Updated misconfiguration test to accommodate both new and ancient versions of Lightning

Author: speediedan

CHANGELOG.md

@@ -8,13 +8,17 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Added
 
-- Support for Lightning ``2.5.2`` and ``2.5.3``
+- Verified support for Lightning ``2.5.2`` and ``2.5.3``
 
 ### Fixed
 
 - Updated explicit pytorch version mapping matrix to include recent PyTorch release
 - Fixed newly failing test dependent on deprecated Lightning class attribute. Resolved [#19](https://github.com/speediedan/finetuning-scheduler/issues/19).
 
+### Changed
+
+- For the examples extra, updated minimum `datasets` version to `4.0.0` to ensure the new API (especially important removal of `trust_remote_code`) is used.
+
 ## [2.5.1] - 2025-03-27
 
 ### Added

requirements/examples.txt

@@ -1,9 +1,9 @@
-datasets
+datasets>=4.0.0  # important removal of trust_remote_code
 evaluate
 transformers>=4.18.0
 scikit-learn
 sentencepiece
 tensorboardX>=2.2
 tabulate
 psutil
-numpy<2.0  # to avoid issues with oldest supported pytorch (2.2)
+numpy  # <2.0  # to avoid issues with oldest supported pytorch (2.2)

src/fts_examples/fts_superglue.py

@@ -126,8 +126,6 @@ def __init__(
             "dataloader_kwargs": dataloader_kwargs,
             "tokenizers_parallelism": tokenizers_parallelism,
         }
-        # starting with HF Datasets v3.x, trust_remote_code must be `True` https://bit.ly/hf_datasets_trust_remote_req
-        self.trust_remote_code = True
         self.save_hyperparameters(self.init_hparams)
         self.dataloader_kwargs = {
             "num_workers": dataloader_kwargs.get("num_workers", 0),
@@ -142,12 +140,11 @@ def prepare_data(self):
         """Load the SuperGLUE dataset."""
         # N.B. PL calls prepare_data from a single process (rank 0) so do not use it to assign
         # state (e.g. self.x=y)
-        datasets.load_dataset("super_glue", self.hparams.task_name, trust_remote_code=self.trust_remote_code)
+        datasets.load_dataset("aps/super_glue", self.hparams.task_name)
 
     def setup(self, stage):
         """Setup our dataset splits for training/validation."""
-        self.dataset = datasets.load_dataset("super_glue", self.hparams.task_name,
-                                             trust_remote_code=self.trust_remote_code)
+        self.dataset = datasets.load_dataset("aps/super_glue", self.hparams.task_name)
         for split in self.dataset.keys():
             self.dataset[split] = self.dataset[split].map(
                 self._convert_to_features, batched=True, remove_columns=["label"]
@@ -203,9 +200,9 @@ def __init__(
         experiment_tag: str = "default",
         log_env_details: bool = True,
     ):
-        """In this example, this :class:`~lightning.pytorch.core.module.LightningModule` is initialized by composing
-        the ./config/fts_defaults.yaml default configuration with various scheduled fine-tuning yaml configurations
-        via the :class:`~lightning.pytorch.cli.LightningCLI` but it can be used like any other
+        """In this example, this :class:`~lightning.pytorch.core.module.LightningModule` is initialized by
+        composing the ./config/fts_defaults.yaml default configuration with various scheduled fine-tuning yaml
+        configurations via the :class:`~lightning.pytorch.cli.LightningCLI` but it can be used like any other
         :class:`~lightning.pytorch.core.module.LightningModule` as well.
 
         Args:

src/fts_examples/ipynb_src/fts_superglue_nb.py

@@ -234,7 +234,8 @@ def __init__(
         tokenizers_parallelism: bool = True,
         **dataloader_kwargs: Any,
     ):
-        r"""Initialize the ``LightningDataModule`` designed for both the RTE or BoolQ SuperGLUE Hugging Face datasets.
+        r"""Initialize the ``LightningDataModule`` designed for both the RTE or BoolQ SuperGLUE Hugging Face
+        datasets.
 
         Args:
             model_name_or_path (str):
@@ -255,8 +256,6 @@ def __init__(
         super().__init__()
         task_name = task_name if task_name in TASK_NUM_LABELS.keys() else DEFAULT_TASK
         self.text_fields = self.TASK_TEXT_FIELD_MAP[task_name]
-        # starting with HF Datasets v3.x, trust_remote_code must be `True` https://bit.ly/hf_datasets_trust_remote_req
-        self.trust_remote_code = True
         self.dataloader_kwargs = {
             "num_workers": dataloader_kwargs.get("num_workers", 0),
             "pin_memory": dataloader_kwargs.get("pin_memory", False),
@@ -271,12 +270,11 @@ def prepare_data(self):
         """Load the SuperGLUE dataset."""
         # N.B. PL calls prepare_data from a single process (rank 0) so do not use it to assign
         # state (e.g. self.x=y)
-        datasets.load_dataset("super_glue", self.hparams.task_name, trust_remote_code=self.trust_remote_code)
+        datasets.load_dataset("aps/super_glue", self.hparams.task_name)
 
     def setup(self, stage):
         """Setup our dataset splits for training/validation."""
-        self.dataset = datasets.load_dataset("super_glue", self.hparams.task_name,
-                                             trust_remote_code=self.trust_remote_code)
+        self.dataset = datasets.load_dataset("aps/super_glue", self.hparams.task_name)
         for split in self.dataset.keys():
             self.dataset[split] = self.dataset[split].map(
                 self._convert_to_features, batched=True, remove_columns=["label"]

tests/test_finetuning_scheduler_callback.py

@@ -2251,8 +2251,7 @@ class TestConnectWarn(Callback, CallbackResolverMixin):
 
     def __init__(self, *args, **kwargs) -> None:
         super().__init__(*args, **kwargs)
-        #self.callback_attrs = ("lr_finder",)
-        self.callback_attrs = ("optimal_lr",)
+        self.callback_attrs = ("_max_lr", "_min_lr", "_early_exit")
         # choosing this callback because it's simple and has an attribute to find during target callback resolution
         self.target_callback_ref = "LearningRateFinder"