Fix for issue #21110: Prevent recursive symlink creation in ModelCheckpoint

- Added a check in _link_checkpoint to compare absolute paths of filepath and linkpath
- Only create symlink if paths differ, avoiding self-linking when save_last='link' and save_top_k=-1
- Updated test to assert the fix prevents the recursive symlink bug

Author: littlebullGit

src/lightning/pytorch/callbacks/model_checkpoint.py

@@ -484,7 +484,7 @@ def _save_checkpoint(self, trainer: "pl.Trainer", filepath: str) -> None:
 
     @staticmethod
     def _link_checkpoint(trainer: "pl.Trainer", filepath: str, linkpath: str) -> None:
-        if trainer.is_global_zero:
+        if trainer.is_global_zero and os.path.abspath(filepath) != os.path.abspath(linkpath):
             if os.path.islink(linkpath) or os.path.isfile(linkpath):
                 os.remove(linkpath)
             elif os.path.isdir(linkpath):

tests/tests_pytorch/callbacks/test_model_checkpoint_additional_cases.py

@@ -206,3 +206,61 @@ def test_model_checkpoint_defer_until_next_validation_when_val_every_2_epochs(tm
     expected = max(val_scores)  # last/maximum value occurs at final validation epoch
     actual = float(ckpt.best_model_score)
     assert math.isclose(actual, expected, rel_tol=0, abs_tol=1e-6)
+
+
+def test_model_checkpoint_save_last_link_symlink_bug(tmp_path):
+    """Reproduce the bug where save_last='link' and save_top_k=-1 creates a recursive symlink."""
+    import os
+
+    class RandomDataset(Dataset):
+        def __init__(self, size, length):
+            self.len = length
+            self.data = torch.randn(length, size)
+
+        def __getitem__(self, index):
+            return self.data[index]
+
+        def __len__(self):
+            return self.len
+
+    class BoringModel(LightningModule):
+        def __init__(self):
+            super().__init__()
+            self.layer = torch.nn.Linear(32, 2)
+
+        def forward(self, x):
+            return self.layer(x)
+
+        def training_step(self, batch, batch_idx):
+            loss = self(batch).sum()
+            self.log("train_loss", loss)
+            return {"loss": loss}
+
+        def validation_step(self, batch, batch_idx):
+            loss = self(batch).sum()
+            self.log("valid_loss", loss)
+
+        def test_step(self, batch, batch_idx):
+            loss = self(batch).sum()
+            self.log("test_loss", loss)
+
+        def configure_optimizers(self):
+            return torch.optim.SGD(self.layer.parameters(), lr=0.1)
+
+    trainer = Trainer(
+        default_root_dir=tmp_path,
+        max_epochs=2,
+        callbacks=[ModelCheckpoint(dirpath=tmp_path, every_n_epochs=10, save_last="link", save_top_k=-1)],
+        enable_checkpointing=True,
+        enable_model_summary=False,
+        logger=False,
+    )
+
+    model = BoringModel()
+    trainer.fit(model, train_dataloaders=DataLoader(RandomDataset(32, 64), batch_size=2))
+
+    last_ckpt = tmp_path / "last.ckpt"
+    assert last_ckpt.exists()
+    # With the fix, it should not be a symlink to itself
+    if os.path.islink(str(last_ckpt)):
+        assert os.readlink(str(last_ckpt)) != str(last_ckpt)