From 41a0e58e5ba4945c244a6141c2adc600fa0a7fdc Mon Sep 17 00:00:00 2001
From: cyyever <cyyever@outlook.com>
Date: Thu, 27 Mar 2025 22:55:50 +0800
Subject: [PATCH] Set weights_only in torch.load (#36991)

---
 .../flax/vision/run_image_classification.py   |  2 +-
 .../multiple_choice/utils_multiple_choice.py  |  2 +-
 examples/legacy/pytorch-lightning/run_glue.py |  2 +-
 examples/legacy/pytorch-lightning/run_ner.py  |  4 +-
 .../legacy/question-answering/run_squad.py    |  6 +-
 examples/legacy/run_swag.py                   |  2 +-
 .../legacy/seq2seq/convert_model_to_fp16.py   |  2 +-
 .../legacy/token-classification/utils_ner.py  |  2 +-
 .../convert_pytorch_checkpoint_to_tf2.py      |  7 +--
 src/transformers/data/datasets/squad.py       |  2 +-
 .../modeling_flax_pytorch_utils.py            |  6 +-
 src/transformers/modeling_tf_pytorch_utils.py |  3 +-
 src/transformers/modeling_utils.py            |  8 +--
 ..._original_pytorch_checkpoint_to_pytorch.py |  2 +-
 .../models/phi/convert_phi_weights_to_hf.py   |  2 +-
 .../models/wav2vec2/modeling_wav2vec2.py      |  3 +-
 src/transformers/trainer.py                   | 57 +++++++++----------
 .../autoformer/test_modeling_autoformer.py    |  2 +-
 .../idefics/test_image_processing_idefics.py  |  2 +-
 .../models/informer/test_modeling_informer.py |  2 +-
 .../llava_next/test_modeling_llava_next.py    |  4 +-
 .../test_modeling_patchtsmixer.py             |  2 +-
 .../models/patchtst/test_modeling_patchtst.py |  2 +-
 .../test_modeling_time_series_transformer.py  |  2 +-
 .../models/videomae/test_modeling_videomae.py |  2 +-
 .../peft_integration/test_peft_integration.py |  8 +--
 tests/trainer/test_trainer.py                 |  2 +-
 tests/utils/test_modeling_utils.py            |  2 +-
 28 files changed, 64 insertions(+), 78 deletions(-)

diff --git a/examples/flax/vision/run_image_classification.py b/examples/flax/vision/run_image_classification.py
index 0228a8797b6..4b50352f580 100644
--- a/examples/flax/vision/run_image_classification.py
+++ b/examples/flax/vision/run_image_classification.py
@@ -36,12 +36,12 @@ import optax
 # for dataset and preprocessing
 import torch
 import torchvision
-import torchvision.transforms as transforms
 from flax import jax_utils
 from flax.jax_utils import pad_shard_unpad, unreplicate
 from flax.training import train_state
 from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
 from huggingface_hub import HfApi
+from torchvision import transforms
 from tqdm import tqdm
 
 import transformers
diff --git a/examples/legacy/multiple_choice/utils_multiple_choice.py b/examples/legacy/multiple_choice/utils_multiple_choice.py
index 6b7559c49e5..2ee901c1bc8 100644
--- a/examples/legacy/multiple_choice/utils_multiple_choice.py
+++ b/examples/legacy/multiple_choice/utils_multiple_choice.py
@@ -113,7 +113,7 @@ if is_torch_available():
             with FileLock(lock_path):
                 if os.path.exists(cached_features_file) and not overwrite_cache:
                     logger.info(f"Loading features from cached file {cached_features_file}")
-                    self.features = torch.load(cached_features_file)
+                    self.features = torch.load(cached_features_file, weights_only=True)
                 else:
                     logger.info(f"Creating features from dataset file at {data_dir}")
                     label_list = processor.get_labels()
diff --git a/examples/legacy/pytorch-lightning/run_glue.py b/examples/legacy/pytorch-lightning/run_glue.py
index 681f633fcd6..00302c5061c 100644
--- a/examples/legacy/pytorch-lightning/run_glue.py
+++ b/examples/legacy/pytorch-lightning/run_glue.py
@@ -81,7 +81,7 @@ class GLUETransformer(BaseTransformer):
 
