Summary

SpikeDetection in src/lightning/fabric/utilities/spike.py stores its running-mean metric on self.running_mean (set in __init__ at line 57) and saves it via state_dict under the key "running":

# line 156
"running": self.running_mean.state_dict(),

But load_state_dict then tries to restore it via self.running, an attribute that is never defined anywhere:

# line 168 (before this PR)
self.running.load_state_dict(state_dict.pop("running"))

So any resume of a SpikeDetection callback's state crashes:

src = SpikeDetection()
dst = SpikeDetection()
dst.load_state_dict(src.state_dict())
# AttributeError: 'SpikeDetection' object has no attribute 'running'

The asymmetry with the very next line — which correctly uses self.running_mean.base_metric — confirms it's a typo for self.running_mean, not a deliberate different attribute. state_dict and load_state_dict were never round-trippable.

Fix

-        self.running.load_state_dict(state_dict.pop("running"))
+        self.running_mean.load_state_dict(state_dict.pop("running"))

Test

Adds test_spike_detection_state_dict_roundtrip next to the existing tests in tests/tests_fabric/utilities/test_spike.py. It constructs two SpikeDetection instances and round-trips state_dict / load_state_dict. Fails on master with the AttributeError above; passes with this change. Reuses the same _TORCHMETRICS_GREATER_EQUAL_1_0_0 skip the surrounding tests use.

ruff check and ruff format --check clean on both files.

Notes

No related open PRs: gh pr list --repo Lightning-AI/pytorch-lightning --state all --search "SpikeDetection" returns #21664 (an Override-decorator typing change, unrelated) and #19282 (a kwarg-typo fix merged in 2024, unrelated). No test pins the current broken behavior.