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/flax/image-captioning/run_image_captioning_flax.py

@@ -198,6 +198,16 @@ 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
@@ -489,17 +499,20 @@ def main():
         seed=training_args.seed,
         dtype=getattr(jnp, model_args.dtype),
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     image_processor = AutoImageProcessor.from_pretrained(
         model_args.model_name_or_path,
         cache_dir=model_args.cache_dir,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.model_name_or_path,
         cache_dir=model_args.cache_dir,
         use_fast=model_args.use_fast_tokenizer,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     tokenizer.pad_token = tokenizer.convert_ids_to_tokens(model.config.pad_token_id)
 

examples/flax/language-modeling/run_clm_flax.py

@@ -185,6 +185,16 @@ 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
@@ -477,12 +487,14 @@ def main():
             model_args.config_name,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            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,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     else:
         config = CONFIG_MAPPING[model_args.model_type]()
@@ -494,6 +506,7 @@ def main():
             cache_dir=model_args.cache_dir,
             use_fast=model_args.use_fast_tokenizer,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     elif model_args.model_name_or_path:
         tokenizer = AutoTokenizer.from_pretrained(
@@ -501,6 +514,7 @@ def main():
             cache_dir=model_args.cache_dir,
             use_fast=model_args.use_fast_tokenizer,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     else:
         raise ValueError(
@@ -515,12 +529,14 @@ def main():
             seed=training_args.seed,
             dtype=getattr(jnp, model_args.dtype),
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     else:
         model = FlaxAutoModelForCausalLM.from_config(
             config,
             seed=training_args.seed,
             dtype=getattr(jnp, model_args.dtype),
+            trust_remote_code=model_args.trust_remote_code,
         )
 
     # Preprocessing the datasets.

examples/flax/language-modeling/run_mlm_flax.py

@@ -190,6 +190,16 @@ 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
@@ -509,12 +519,14 @@ def main():
             model_args.config_name,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            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,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     else:
         config = CONFIG_MAPPING[model_args.model_type]()
@@ -526,6 +538,7 @@ def main():
             cache_dir=model_args.cache_dir,
             use_fast=model_args.use_fast_tokenizer,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     elif model_args.model_name_or_path:
         tokenizer = AutoTokenizer.from_pretrained(
@@ -533,6 +546,7 @@ def main():
             cache_dir=model_args.cache_dir,
             use_fast=model_args.use_fast_tokenizer,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     else:
         raise ValueError(
@@ -652,12 +666,14 @@ def main():
             seed=training_args.seed,
             dtype=getattr(jnp, model_args.dtype),
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     else:
         model = FlaxAutoModelForMaskedLM.from_config(
             config,
             seed=training_args.seed,
             dtype=getattr(jnp, model_args.dtype),
+            trust_remote_code=model_args.trust_remote_code,
         )
 
     if training_args.gradient_checkpointing:

examples/flax/question-answering/run_qa.py

@@ -171,6 +171,16 @@ 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."
+            )
+        },
+    )
     dtype: Optional[str] = field(
         default="float32",
         metadata={
@@ -534,6 +544,7 @@ 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,
@@ -541,6 +552,7 @@ def main():
         use_fast=True,
         revision=model_args.model_revision,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     # endregion
 
@@ -888,6 +900,7 @@ 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,
         seed=training_args.seed,
         dtype=getattr(jnp, model_args.dtype),
     )

examples/flax/summarization/run_summarization_flax.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."
+            )
+        },
+    )
 
 
 @dataclass
@@ -517,12 +527,14 @@ def main():
             model_args.config_name,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            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,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     else:
         config = CONFIG_MAPPING[model_args.model_type]()
@@ -534,6 +546,7 @@ def main():
             cache_dir=model_args.cache_dir,
             use_fast=model_args.use_fast_tokenizer,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     elif model_args.model_name_or_path:
         tokenizer = AutoTokenizer.from_pretrained(
@@ -541,6 +554,7 @@ def main():
             cache_dir=model_args.cache_dir,
             use_fast=model_args.use_fast_tokenizer,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     else:
         raise ValueError(
@@ -555,12 +569,14 @@ def main():
             seed=training_args.seed,
             dtype=getattr(jnp, model_args.dtype),
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     else:
         model = FlaxAutoModelForSeq2SeqLM.from_config(
             config,
             seed=training_args.seed,
             dtype=getattr(jnp, model_args.dtype),
+            trust_remote_code=model_args.trust_remote_code,
         )
 
     if training_args.gradient_checkpointing:

examples/flax/text-classification/run_flax_glue.py

@@ -117,6 +117,16 @@ 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
@@ -425,16 +435,19 @@ def main():
         num_labels=num_labels,
         finetuning_task=data_args.task_name,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.model_name_or_path,
         use_fast=not model_args.use_slow_tokenizer,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
     model = FlaxAutoModelForSequenceClassification.from_pretrained(
         model_args.model_name_or_path,
         config=config,
         token=model_args.token,
+        trust_remote_code=model_args.trust_remote_code,
     )
 
