Reset dataloaders on failure in tuner

src/pytorch_lightning/CHANGELOG.md

@@ -141,6 +141,12 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed incorrect values after transferring data to a MPS device ([#13285](https://github.com/Lightning-AI/lightning/issues/13285))
 
 
+- Reset the dataloaders on OOM failure in batch size finder to use the last successful batch size ([#14372](https://github.com/Lightning-AI/lightning/pull/14372))
+
+
+- Fixed an issue to keep downscaling the batch size in case there hasn't been even a single successful optimal batch size with `mode="power"` ([#14372](https://github.com/Lightning-AI/lightning/pull/14372))
+
+
 - Fixed an issue to avoid the impact of sanity check on `reload_dataloaders_every_n_epochs` for validation ([#13964](https://github.com/Lightning-AI/lightning/pull/13964))
 
 

src/pytorch_lightning/tuner/batch_size_scaling.py

@@ -126,6 +126,9 @@ def _run_power_scaling(
     trainer: "pl.Trainer", model: "pl.LightningModule", new_size: int, batch_arg_name: str, max_trials: int
 ) -> int:
     """Batch scaling mode where the size is doubled at each iteration until an OOM error is encountered."""
+    # this flag is used to determine whether the previously scaled batch size, right before OOM, was a success or not
+    # if it was we exit, else we continue downscaling in case we haven't encountered a single optimal batch size
+    any_success = False
     for _ in range(max_trials):
         garbage_collection_cuda()
 
@@ -137,22 +140,28 @@ def _run_power_scaling(
             trainer.tuner._run(model)
             # Double in size
             new_size, changed = _adjust_batch_size(trainer, batch_arg_name, factor=2.0, desc="succeeded")
+
+            if not changed:
+                break
+
+            # Force the train dataloader to reset as the batch size has changed
+            trainer.reset_train_dataloader(model)
+            trainer.reset_val_dataloader(model)
+            any_success = True
         except RuntimeError as exception:
             # Only these errors should trigger an adjustment
             if is_oom_error(exception):
                 # If we fail in power mode, half the size and return
                 garbage_collection_cuda()
                 new_size, _ = _adjust_batch_size(trainer, batch_arg_name, factor=0.5, desc="failed")
-                break
+                # Force the train dataloader to reset as the batch size has changed
+                trainer.reset_train_dataloader(model)
+                trainer.reset_val_dataloader(model)
+                if any_success:
+                    break
             else:
                 raise  # some other error not memory related
 
-        if changed:
-            # Force the train dataloader to reset as the batch size has changed
-            trainer.reset_train_dataloader(model)
-            trainer.reset_val_dataloader(model)
-        else:
-            break
     return new_size
 
 
@@ -189,13 +198,13 @@ def _run_binsearch_scaling(
             else:
                 new_size, changed = _adjust_batch_size(trainer, batch_arg_name, factor=2.0, desc="succeeded")
 
-            if changed:
-                # Force the train dataloader to reset as the batch size has changed
-                trainer.reset_train_dataloader(model)
-                trainer.reset_val_dataloader(model)
-            else:
+            if not changed:
                 break
 
+            # Force the train dataloader to reset as the batch size has changed
+            trainer.reset_train_dataloader(model)
+            trainer.reset_val_dataloader(model)
+
         except RuntimeError as exception:
             # Only these errors should trigger an adjustment
             if is_oom_error(exception):
@@ -204,6 +213,11 @@ def _run_binsearch_scaling(
                 high = new_size
                 midval = (high + low) // 2
                 new_size, _ = _adjust_batch_size(trainer, batch_arg_name, value=midval, desc="failed")
+
+                # Force the train dataloader to reset as the batch size has changed
+                trainer.reset_train_dataloader(model)
+                trainer.reset_val_dataloader(model)
+
                 if high - low <= 1:
                     break
             else:

tests/tests_pytorch/tuner/test_scale_batch_size.py

@@ -319,3 +319,26 @@ def test_dataloader_reset_with_scale_batch_size(tmpdir, scale_method):
 
     assert trainer.train_dataloader.loaders.batch_size == new_batch_size
     assert trainer.val_dataloaders[0].batch_size == new_batch_size
+
+
+@pytest.mark.parametrize("scale_method, expected_batch_size", [("power", 62), ("binsearch", 100)])
+@patch("pytorch_lightning.tuner.batch_size_scaling.is_oom_error", return_value=True)
+def test_dataloader_batch_size_updated_on_failure(_, tmpdir, scale_method, expected_batch_size):
+    class CustomBatchSizeModel(BatchSizeModel):
+        def training_step(self, *_, **__):
+            if self.batch_size > 100:
+                raise RuntimeError
+
+        def train_dataloader(self):
+            return DataLoader(RandomDataset(32, 1000), batch_size=self.batch_size)
+
+    model = CustomBatchSizeModel(batch_size=16)
+    model.validation_step = None
+    model.training_epoch_end = None
+    scale_batch_size_kwargs = {"max_trials": 10, "steps_per_trial": 1, "init_val": 500, "mode": scale_method}
+
+    trainer = Trainer(default_root_dir=tmpdir, max_epochs=2, auto_scale_batch_size=True)
+    new_batch_size = trainer.tune(model, scale_batch_size_kwargs=scale_batch_size_kwargs)["scale_batch_size"]
+    assert new_batch_size == model.batch_size
+    assert new_batch_size == expected_batch_size
+    assert trainer.train_dataloader.loaders.batch_size == expected_batch_size