add test coverage

Author: littlebullGit

src/lightning/pytorch/CHANGELOG.md

@@ -82,6 +82,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed FSDP mixed precision semantics and added user warning ([#21361](https://github.com/Lightning-AI/pytorch-lightning/pull/21361))
 
 
+- Fixed `ModelCheckpoint.file_exists` using broadcast in DDP, reducing memory usage when checking for existing checkpoints ([#19674](https://github.com/Lightning-AI/pytorch-lightning/issues/19674))
+
+
 ---
 
 ## [2.5.6] - 2025-11-05

src/lightning/pytorch/callbacks/model_checkpoint.py

@@ -999,10 +999,6 @@ def to_yaml(self, filepath: Optional[_PATH] = None) -> None:
     def file_exists(self, filepath: _PATH, trainer: "pl.Trainer") -> bool:
         """Checks if a file exists on rank 0 and broadcasts the result to all other ranks, preventing the internal
         state to diverge between ranks."""
-        # Single-process or strategies without distributed world size: no need for coordination
-        if trainer.world_size == 1:
-            return self._fs.exists(filepath)
-
         # In distributed setups, only global rank 0 touches the filesystem
         local_decision = self._fs.exists(filepath) if trainer.is_global_zero else False
         # Reduce the decision across ranks using an "any"-style reduction to decide if the file exists anywhere

tests/tests_pytorch/checkpointing/test_model_checkpoint.py

@@ -2180,3 +2180,36 @@ def on_validation_epoch_end(self):
     assert len(checkpoint_files) == expected_files, (
         f"Expected {expected_files} files, got {len(checkpoint_files)}: {checkpoint_names}"
     )
+
+
+def test_model_checkpoint_file_exists_distributed_branch(tmp_path):
+    """Ensure the distributed branch of ModelCheckpoint.file_exists uses reduce_boolean_decision."""
+
+    checkpoint = ModelCheckpoint(dirpath=tmp_path)
+    calls = []
+
+    class DummyStrategy:
+        def reduce_boolean_decision(self, decision, all=True):
+            calls.append((decision, all))
+            return decision
+
+    class DummyTrainer:
+        def __init__(self, is_global_zero: bool):
+            self.world_size = 2
+            self.is_global_zero = is_global_zero
+            self.strategy = DummyStrategy()
+
+    # global rank 0: filesystem is touched and decision=True is reduced with all=False
+    checkpoint._fs.exists = Mock(return_value=True)
+    trainer = DummyTrainer(is_global_zero=True)
+    assert checkpoint.file_exists("ignored", trainer)
+    checkpoint._fs.exists.assert_called_once_with("ignored")
+    assert calls == [(True, False)]
+
+    # non-global ranks: filesystem is not touched and local decision is False
+    calls.clear()
+    checkpoint._fs.exists = Mock(return_value=True)
+    trainer = DummyTrainer(is_global_zero=False)
+    assert not checkpoint.file_exists("ignored", trainer)
+    checkpoint._fs.exists.assert_not_called()
+    assert calls == [(False, False)]