
     if args.tokenizer_name:
-        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
+        tokenizer = AutoTokenizer.from_pretrained(
+            args.tokenizer_name, use_fast=not args.use_slow_tokenizer, trust_remote_code=args.trust_remote_code
+        )
     elif args.model_name_or_path:
-        tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
+        tokenizer = AutoTokenizer.from_pretrained(
+            args.model_name_or_path, use_fast=not args.use_slow_tokenizer, trust_remote_code=args.trust_remote_code
+        )
     else:
         raise ValueError(
             "You are instantiating a new tokenizer from scratch. This is not supported by this script."
@@ -408,10 +422,11 @@ def main():
             args.model_name_or_path,
             from_tf=bool(".ckpt" in args.model_name_or_path),
             config=config,
+            trust_remote_code=args.trust_remote_code,
         )
     else:
         logger.info("Training new model from scratch")
-        model = AutoModelForSeq2SeqLM.from_config(config)
+        model = AutoModelForSeq2SeqLM.from_config(config, trust_remote_code=args.trust_remote_code)
 
     # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
     # on a small vocab and want a smaller embedding size, remove this test.

examples/tensorflow/image-classification/run_image_classification.py

@@ -173,6 +173,16 @@ class ModelArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
     ignore_mismatched_sizes: bool = field(
         default=False,
         metadata={"help": "Will enable to load a pretrained model whose head dimensions are different."},
@@ -323,12 +333,14 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     image_processor = AutoImageProcessor.from_pretrained(
         model_args.image_processor_name or model_args.model_name_or_path,
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
 
     # If we don't have a validation split, split off a percentage of train as validation.
@@ -449,6 +461,7 @@ def main():
             cache_dir=model_args.cache_dir,
             revision=model_args.model_revision,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
             ignore_mismatched_sizes=model_args.ignore_mismatched_sizes,
         )
         num_replicas = training_args.strategy.num_replicas_in_sync

examples/tensorflow/language-modeling/run_clm.py

@@ -128,6 +128,16 @@ class ModelArguments:
             "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token`."
         },
     )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will"
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
 
     def __post_init__(self):
         if self.config_overrides is not None and (self.config_name is not None or self.model_name_or_path is not None):
@@ -366,17 +376,26 @@ def main():
     # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
     # download model & vocab.
     if model_args.config_name:
-        config = AutoConfig.from_pretrained(model_args.config_name)
+        config = AutoConfig.from_pretrained(
+            model_args.config_name,
+            trust_remote_code=model_args.trust_remote_code,
+        )
     elif model_args.model_name_or_path:
-        config = AutoConfig.from_pretrained(model_args.model_name_or_path)
+        config = AutoConfig.from_pretrained(
+            model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code
+        )
     else:
         config = CONFIG_MAPPING[model_args.model_type]()
         logger.warning("You are instantiating a new config instance from scratch.")
 
     if model_args.tokenizer_name:
-        tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name)
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, trust_remote_code=model_args.trust_remote_code
+        )
     elif model_args.model_name_or_path:
-        tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path)
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code
+        )
     else:
         raise ValueError(
             "You are instantiating a new tokenizer from scratch. This is not supported by this script."
@@ -479,12 +498,16 @@ def main():
     with training_args.strategy.scope():
         # region Prepare model
         if checkpoint is not None:
-            model = TFAutoModelForCausalLM.from_pretrained(checkpoint, config=config)
+            model = TFAutoModelForCausalLM.from_pretrained(
+                checkpoint, config=config, trust_remote_code=model_args.trust_remote_code
+            )
         elif model_args.model_name_or_path:
-            model = TFAutoModelForCausalLM.from_pretrained(model_args.model_name_or_path, config=config)
+            model = TFAutoModelForCausalLM.from_pretrained(
+                model_args.model_name_or_path, config=config, trust_remote_code=model_args.trust_remote_code
+            )
         else:
             logger.info("Training new model from scratch")
-            model = TFAutoModelForCausalLM.from_config(config)
+            model = TFAutoModelForCausalLM.from_config(config, trust_remote_code=model_args.trust_remote_code)
 
         # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
         # on a small vocab and want a smaller embedding size, remove this test.