🚨 Fully revert atomic checkpointing 🚨 (#29370)

Fully revert atomic checkpointing

tests/trainer/test_trainer.py

@@ -84,8 +84,7 @@ from transformers.testing_utils import (
     slow,
     torch_device,
 )
-from transformers.tokenization_utils_base import PreTrainedTokenizerBase
-from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, HPSearchBackend, get_last_checkpoint
+from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, HPSearchBackend
 from transformers.training_args import OptimizerNames
 from transformers.utils import (
     SAFE_WEIGHTS_INDEX_NAME,
@@ -1406,19 +1405,6 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
             trainer.train()
             self.check_saved_checkpoints(tmpdir, 5, int(self.n_epochs * 64 / self.batch_size), False)
 
-    def test_save_checkpoints_is_atomic(self):
-        class UnsaveableTokenizer(PreTrainedTokenizerBase):
-            def save_pretrained(self, *args, **kwargs):
-                raise OSError("simulated file write error")
-
-        with tempfile.TemporaryDirectory() as tmpdir:
-            trainer = get_regression_trainer(output_dir=tmpdir, save_steps=5)
-            # Attach unsaveable tokenizer to partially fail checkpointing
-            trainer.tokenizer = UnsaveableTokenizer()
-            with self.assertRaises(OSError) as _context:
-                trainer.train()
-            assert get_last_checkpoint(tmpdir) is None
-
     @require_safetensors
     def test_safe_checkpoints(self):
         for save_safetensors in [True, False]:

tests/trainer/test_trainer_distributed.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from pathlib import Path
 from typing import Dict
 
 import numpy as np
@@ -237,20 +236,6 @@ if __name__ == "__main__":
 
         trainer.args.eval_accumulation_steps = None
 
-    # Check that saving does indeed work with temp dir rotation
-    # If this fails, will see a FileNotFoundError
-    model = RegressionModel()
-    training_args.max_steps = 1
-    opt = torch.optim.Adam(model.parameters(), lr=1e-3)
-    sched = torch.optim.lr_scheduler.LambdaLR(opt, lambda x: 1)
-    trainer = Trainer(
-        model, training_args, optimizers=(opt, sched), data_collator=DummyDataCollator(), eval_dataset=dataset
-    )
-    trainer._save_checkpoint(model=None, trial=None)
-    # Check that the temp folder does not exist
-    assert not (Path(training_args.output_dir) / "tmp-checkpoint-0").exists()
-    assert (Path(training_args.output_dir) / "checkpoint-0").exists()
-
     # Check that `dispatch_batches=False` will work on a finite iterable dataset
 
     train_dataset = FiniteIterableDataset(label_names=["labels", "extra"], length=1)