         cached_features_file = self._feature_file(mode)
         logger.info("Loading features from cached file %s", cached_features_file)
-        features = torch.load(cached_features_file)
+        features = torch.load(cached_features_file, weights_only=True)
         all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
         all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
         all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
diff --git a/examples/legacy/pytorch-lightning/run_ner.py b/examples/legacy/pytorch-lightning/run_ner.py
index fc6f812275e..144759d36aa 100644
--- a/examples/legacy/pytorch-lightning/run_ner.py
+++ b/examples/legacy/pytorch-lightning/run_ner.py
@@ -63,7 +63,7 @@ class NERTransformer(BaseTransformer):
             cached_features_file = self._feature_file(mode)
             if os.path.exists(cached_features_file) and not args.overwrite_cache:
                 logger.info("Loading features from cached file %s", cached_features_file)
-                features = torch.load(cached_features_file)
+                features = torch.load(cached_features_file, weights_only=True)
             else:
                 logger.info("Creating features from dataset file at %s", args.data_dir)
                 examples = self.token_classification_task.read_examples_from_file(args.data_dir, mode)
@@ -89,7 +89,7 @@ class NERTransformer(BaseTransformer):
         "Load datasets. Called after prepare data."
         cached_features_file = self._feature_file(mode)
         logger.info("Loading features from cached file %s", cached_features_file)
-        features = torch.load(cached_features_file)
+        features = torch.load(cached_features_file, weights_only=True)
         all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
         all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
         if features[0].token_type_ids is not None:
diff --git a/examples/legacy/question-answering/run_squad.py b/examples/legacy/question-answering/run_squad.py
index 19d3c6f49e7..007e9e7ab27 100644
--- a/examples/legacy/question-answering/run_squad.py
+++ b/examples/legacy/question-answering/run_squad.py
@@ -105,8 +105,8 @@ def train(args, train_dataset, model, tokenizer):
         os.path.join(args.model_name_or_path, "scheduler.pt")
     ):
         # Load in optimizer and scheduler states
-        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
-        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
+        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt"), weights_only=True))
+        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt"), weights_only=True))
 
     if args.fp16:
         try:
@@ -417,7 +417,7 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
     # Init features and dataset from cache if it exists
     if os.path.exists(cached_features_file) and not args.overwrite_cache:
         logger.info("Loading features from cached file %s", cached_features_file)
-        features_and_dataset = torch.load(cached_features_file)
+        features_and_dataset = torch.load(cached_features_file, weights_only=True)
         features, dataset, examples = (
             features_and_dataset["features"],
             features_and_dataset["dataset"],
diff --git a/examples/legacy/run_swag.py b/examples/legacy/run_swag.py
index 8fca5507bc5..55fd0aa0520 100755
--- a/examples/legacy/run_swag.py
+++ b/examples/legacy/run_swag.py
@@ -244,7 +244,7 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
     )
     if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
         logger.info("Loading features from cached file %s", cached_features_file)
-        features = torch.load(cached_features_file)
+        features = torch.load(cached_features_file, weights_only=True)
     else:
         logger.info("Creating features from dataset file at %s", input_file)
         examples = read_swag_examples(input_file)
diff --git a/examples/legacy/seq2seq/convert_model_to_fp16.py b/examples/legacy/seq2seq/convert_model_to_fp16.py
index 7fffbde79df..8d568a7e4af 100755
--- a/examples/legacy/seq2seq/convert_model_to_fp16.py
+++ b/examples/legacy/seq2seq/convert_model_to_fp16.py
@@ -22,7 +22,7 @@ from tqdm import tqdm
 
 def convert(src_path: str, map_location: str = "cpu", save_path: Union[str, None] = None) -> None:
     """Convert a pytorch_model.bin or model.pt file to torch.float16 for faster downloads, less disk space."""
-    state_dict = torch.load(src_path, map_location=map_location)
+    state_dict = torch.load(src_path, map_location=map_location, weights_only=True)
     for k, v in tqdm(state_dict.items()):
         if not isinstance(v, torch.Tensor):
             raise TypeError("FP16 conversion only works on paths that are saved state dicts, like pytorch_model.bin")
diff --git a/examples/legacy/token-classification/utils_ner.py b/examples/legacy/token-classification/utils_ner.py
index da4d8c3b605..9167ce15161 100644
--- a/examples/legacy/token-classification/utils_ner.py
+++ b/examples/legacy/token-classification/utils_ner.py
@@ -242,7 +242,7 @@ if is_torch_available():
             with FileLock(lock_path):
                 if os.path.exists(cached_features_file) and not overwrite_cache:
                     logger.info(f"Loading features from cached file {cached_features_file}")
-                    self.features = torch.load(cached_features_file)
+                    self.features = torch.load(cached_features_file, weights_only=True)
                 else:
                     logger.info(f"Creating features from dataset file at {data_dir}")
                     examples = token_classification_task.read_examples_from_file(data_dir, mode)
diff --git a/src/transformers/convert_pytorch_checkpoint_to_tf2.py b/src/transformers/convert_pytorch_checkpoint_to_tf2.py
index 46c7ba12a54..fad6463e982 100755
--- a/src/transformers/convert_pytorch_checkpoint_to_tf2.py
+++ b/src/transformers/convert_pytorch_checkpoint_to_tf2.py
@@ -277,12 +277,7 @@ def convert_pt_checkpoint_to_tf(
     if compare_with_pt_model:
         tfo = tf_model(tf_model.dummy_inputs, training=False)  # build the network
 
