fix: convert step to int when logging

src/lightning/pytorch/CHANGELOG.md

@@ -24,7 +24,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
--
+- Fixed logger_connector has edge case where step can be a float ([#20692](https://github.com/Lightning-AI/pytorch-lightning/issues/20692))
 
 
 ---

src/lightning/pytorch/trainer/connectors/logger_connector/logger_connector.py

@@ -106,12 +106,13 @@ def log_metrics(self, metrics: _OUT_DICT, step: Optional[int] = None) -> None:
         scalar_metrics = convert_tensors_to_scalars(metrics)
 
         if step is None:
-            step = scalar_metrics.pop("step", None)
-
-        if step is None:
-            # added metrics for convenience
-            scalar_metrics.setdefault("epoch", self.trainer.current_epoch)
-            step = self.trainer.fit_loop.epoch_loop._batches_that_stepped
+            step_metric = scalar_metrics.pop("step", None)
+            if step_metric is not None:
+                step = int(step_metric)
+            else:
+                # added metrics for convenience
+                scalar_metrics.setdefault("epoch", self.trainer.current_epoch)
+                step = self.trainer.fit_loop.epoch_loop._batches_that_stepped
 
         # log actual metrics
         for logger in self.trainer.loggers:

tests/tests_pytorch/trainer/connectors/test_logger_connector.py

@@ -0,0 +1,61 @@
+from unittest.mock import MagicMock, patch
+
+from lightning.pytorch import Trainer
+from lightning.pytorch.loggers import Logger
+from lightning.pytorch.trainer.connectors.logger_connector import _LoggerConnector
+
+
+@patch("lightning.pytorch.trainer.connectors.logger_connector.logger_connector.convert_tensors_to_scalars")
+def test_uses_provided_step(mock_convert):
+    """Test that the LoggerConnector uses explicitly provided step to log metrics."""
+
+    trainer = MagicMock(spec=Trainer)
+    trainer.loggers = [logger := MagicMock(spec=Logger)]
+    connector = _LoggerConnector(trainer)
+    mock_convert.return_value.pop.return_value = step = 42
+
+    connector.log_metrics((metrics := {"some_metric": 123}), step=step)
+
+    assert connector._logged_metrics == metrics
+    mock_convert.assert_called_once_with(metrics)
+    logger.log_metrics.assert_called_once_with(metrics=mock_convert.return_value, step=step)
+    logger.save.assert_called_once_with()
+
+
+@patch("lightning.pytorch.trainer.connectors.logger_connector.logger_connector.convert_tensors_to_scalars")
+def test_uses_step_metric(mock_convert):
+    """Test that the LoggerConnector uses explicitly provided step metric to log metrics."""
+
+    trainer = MagicMock(spec=Trainer)
+    trainer.loggers = [logger := MagicMock(spec=Logger)]
+    connector = _LoggerConnector(trainer)
+    mock_convert.return_value.pop.return_value = step = 42.0
+
+    metrics = {"some_metric": 123}
+    connector.log_metrics(logged_metrics := {**metrics, "step": step})
+
+    assert connector._logged_metrics == logged_metrics
+    mock_convert.assert_called_once_with(logged_metrics)
+    logger.log_metrics.assert_called_once_with(metrics=mock_convert.return_value, step=int(step))
+    logger.save.assert_called_once_with()
+
+
+@patch("lightning.pytorch.trainer.connectors.logger_connector.logger_connector.convert_tensors_to_scalars")
+def test_uses_batches_that_stepped(mock_convert):
+    """Test that the LoggerConnector uses implicitly provided batches_that_stepped to log metrics."""
+
+    trainer = MagicMock(spec=Trainer)
+    trainer.fit_loop = MagicMock()
+    trainer.loggers = [logger := MagicMock(spec=Logger)]
+    connector = _LoggerConnector(trainer)
+    mock_convert.return_value.pop.return_value = None
+
+    connector.log_metrics(metrics := {"some_metric": 123})
+
+    assert connector._logged_metrics == metrics
+    mock_convert.assert_called_once_with(metrics)
+    logger.log_metrics.assert_called_once_with(
+        metrics=mock_convert.return_value, step=trainer.fit_loop.epoch_loop._batches_that_stepped
+    )
+    logger.save.assert_called_once_with()
+    mock_convert.return_value.setdefault.assert_called_once_with("epoch", trainer.current_epoch)