Merge branch 'bugfix_model_checkpoint_save_last_not_exists' of https://github.com/SkafteNicki/pytorch-lightning into bugfix_model_checkpoint_save_last_not_exists

Author: SkafteNicki

src/lightning/pytorch/callbacks/model_checkpoint.py

@@ -381,8 +381,9 @@ def on_train_epoch_end(self, trainer: "pl.Trainer", pl_module: "pl.LightningModu
             if self._every_n_epochs >= 1 and (trainer.current_epoch + 1) % self._every_n_epochs == 0:
                 self._save_topk_checkpoint(trainer, monitor_candidates)
             # Only save last checkpoint if a checkpoint was actually saved in this step or if save_last="link"
-            if (self._last_global_step_saved == trainer.global_step or 
-                (self.save_last == "link" and self._last_checkpoint_saved)):
+            if self._last_global_step_saved == trainer.global_step or (
+                self.save_last == "link" and self._last_checkpoint_saved
+            ):
                 self._save_last_checkpoint(trainer, monitor_candidates)
 
     @override
@@ -401,8 +402,9 @@ def on_validation_end(self, trainer: "pl.Trainer", pl_module: "pl.LightningModul
             if self._every_n_epochs >= 1 and (trainer.current_epoch + 1) % self._every_n_epochs == 0:
                 self._save_topk_checkpoint(trainer, monitor_candidates)
             # Only save last checkpoint if a checkpoint was actually saved in this step or if save_last="link"
-            if (self._last_global_step_saved == trainer.global_step or 
-                (self.save_last == "link" and self._last_checkpoint_saved)):
+            if self._last_global_step_saved == trainer.global_step or (
+                self.save_last == "link" and self._last_checkpoint_saved
+            ):
                 self._save_last_checkpoint(trainer, monitor_candidates)
 
     @override

tests/tests_pytorch/checkpointing/test_model_checkpoint.py

@@ -2128,32 +2128,32 @@ def test_save_last_without_save_on_train_epoch_and_without_val(tmp_path):
 
 def test_save_last_only_when_checkpoint_saved(tmp_path):
     """Test that save_last only creates last.ckpt when another checkpoint is actually saved."""
-    
+
     class SelectiveModel(BoringModel):
         def __init__(self):
             super().__init__()
             self.validation_step_outputs = []
-            
+
         def validation_step(self, batch, batch_idx):
             outputs = super().validation_step(batch, batch_idx)
             epoch = self.trainer.current_epoch
             loss = torch.tensor(1.0 - epoch * 0.1) if epoch % 2 == 0 else torch.tensor(1.0 + epoch * 0.1)
             outputs["val_loss"] = loss
             self.validation_step_outputs.append(outputs)
             return outputs
-            
+
         def on_validation_epoch_end(self):
             if self.validation_step_outputs:
                 avg_loss = torch.stack([x["val_loss"] for x in self.validation_step_outputs]).mean()
                 self.log("val_loss", avg_loss)
                 self.validation_step_outputs.clear()
 
     model = SelectiveModel()
-    
+
     checkpoint_callback = ModelCheckpoint(
         dirpath=tmp_path,
         filename="best-{epoch}-{val_loss:.2f}",
-        monitor="val_loss", 
+        monitor="val_loss",
         save_last=True,
         save_top_k=1,
         mode="min",
@@ -2177,4 +2177,6 @@ def on_validation_epoch_end(self):
     checkpoint_names = [f.name for f in checkpoint_files]
     assert "last.ckpt" in checkpoint_names, "last.ckpt should exist since checkpoints were saved"
     expected_files = 2  # best checkpoint + last.ckpt
-    assert len(checkpoint_files) == expected_files, f"Expected {expected_files} files, got {len(checkpoint_files)}: {checkpoint_names}"
+    assert len(checkpoint_files) == expected_files, (
+        f"Expected {expected_files} files, got {len(checkpoint_files)}: {checkpoint_names}"
+    )