-        weights_only_kwarg = {"weights_only": True}
-        state_dict = torch.load(
-            pytorch_checkpoint_path,
-            map_location="cpu",
-            **weights_only_kwarg,
-        )
+        state_dict = torch.load(pytorch_checkpoint_path, map_location="cpu", weights_only=True)
         pt_model = pt_model_class.from_pretrained(
             pretrained_model_name_or_path=None, config=config, state_dict=state_dict
         )
diff --git a/src/transformers/data/datasets/squad.py b/src/transformers/data/datasets/squad.py
index d81217d818a..7546d7b49ed 100644
--- a/src/transformers/data/datasets/squad.py
+++ b/src/transformers/data/datasets/squad.py
@@ -148,7 +148,7 @@ class SquadDataset(Dataset):
         with FileLock(lock_path):
             if os.path.exists(cached_features_file) and not args.overwrite_cache:
                 start = time.time()
-                self.old_features = torch.load(cached_features_file)
+                self.old_features = torch.load(cached_features_file, weights_only=True)
 
                 # Legacy cache files have only features, while new cache files
                 # will have dataset and examples also.
diff --git a/src/transformers/modeling_flax_pytorch_utils.py b/src/transformers/modeling_flax_pytorch_utils.py
index 8fbba8a1651..07285065772 100644
--- a/src/transformers/modeling_flax_pytorch_utils.py
+++ b/src/transformers/modeling_flax_pytorch_utils.py
@@ -71,8 +71,7 @@ def load_pytorch_checkpoint_in_flax_state_dict(
                 )
                 raise
 
-            weights_only_kwarg = {"weights_only": True}
-            pt_state_dict = torch.load(pt_path, map_location="cpu", **weights_only_kwarg)
+            pt_state_dict = torch.load(pt_path, map_location="cpu", weights_only=True)
             logger.info(f"PyTorch checkpoint contains {sum(t.numel() for t in pt_state_dict.values()):,} parameters.")
 
         flax_state_dict = convert_pytorch_state_dict_to_flax(pt_state_dict, flax_model)
@@ -248,8 +247,7 @@ def convert_pytorch_sharded_state_dict_to_flax(shard_filenames, flax_model):
     flax_state_dict = {}
     for shard_file in shard_filenames:
         # load using msgpack utils
-        weights_only_kwarg = {"weights_only": True}
-        pt_state_dict = torch.load(shard_file, **weights_only_kwarg)
+        pt_state_dict = torch.load(shard_file, weights_only=True)
         weight_dtypes = {k: v.dtype for k, v in pt_state_dict.items()}
         pt_state_dict = {
             k: v.numpy() if v.dtype != torch.bfloat16 else v.float().numpy() for k, v in pt_state_dict.items()
diff --git a/src/transformers/modeling_tf_pytorch_utils.py b/src/transformers/modeling_tf_pytorch_utils.py
index 416db528880..84a6ddaebcc 100644
--- a/src/transformers/modeling_tf_pytorch_utils.py
+++ b/src/transformers/modeling_tf_pytorch_utils.py
@@ -198,8 +198,7 @@ def load_pytorch_checkpoint_in_tf2_model(
         if pt_path.endswith(".safetensors"):
             state_dict = safe_load_file(pt_path)
         else:
-            weights_only_kwarg = {"weights_only": True}
-            state_dict = torch.load(pt_path, map_location="cpu", **weights_only_kwarg)
+            state_dict = torch.load(pt_path, map_location="cpu", weights_only=True)
 
         pt_state_dict.update(state_dict)
 
diff --git a/src/transformers/modeling_utils.py b/src/transformers/modeling_utils.py
index abf8dec55cc..3c4526eb1e6 100644
--- a/src/transformers/modeling_utils.py
+++ b/src/transformers/modeling_utils.py
@@ -504,8 +504,7 @@ def load_sharded_checkpoint(model, folder, strict=True, prefer_safe=True):
             error_message += f"\nMissing key(s): {str_unexpected_keys}."
         raise RuntimeError(error_message)
 
