Fix LightningDataModule zero-length attribute

src/lightning/pytorch/CHANGELOG.md

@@ -6,19 +6,12 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ---
 
-## [Unreleased] - YYYY-MM-DD
 
-### Added
-
--
-
-### Changed
-
--
+### Fixed
 
-### Removed
+- Ensured ``LightningDataModule`` always exposes ``allow_zero_length_dataloader_with_multiple_devices`` so Trainer zero-length checks don't raise ``AttributeError`` when subclasses skip ``super().__init__`` ([#21358](https://github.com/Lightning-AI/pytorch-lightning/issues/21358))
 
--
+---
 
 
 ## [2.6.0] - 2025-11-28

src/lightning/pytorch/core/datamodule.py

@@ -76,6 +76,9 @@ def teardown(self):
     """
 
     name: Optional[str] = None
+    # Fallback for subclasses that don't call ``super().__init__``. The attribute normally gets initialized in
+    # ``DataHooks.__init__`` but Trainer loops expect it to exist regardless.
+    allow_zero_length_dataloader_with_multiple_devices: bool = False
     CHECKPOINT_HYPER_PARAMS_KEY = "datamodule_hyper_parameters"
     CHECKPOINT_HYPER_PARAMS_NAME = "datamodule_hparams_name"
     CHECKPOINT_HYPER_PARAMS_TYPE = "datamodule_hparams_type"

tests/tests_pytorch/core/test_datamodules.py

@@ -21,6 +21,7 @@
 
 import pytest
 import torch
+from torch.utils.data import DataLoader
 
 from lightning.pytorch import LightningDataModule, Trainer, seed_everything
 from lightning.pytorch.callbacks import ModelCheckpoint
@@ -29,6 +30,7 @@
     BoringDataModuleNoLen,
     BoringModel,
     IterableBoringDataModule,
+    RandomDataset,
 )
 from lightning.pytorch.profilers.simple import SimpleProfiler
 from lightning.pytorch.trainer.states import TrainerFn
@@ -274,6 +276,29 @@ def train_dataloader(self):
     trainer.fit(model, dm)
 
 
+def test_datamodule_allow_zero_length_attr_without_super(tmp_path):
+    class DataModuleWithoutSuper(LightningDataModule):
+        def __init__(self):
+            self.data = RandomDataset(32, 4)
+
+        def val_dataloader(self):
+            return DataLoader(self.data, batch_size=2)
+
+    dm = DataModuleWithoutSuper()
+    assert dm.allow_zero_length_dataloader_with_multiple_devices is False
+
+    model = BoringModel()
+    trainer = Trainer(
+        default_root_dir=tmp_path,
+        fast_dev_run=1,
+        enable_model_summary=False,
+        enable_checkpointing=False,
+        logger=False,
+    )
+
+    trainer.validate(model, datamodule=dm)
+
+
 class DummyDS(torch.utils.data.Dataset):
     def __getitem__(self, index):
         return 1