fix: use del instead of explicit shutdown in CombinedLoader.reset()

src/lightning/pytorch/CHANGELOG.md

@@ -24,6 +24,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
+- Fixed non-zero process exits in `CombinedLoader.reset()` with large tensors and persistent spawned workers by avoiding explicit `_shutdown_workers()` calls and relying on iterator cleanup via `del` [#21708](https://github.com/Lightning-AI/pytorch-lightning/issues/21708)
+
 - Fixed `SIGTERMException` producing a zero exit code instead of 143 (128 + SIGTERM) ([#21623](https://github.com/Lightning-AI/pytorch-lightning/issues/21623))
 
 ---

src/lightning/pytorch/utilities/combined_loader.py

@@ -397,7 +397,7 @@ def _load_state_dicts(self, states: list[dict[str, Any]]) -> None:
 def _shutdown_workers_and_reset_iterator(dataloader: object) -> None:
     if hasattr(dataloader, "_iterator"):
         if isinstance(dataloader._iterator, _MultiProcessingDataLoaderIter):
-            dataloader._iterator._shutdown_workers()
+            del dataloader._iterator
         dataloader._iterator = None
 
 

tests/tests_pytorch/loops/test_loops.py

@@ -1111,9 +1111,9 @@ def __init__(self, *args, dataloader, **kwargs):
             super().__init__(*args, **kwargs)
             self.dataloader = dataloader
 
-        def _shutdown_workers(self):
+        def __del__(self):
             self.dataloader.shutdown_workers_epochs.append(trainer.current_epoch)
-            super()._shutdown_workers()
+            super().__del__()
 
     class TestDataLoader(DataLoader):
         def __init__(self, *args, **kwargs):
@@ -1137,8 +1137,8 @@ def _get_iterator(self):
         trainer.fit(model, train_dataloader, val_dataloader)
 
     if persistent_workers:
-        # workers get created and persist until the teardown in the final epoch
-        expected = [trainer.current_epoch, trainer.current_epoch]  # once epoch end, once on teardown
+        # workers persist across epochs and are shut down exactly once via __del__.
+        expected = [trainer.current_epoch]
     elif should_fail:
         expected = [
             # <-- iter() on epoch 0, workers get created
@@ -1155,8 +1155,7 @@ def _get_iterator(self):
     assert train_dataloader.shutdown_workers_epochs == expected
 
     if persistent_workers:
-        # workers get created and persist until the teardown in the final epoch
-        expected = [trainer.current_epoch, trainer.current_epoch]  # once epoch end, once on teardown
+        expected = [trainer.current_epoch]
     elif should_fail:
         expected = [
             # <-- iter() on sanity check, workers get created

tests/tests_pytorch/utilities/test_combined_loader.py

@@ -657,3 +657,27 @@ def test_load_state_dicts():
     cl._load_state_dicts([state1, state2])
     stateful1.load_state_dict.assert_called_with(state1)
     stateful2.load_state_dict.assert_called_with(state2)
+
+
+def test_combined_loader_reset_uses_del_not_shutdown_workers():
+    """Test that `combined_loader.reset()` uses `del` to reset the dataloader iterator instead of calling
+    `_shutdown_workers()` explicitly.
+
+    This is a regression test for https://github.com/Lightning-AI/pytorch-lightning/issues/21703
+
+    """
+    from torch.utils.data.dataloader import _MultiProcessingDataLoaderIter
+
+    dataloader = DataLoader(range(10), num_workers=2, persistent_workers=True, multiprocessing_context="spawn")
+    combined_loader = CombinedLoader([dataloader])
+
+    mock_iterator = Mock(spec=_MultiProcessingDataLoaderIter)
+    mock_iterator._shutdown_workers = Mock()
+    dataloader._iterator = mock_iterator
+
+    iterator_ref = dataloader._iterator
+
+    combined_loader.reset()
+
+    iterator_ref._shutdown_workers.assert_not_called()
+    assert dataloader._iterator is None