-    weights_only_kwarg = {"weights_only": True}
-    loader = safe_load_file if load_safe else partial(torch.load, map_location="cpu", **weights_only_kwarg)
+    loader = safe_load_file if load_safe else partial(torch.load, map_location="cpu", weights_only=True)
 
     for shard_file in shard_files:
         state_dict = loader(os.path.join(folder, shard_file))
@@ -598,11 +597,10 @@ def load_state_dict(
             and is_zipfile(checkpoint_file)
         ):
             extra_args = {"mmap": True}
-        weights_only_kwarg = {"weights_only": weights_only}
         return torch.load(
             checkpoint_file,
             map_location=map_location,
-            **weights_only_kwarg,
+            weights_only=weights_only,
             **extra_args,
         )
     except Exception as e:
@@ -1216,7 +1214,7 @@ def _get_torch_dtype(
     weights_only: bool,
 ) -> Tuple[PretrainedConfig, Optional[torch.dtype], Optional[torch.dtype]]:
     """Find the correct `torch_dtype` to use based on provided arguments. Also update the `config` based on the
-    infered dtype. We do the following:
+    inferred dtype. We do the following:
     1. If torch_dtype is not None, we use that dtype
     2. If torch_dtype is "auto", we auto-detect dtype from the loaded state_dict, by checking its first
         weights entry that is of a floating type - we assume all floating dtype weights are of the same dtype
diff --git a/src/transformers/models/data2vec/convert_data2vec_audio_original_pytorch_checkpoint_to_pytorch.py b/src/transformers/models/data2vec/convert_data2vec_audio_original_pytorch_checkpoint_to_pytorch.py
index 5339f1671b0..4ecc3335514 100644
--- a/src/transformers/models/data2vec/convert_data2vec_audio_original_pytorch_checkpoint_to_pytorch.py
+++ b/src/transformers/models/data2vec/convert_data2vec_audio_original_pytorch_checkpoint_to_pytorch.py
@@ -207,7 +207,7 @@ def convert_wav2vec2_checkpoint(
         hf_wav2vec = Data2VecAudioModel(config)
         data2vec_checkpoint_dir = os.path.dirname(checkpoint_path)
 
-        state_dict = torch.load(checkpoint_path)
+        state_dict = torch.load(checkpoint_path, weights_only=True)
         state_dict["model"]["final_proj.weight"] = state_dict["model"].pop("final_proj.0.weight")
         state_dict["model"]["final_proj.bias"] = state_dict["model"].pop("final_proj.0.bias")
         converted_ckpt = os.path.join(data2vec_checkpoint_dir, "converted.pt")
diff --git a/src/transformers/models/phi/convert_phi_weights_to_hf.py b/src/transformers/models/phi/convert_phi_weights_to_hf.py
index 69ef4c5919e..09742336611 100644
--- a/src/transformers/models/phi/convert_phi_weights_to_hf.py
+++ b/src/transformers/models/phi/convert_phi_weights_to_hf.py
@@ -121,7 +121,7 @@ def convert_phi_weights(
             if model_path.endswith("safetensors"):
                 loaded_weights = safetensors.torch.load_file(model_path, device=device)
             else:
-                loaded_weights = torch.load(model_path, map_location=device)
+                loaded_weights = torch.load(model_path, map_location=device, weights_only=True)
             model_checkpoint.update(**loaded_weights)
 
         model_type = model_name.split("/")[1]  # phi-1 or phi-1_5 or phi-2
diff --git a/src/transformers/models/wav2vec2/modeling_wav2vec2.py b/src/transformers/models/wav2vec2/modeling_wav2vec2.py
index f9043eba0eb..b7b37e47a67 100755
--- a/src/transformers/models/wav2vec2/modeling_wav2vec2.py
+++ b/src/transformers/models/wav2vec2/modeling_wav2vec2.py
@@ -1589,11 +1589,10 @@ class Wav2Vec2PreTrainedModel(PreTrainedModel):
                     cache_dir=cache_dir,
                 )
 