     # Preprocessing the datasets

examples/flax/token-classification/run_flax_ner.py

@@ -165,6 +165,16 @@ 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
@@ -504,6 +514,7 @@ 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
     if config.model_type in {"gpt2", "roberta"}:
@@ -512,6 +523,7 @@ 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,
             add_prefix_space=True,
         )
     else:
@@ -520,6 +532,7 @@ 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 = FlaxAutoModelForTokenClassification.from_pretrained(
         model_args.model_name_or_path,
@@ -527,6 +540,7 @@ 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,
     )
 
     # Preprocessing the datasets

examples/flax/vision/run_image_classification.py

@@ -175,6 +175,16 @@ 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
@@ -352,6 +362,7 @@ def main():
             image_size=data_args.image_size,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     elif model_args.model_name_or_path:
         config = AutoConfig.from_pretrained(
@@ -360,6 +371,7 @@ def main():
             image_size=data_args.image_size,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     else:
         config = CONFIG_MAPPING[model_args.model_type]()
@@ -372,12 +384,14 @@ def main():
             seed=training_args.seed,
             dtype=getattr(jnp, model_args.dtype),
             token=model_args.token,
+            trust_remote_code=model_args.trust_remote_code,
         )
     else:
         model = FlaxAutoModelForImageClassification.from_config(
             config,
             seed=training_args.seed,
             dtype=getattr(jnp, model_args.dtype),
+            trust_remote_code=model_args.trust_remote_code,
         )
 
     # Store some constant

examples/pytorch/audio-classification/run_audio_classification.py

@@ -167,6 +167,16 @@ 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_extractor: Optional[bool] = field(
         default=None, metadata={"help": "Whether to freeze the feature extractor layers of the model."}
     )
@@ -293,6 +303,7 @@ 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,
     )
 
     # `datasets` takes care of automatically loading and resampling the audio,
@@ -353,6 +364,7 @@ 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 = AutoModelForAudioClassification.from_pretrained(
         model_args.model_name_or_path,
@@ -361,6 +373,7 @@ 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/contrastive-image-text/run_clip.py

@@ -102,6 +102,16 @@ 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_vision_model: bool = field(
         default=False, metadata={"help": "Whether to freeze the vision model parameters or not."}
     )
@@ -350,6 +360,7 @@ 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 = AutoModel.from_pretrained(
@@ -357,6 +368,7 @@ 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 = model.config
 

examples/pytorch/image-classification/run_image_classification.py

@@ -158,6 +158,16 @@ 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."},
@@ -290,6 +300,7 @@ 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 = AutoModelForImageClassification.from_pretrained(
         model_args.model_name_or_path,
@@ -298,6 +309,7 @@ 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,
     )
     image_processor = AutoImageProcessor.from_pretrained(
@@ -305,6 +317,7 @@ 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,
     )
 
     # Define torchvision transforms to be applied to each image.

examples/pytorch/image-classification/run_image_classification_no_trainer.py

@@ -146,6 +146,16 @@ 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,
@@ -300,13 +310,18 @@ def main():
         i2label=id2label,
         label2id=label2id,
         finetuning_task="image-classification",
+        trust_remote_code=args.trust_remote_code,
+    )
+    image_processor = AutoImageProcessor.from_pretrained(
+        args.model_name_or_path,
+        trust_remote_code=args.trust_remote_code,
     )
-    image_processor = AutoImageProcessor.from_pretrained(args.model_name_or_path)
     model = AutoModelForImageClassification.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/image-pretraining/run_mim.py

@@ -169,6 +169,16 @@ 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."
+            )
+        },
+    )
     image_size: Optional[int] = field(
         default=None,
         metadata={
@@ -319,6 +329,7 @@ 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_args.config_name_or_path:
         config = AutoConfig.from_pretrained(model_args.config_name_or_path, **config_kwargs)
@@ -371,10 +382,11 @@ 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,
         )
     else:
         logger.info("Training new model from scratch")
-        model = AutoModelForMaskedImageModeling.from_config(config)
+        model = AutoModelForMaskedImageModeling.from_config(config, trust_remote_code=model_args.trust_remote_code)
 
