fix: TorchWriter.reset() closes SummaryWriter (#5375)

next write will reopen the writers automatically so the net
effect of the change is that new files are created for the first
write after each reset.

harness/determined/tensorboard/metric_writers/pytorch.py

@@ -44,5 +44,5 @@ def add_scalar(self, name: str, value: Union[int, float, np.number], step: int)
         self.writer.add_scalar(name, value, step)
 
     def reset(self) -> None:
-        if "flush" in dir(self.writer):
-            self.writer.flush()
+        # flush AND close the writer so that the next attempt to write will create a new file
+        self.writer.close()

harness/tests/tensorboard/test_torch_writer.py

@@ -0,0 +1,22 @@
+import pathlib
+from typing import Any, Dict
+
+from _pytest import monkeypatch
+
+from determined import tensorboard
+from determined.tensorboard.metric_writers import pytorch
+
+
+def test_torch_writer(monkeypatch: monkeypatch.MonkeyPatch, tmp_path: pathlib.Path) -> None:
+    def mock_get_base_path(dummy: Dict[str, Any]) -> pathlib.Path:
+        return tmp_path
+
+    monkeypatch.setattr(tensorboard, "get_base_path", mock_get_base_path)
+    logger = pytorch.TorchWriter()
+    logger.add_scalar("foo", 7, 0)
+    logger.reset()
+    logger.add_scalar("foo", 8, 1)
+    logger.reset()
+
+    files = list(tmp_path.iterdir())
+    assert len(files) == 2