-                weights_only_kwarg = {"weights_only": True}
                 state_dict = torch.load(
                     weight_path,
                     map_location="cpu",
-                    **weights_only_kwarg,
+                    weights_only=True,
                 )
 
             except EnvironmentError:
diff --git a/src/transformers/trainer.py b/src/transformers/trainer.py
index 17ad6145045..db8a5615cc0 100755
--- a/src/transformers/trainer.py
+++ b/src/transformers/trainer.py
@@ -2820,7 +2820,6 @@ class Trainer:
                 )
 
         if os.path.isfile(weights_file) or os.path.isfile(safe_weights_file) or is_fsdp_ckpt:
-            weights_only_kwarg = {"weights_only": True}
             # If the model is on the GPU, it still works!
             if is_sagemaker_mp_enabled():
                 if os.path.isfile(os.path.join(resume_from_checkpoint, "user_content.pt")):
@@ -2836,11 +2835,7 @@ class Trainer:
                         logger.warning(
                             "Enabling FP16 and loading from smp < 1.10 checkpoint together is not supported."
                         )
-                    state_dict = torch.load(
-                        weights_file,
-                        map_location="cpu",
-                        **weights_only_kwarg,
-                    )
+                    state_dict = torch.load(weights_file, map_location="cpu", weights_only=True)
                     # Required for smp to not auto-translate state_dict from hf to smp (is already smp).
                     state_dict["_smp_is_partial"] = False
                     load_result = model.load_state_dict(state_dict, strict=True)
@@ -2859,11 +2854,7 @@ class Trainer:
                 if self.args.save_safetensors and os.path.isfile(safe_weights_file):
                     state_dict = safetensors.torch.load_file(safe_weights_file, device="cpu")
                 else:
-                    state_dict = torch.load(
-                        weights_file,
-                        map_location="cpu",
-                        **weights_only_kwarg,
-                    )
+                    state_dict = torch.load(weights_file, map_location="cpu", weights_only=True)
 
                 # workaround for FSDP bug https://github.com/pytorch/pytorch/issues/82963
                 # which takes *args instead of **kwargs
@@ -2941,7 +2932,6 @@ class Trainer:
             or os.path.exists(best_safe_adapter_model_path)
         ):
             has_been_loaded = True
-            weights_only_kwarg = {"weights_only": True}
             if is_sagemaker_mp_enabled():
                 if os.path.isfile(os.path.join(self.state.best_model_checkpoint, "user_content.pt")):
                     # If the 'user_content.pt' file exists, load with the new smp api.
@@ -2958,11 +2948,7 @@ class Trainer:
                     if self.args.save_safetensors and os.path.isfile(best_safe_model_path):
                         state_dict = safetensors.torch.load_file(best_safe_model_path, device="cpu")
                     else:
-                        state_dict = torch.load(
-                            best_model_path,
-                            map_location="cpu",
-                            **weights_only_kwarg,
-                        )
+                        state_dict = torch.load(best_model_path, map_location="cpu", weights_only=True)
 
                     state_dict["_smp_is_partial"] = False
                     load_result = model.load_state_dict(state_dict, strict=True)
@@ -3017,11 +3003,7 @@ class Trainer:
                     if self.args.save_safetensors and os.path.isfile(best_safe_model_path):
                         state_dict = safetensors.torch.load_file(best_safe_model_path, device="cpu")
                     else:
-                        state_dict = torch.load(
-                            best_model_path,
-                            map_location="cpu",
-                            **weights_only_kwarg,
-                        )
+                        state_dict = torch.load(best_model_path, map_location="cpu", weights_only=True)
 
                     # If the model is on the GPU, it still works!
                     # workaround for FSDP bug https://github.com/pytorch/pytorch/issues/82963
@@ -3142,7 +3124,7 @@ class Trainer:
                 return
 
         with safe_globals():
-            checkpoint_rng_state = torch.load(rng_file)
+            checkpoint_rng_state = torch.load(rng_file, weights_only=True)
         random.setstate(checkpoint_rng_state["python"])
         np.random.set_state(checkpoint_rng_state["numpy"])
         torch.random.set_rng_state(checkpoint_rng_state["cpu"])
@@ -3375,7 +3357,9 @@ class Trainer:
             # deepspeed loads optimizer/lr_scheduler together with the model in deepspeed_init
             if not isinstance(self.lr_scheduler, DeepSpeedSchedulerWrapper):
                 with warnings.catch_warnings(record=True) as caught_warnings:
-                    self.lr_scheduler.load_state_dict(torch.load(os.path.join(checkpoint, SCHEDULER_NAME)))
+                    self.lr_scheduler.load_state_dict(
+                        torch.load(os.path.join(checkpoint, SCHEDULER_NAME), weights_only=True)
+                    )
                 reissue_pt_warnings(caught_warnings)
             return
 