     if training_args.do_train:
         column_names = ds["train"].column_names

examples/pytorch/image-pretraining/run_mim_no_trainer.py

@@ -195,6 +195,16 @@ def parse_args():
             "with private models)."
         ),
     )
+    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(
         "--image_size",
         type=int,
@@ -448,6 +458,7 @@ def main():
         "cache_dir": args.cache_dir,
         "revision": args.model_revision,
         "use_auth_token": True if args.use_auth_token else None,
+        "trust_remote_code": args.trust_remote_code,
     }
     if args.config_name_or_path:
         config = AutoConfig.from_pretrained(args.config_name_or_path, **config_kwargs)
@@ -498,10 +509,14 @@ def main():
             cache_dir=args.cache_dir,
             revision=args.model_revision,
             token=True if args.use_auth_token else None,
+            trust_remote_code=args.trust_remote_code,
         )
     else:
         logger.info("Training new model from scratch")
-        model = AutoModelForMaskedImageModeling.from_config(config)
+        model = AutoModelForMaskedImageModeling.from_config(
+            config,
+            trust_remote_code=args.trust_remote_code,
+        )
 
     column_names = ds["train"].column_names
 

examples/pytorch/language-modeling/run_clm.py

@@ -127,6 +127,16 @@ 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."
+            )
+        },
+    )
     torch_dtype: Optional[str] = field(
         default=None,
         metadata={
@@ -387,6 +397,7 @@ 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_args.config_name:
         config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
@@ -405,6 +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,
     }
     if model_args.tokenizer_name:
         tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
@@ -429,11 +441,12 @@ 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,
             torch_dtype=torch_dtype,
             low_cpu_mem_usage=model_args.low_cpu_mem_usage,
         )
     else:
-        model = AutoModelForCausalLM.from_config(config)
+        model = AutoModelForCausalLM.from_config(config, trust_remote_code=model_args.trust_remote_code)
         n_params = sum({p.data_ptr(): p.numel() for p in model.parameters()}.values())
         logger.info(f"Training new model from scratch - Total size={n_params/2**20:.2f}M params")
 

examples/pytorch/language-modeling/run_clm_no_trainer.py

@@ -193,6 +193,16 @@ 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,
@@ -362,17 +372,27 @@ 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."
@@ -385,10 +405,11 @@ def main():
             from_tf=bool(".ckpt" in args.model_name_or_path),
             config=config,
             low_cpu_mem_usage=args.low_cpu_mem_usage,
+            trust_remote_code=args.trust_remote_code,
         )
     else:
         logger.info("Training new model from scratch")
-        model = AutoModelForCausalLM.from_config(config)
+        model = AutoModelForCausalLM.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/language-modeling/run_mlm.py

@@ -123,6 +123,16 @@ 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."
+            )
+        },
+    )
     low_cpu_mem_usage: bool = field(
         default=False,
         metadata={
@@ -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,
     }
     if model_args.config_name:
         config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
@@ -398,6 +409,7 @@ 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,
     }
     if model_args.tokenizer_name:
         tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
@@ -417,11 +429,12 @@ 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,
             low_cpu_mem_usage=model_args.low_cpu_mem_usage,
         )
     else:
         logger.info("Training new model from scratch")
-        model = AutoModelForMaskedLM.from_config(config)
+        model = AutoModelForMaskedLM.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.

examples/pytorch/language-modeling/run_mlm_no_trainer.py

@@ -200,6 +200,16 @@ 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,
@@ -367,17 +377,21 @@ 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."
@@ -390,10 +404,11 @@ def main():
             from_tf=bool(".ckpt" in args.model_name_or_path),
             config=config,
             low_cpu_mem_usage=args.low_cpu_mem_usage,
+            trust_remote_code=args.trust_remote_code,
         )
     else:
         logger.info("Training new model from scratch")
-        model = AutoModelForMaskedLM.from_config(config)
+        model = AutoModelForMaskedLM.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/multiple-choice/run_swag.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
@@ -328,6 +338,7 @@ 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,
@@ -335,6 +346,7 @@ 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 = AutoModelForMultipleChoice.from_pretrained(
         model_args.model_name_or_path,
@@ -343,6 +355,7 @@ 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,
     )
 
     # When using your own dataset or a different dataset from swag, you will probably need to change this.

examples/pytorch/multiple-choice/run_swag_no_trainer.py

@@ -182,6 +182,16 @@ 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,
@@ -374,17 +384,21 @@ 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.model_name_or_path)
+        config = AutoConfig.from_pretrained(args.model_name_or_path, 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."
@@ -396,10 +410,11 @@ 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 = AutoModelForMultipleChoice.from_config(config)
+        model = AutoModelForMultipleChoice.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.