fix: AdvancedProfiler no longer crashes when stop() is called after teardown

src/lightning/pytorch/profilers/advanced.py

@@ -82,7 +82,7 @@ def start(self, action_name: str) -> None:
     def stop(self, action_name: str) -> None:
         pr = self.profiled_actions.get(action_name)
         if pr is None:
-            raise ValueError(f"Attempting to stop recording an action ({action_name}) which was never started.")
+            return
         pr.disable()
 
     def _dump_stats(self, action_name: str, profile: cProfile.Profile) -> None:
@@ -116,6 +116,8 @@ def summary(self) -> str:
     @override
     def teardown(self, stage: Optional[str]) -> None:
         super().teardown(stage=stage)
+        for pr in self.profiled_actions.values():
+            pr.disable()
         self.profiled_actions.clear()
 
     def __reduce__(self) -> tuple:

tests/tests_pytorch/profilers/test_profiler.py

@@ -375,13 +375,33 @@ def test_advanced_profiler_dump_states_sanitizes_filename(tmp_path, char):
 def test_advanced_profiler_value_errors(advanced_profiler):
     """Ensure errors are raised where expected."""
     action = "test"
-    with pytest.raises(ValueError, match="Attempting to stop recording*"):
-        advanced_profiler.stop(action)
-
+    # Starting an action already started does not raise an error since defaultdict handles it silently
     advanced_profiler.start(action)
     advanced_profiler.stop(action)
 
 
+def test_advanced_profiler_stop_on_never_started_action(advanced_profiler):
+    """Ensure AdvancedProfiler.stop() does not crash when stopping an action that was never started.
+
+    Regression test for https://github.com/Lightning-AI/pytorch-lightning/issues/9136
+
+    """
+    # Should not raise ValueError
+    advanced_profiler.stop("never_started_action")
+
+
+def test_advanced_profiler_teardown_with_active_action(advanced_profiler):
+    """Ensure AdvancedProfiler.stop() does not crash after teardown clears profiled_actions.
+
+    Regression test for https://github.com/Lightning-AI/pytorch-lightning/issues/9136
+
+    """
+    with advanced_profiler.profile("action"):
+        # teardown clears the recorded actions, mimicking describe()
+        advanced_profiler.teardown(stage=None)
+    # Exiting the 'with' block will call stop("action") which should not raise an error
+
+
 def test_advanced_profiler_deepcopy(advanced_profiler):
     advanced_profiler.describe()
     assert deepcopy(advanced_profiler)