@@ -3410,13 +3394,18 @@ class Trainer:
                             checkpoint, f"rank{self.args.process_index}-of-{self.args.world_size}-{OPTIMIZER_NAME}"
                         ),
                         map_location="cpu",
+                        weights_only=True,
                     )
                     # We only need `optimizer` when resuming from checkpoint
                     optimizer_state = optimizer_state["optimizer"]
                 else:
-                    optimizer_state = torch.load(os.path.join(checkpoint, OPTIMIZER_NAME), map_location="cpu")
+                    optimizer_state = torch.load(
+                        os.path.join(checkpoint, OPTIMIZER_NAME), map_location="cpu", weights_only=True
+                    )
                 with warnings.catch_warnings(record=True) as caught_warnings:
-                    lr_scheduler_state = torch.load(os.path.join(checkpoint, SCHEDULER_NAME), map_location="cpu")
+                    lr_scheduler_state = torch.load(
+                        os.path.join(checkpoint, SCHEDULER_NAME), map_location="cpu", weights_only=True
+                    )
                 reissue_pt_warnings(caught_warnings)
 
                 xm.send_cpu_data_to_device(optimizer_state, self.args.device)
@@ -3458,10 +3447,14 @@ class Trainer:
                         )
                     else:
                         self.optimizer.load_state_dict(
-                            torch.load(os.path.join(checkpoint, OPTIMIZER_NAME), map_location=map_location)
+                            torch.load(
+                                os.path.join(checkpoint, OPTIMIZER_NAME), map_location=map_location, weights_only=True
+                            )
                         )
                 with warnings.catch_warnings(record=True) as caught_warnings:
-                    self.lr_scheduler.load_state_dict(torch.load(os.path.join(checkpoint, SCHEDULER_NAME)))
+                    self.lr_scheduler.load_state_dict(
+                        torch.load(os.path.join(checkpoint, SCHEDULER_NAME), weights_only=True)
+                    )
                 reissue_pt_warnings(caught_warnings)
 
     def _save_scaler(self, output_dir):
@@ -3496,13 +3489,17 @@ class Trainer:
             # Load in scaler states
             if is_torch_xla_available():
                 with warnings.catch_warnings(record=True) as caught_warnings:
-                    scaler_state = torch.load(os.path.join(checkpoint, SCALER_NAME), map_location="cpu")
+                    scaler_state = torch.load(
+                        os.path.join(checkpoint, SCALER_NAME), map_location="cpu", weights_only=True
+                    )
                 reissue_pt_warnings(caught_warnings)
                 xm.send_cpu_data_to_device(scaler_state, self.args.device)
                 self.accelerator.scaler.load_state_dict(scaler_state)
             else:
                 with warnings.catch_warnings(record=True) as caught_warnings:
-                    self.accelerator.scaler.load_state_dict(torch.load(os.path.join(checkpoint, SCALER_NAME)))
+                    self.accelerator.scaler.load_state_dict(
+                        torch.load(os.path.join(checkpoint, SCALER_NAME), weights_only=True)
+                    )
                 reissue_pt_warnings(caught_warnings)
 
     def _load_callback_state(self):
diff --git a/tests/models/autoformer/test_modeling_autoformer.py b/tests/models/autoformer/test_modeling_autoformer.py
index 92aa1ad4c9d..cca8f3b3ac8 100644
--- a/tests/models/autoformer/test_modeling_autoformer.py
+++ b/tests/models/autoformer/test_modeling_autoformer.py
@@ -415,7 +415,7 @@ class AutoformerModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCa
 
 def prepare_batch(filename="train-batch.pt"):
     file = hf_hub_download(repo_id="hf-internal-testing/tourism-monthly-batch", filename=filename, repo_type="dataset")
-    batch = torch.load(file, map_location=torch_device)
+    batch = torch.load(file, map_location=torch_device, weights_only=True)
     return batch
 
 
diff --git a/tests/models/idefics/test_image_processing_idefics.py b/tests/models/idefics/test_image_processing_idefics.py
index ad208881578..5b40a0393e5 100644
--- a/tests/models/idefics/test_image_processing_idefics.py
+++ b/tests/models/idefics/test_image_processing_idefics.py
@@ -28,7 +28,7 @@ if is_torch_available():
     import torch
 
 if is_torchvision_available():
-    import torchvision.transforms as transforms
+    from torchvision import transforms
 
 if is_vision_available():
     from PIL import Image
diff --git a/tests/models/informer/test_modeling_informer.py b/tests/models/informer/test_modeling_informer.py
index fb91a652004..b5e35490f6f 100644
--- a/tests/models/informer/test_modeling_informer.py
+++ b/tests/models/informer/test_modeling_informer.py
@@ -476,7 +476,7 @@ class InformerModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase
 
 def prepare_batch(filename="train-batch.pt"):
     file = hf_hub_download(repo_id="hf-internal-testing/tourism-monthly-batch", filename=filename, repo_type="dataset")
-    batch = torch.load(file, map_location=torch_device)
+    batch = torch.load(file, map_location=torch_device, weights_only=True)
     return batch
 
 
diff --git a/tests/models/llava_next/test_modeling_llava_next.py b/tests/models/llava_next/test_modeling_llava_next.py
index 8d517c644d1..a9758bd4230 100644
--- a/tests/models/llava_next/test_modeling_llava_next.py
+++ b/tests/models/llava_next/test_modeling_llava_next.py
@@ -408,7 +408,7 @@ class LlavaNextForConditionalGenerationIntegrationTest(unittest.TestCase):
             filename="llava_1_6_input_ids.pt",
             repo_type="dataset",
         )
-        original_input_ids = torch.load(filepath, map_location="cpu")
+        original_input_ids = torch.load(filepath, map_location="cpu", weights_only=True)
         # replace -200 by image_token_index (since we use token ID = 32000 for the image token)
         # remove image token indices because HF impl expands image tokens `image_seq_length` times
         original_input_ids = original_input_ids[original_input_ids != -200]
@@ -420,7 +420,7 @@ class LlavaNextForConditionalGenerationIntegrationTest(unittest.TestCase):
             filename="llava_1_6_pixel_values.pt",
             repo_type="dataset",
         )
-        original_pixel_values = torch.load(filepath, map_location="cpu")
+        original_pixel_values = torch.load(filepath, map_location="cpu", weights_only=True)
         assert torch.allclose(original_pixel_values, inputs.pixel_values.half())
 
         # verify generation
diff --git a/tests/models/patchtsmixer/test_modeling_patchtsmixer.py b/tests/models/patchtsmixer/test_modeling_patchtsmixer.py
index 939072d825f..2f39b9e6800 100644
--- a/tests/models/patchtsmixer/test_modeling_patchtsmixer.py
+++ b/tests/models/patchtsmixer/test_modeling_patchtsmixer.py
@@ -452,7 +452,7 @@ class PatchTSMixerModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.Test
 def prepare_batch(repo_id="ibm/patchtsmixer-etth1-test-data", file="pretrain_batch.pt"):
     # TODO: Make repo public
     file = hf_hub_download(repo_id=repo_id, filename=file, repo_type="dataset")
-    batch = torch.load(file, map_location=torch_device)
+    batch = torch.load(file, map_location=torch_device, weights_only=True)
     return batch
 
 
diff --git a/tests/models/patchtst/test_modeling_patchtst.py b/tests/models/patchtst/test_modeling_patchtst.py
index 0956386f0d3..8f48a6f904a 100644
--- a/tests/models/patchtst/test_modeling_patchtst.py
+++ b/tests/models/patchtst/test_modeling_patchtst.py
@@ -303,7 +303,7 @@ class PatchTSTModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase
 
 def prepare_batch(repo_id="hf-internal-testing/etth1-hourly-batch", file="train-batch.pt"):
     file = hf_hub_download(repo_id=repo_id, filename=file, repo_type="dataset")
-    batch = torch.load(file, map_location=torch_device)
+    batch = torch.load(file, map_location=torch_device, weights_only=True)
     return batch
 
 
diff --git a/tests/models/time_series_transformer/test_modeling_time_series_transformer.py b/tests/models/time_series_transformer/test_modeling_time_series_transformer.py
index 50abfaa765e..f9dfefe3767 100644
--- a/tests/models/time_series_transformer/test_modeling_time_series_transformer.py
+++ b/tests/models/time_series_transformer/test_modeling_time_series_transformer.py
@@ -481,7 +481,7 @@ class TimeSeriesTransformerModelTest(ModelTesterMixin, PipelineTesterMixin, unit
 
 def prepare_batch(filename="train-batch.pt"):
     file = hf_hub_download(repo_id="hf-internal-testing/tourism-monthly-batch", filename=filename, repo_type="dataset")
-    batch = torch.load(file, map_location=torch_device)
+    batch = torch.load(file, map_location=torch_device, weights_only=True)
     return batch
 
 
diff --git a/tests/models/videomae/test_modeling_videomae.py b/tests/models/videomae/test_modeling_videomae.py
index 4b1abab2067..8f1ab7de053 100644
--- a/tests/models/videomae/test_modeling_videomae.py
+++ b/tests/models/videomae/test_modeling_videomae.py
@@ -456,7 +456,7 @@ class VideoMAEModelIntegrationTest(unittest.TestCase):
 
         # add boolean mask, indicating which patches to mask
         local_path = hf_hub_download(repo_id="hf-internal-testing/bool-masked-pos", filename="bool_masked_pos.pt")
-        inputs["bool_masked_pos"] = torch.load(local_path)
+        inputs["bool_masked_pos"] = torch.load(local_path, weights_only=True)
 
         # forward pass
         with torch.no_grad():
diff --git a/tests/peft_integration/test_peft_integration.py b/tests/peft_integration/test_peft_integration.py
index 5ebb53c5be4..22282aa07f3 100644
--- a/tests/peft_integration/test_peft_integration.py
+++ b/tests/peft_integration/test_peft_integration.py
@@ -554,7 +554,7 @@ class PeftIntegrationTester(unittest.TestCase, PeftTesterMixin):
 
                 state_dict_path = hf_hub_download(peft_model_id, "adapter_model.bin")
 
-                dummy_state_dict = torch.load(state_dict_path)
+                dummy_state_dict = torch.load(state_dict_path, weights_only=True)
 
                 model.load_adapter(adapter_state_dict=dummy_state_dict, peft_config=peft_config)
                 with self.assertRaises(ValueError):
@@ -579,7 +579,7 @@ class PeftIntegrationTester(unittest.TestCase, PeftTesterMixin):
 
                 peft_config = LoraConfig()
                 state_dict_path = hf_hub_download(peft_model_id, "adapter_model.bin")
-                dummy_state_dict = torch.load(state_dict_path)
+                dummy_state_dict = torch.load(state_dict_path, weights_only=True)
 
                 # this should always work
                 model.load_adapter(
@@ -647,7 +647,7 @@ class PeftIntegrationTester(unittest.TestCase, PeftTesterMixin):
 
                 peft_config = LoraConfig()
                 state_dict_path = hf_hub_download(peft_model_id, "adapter_model.bin")
-                dummy_state_dict = torch.load(state_dict_path)
+                dummy_state_dict = torch.load(state_dict_path, weights_only=True)
 
                 # add unexpected key
                 dummy_state_dict["foobar"] = next(iter(dummy_state_dict.values()))
@@ -674,7 +674,7 @@ class PeftIntegrationTester(unittest.TestCase, PeftTesterMixin):
 
                 peft_config = LoraConfig()
                 state_dict_path = hf_hub_download(peft_model_id, "adapter_model.bin")
-                dummy_state_dict = torch.load(state_dict_path)
+                dummy_state_dict = torch.load(state_dict_path, weights_only=True)
 
                 # remove a key so that we have missing keys
                 key = next(iter(dummy_state_dict.keys()))
diff --git a/tests/trainer/test_trainer.py b/tests/trainer/test_trainer.py
index 7e157550589..f7e16926f2d 100644
--- a/tests/trainer/test_trainer.py
+++ b/tests/trainer/test_trainer.py
@@ -648,7 +648,7 @@ class TrainerIntegrationCommon:
         else:
             best_model = RegressionModel()
             if not safe_weights:
-                state_dict = torch.load(os.path.join(checkpoint, WEIGHTS_NAME))
+                state_dict = torch.load(os.path.join(checkpoint, WEIGHTS_NAME), weights_only=True)
             else:
                 state_dict = safetensors.torch.load_file(os.path.join(checkpoint, SAFE_WEIGHTS_NAME))
             best_model.load_state_dict(state_dict)
diff --git a/tests/utils/test_modeling_utils.py b/tests/utils/test_modeling_utils.py
index 96cbd4f77bb..212969ce150 100644
--- a/tests/utils/test_modeling_utils.py
+++ b/tests/utils/test_modeling_utils.py
@@ -765,7 +765,7 @@ class ModelUtilsTest(TestCasePlus):
                     # Note: pickle adds some junk so the weight of the file can end up being slightly bigger than
                     # the size asked for (since we count parameters)
                     if size >= max_size_int + 50000:
-                        state_dict = torch.load(shard_file)
+                        state_dict = torch.load(shard_file, weights_only=True)
                         self.assertEqual(len(state_dict), 1)
 
                 # Check the index and the shard files found match