Skip to content

feat(cache): version-aware response cache keys + cleanup#1217

Merged
Luodian merged 2 commits intomainfrom
fix/cache-version-tracking
Mar 2, 2026
Merged

feat(cache): version-aware response cache keys + cleanup#1217
Luodian merged 2 commits intomainfrom
fix/cache-version-tracking

Conversation

@Luodian
Copy link
Contributor

@Luodian Luodian commented Mar 2, 2026

Summary

  • Cache version isolation: Response cache keys now include eval_version (git commit for dev installs, PyPI version for pip installs), preventing stale cache reuse across lmms-eval upgrades. Schema bumped v2 -> v3.
  • Version mismatch warning: On startup, warns if the DB was written by a different lmms-eval version.
  • Audit log: eval_version field added to JSONL audit records for traceability.
  • Cleanup: Removed unused .claude-plugin/marketplace.json.

Changes

File Change
lmms_eval/utils.py Added get_lmms_eval_cache_version() — git commit hash for dev, PyPI version for pip
lmms_eval/caching/response_cache.py Schema v3, eval_version in cache key / meta table / audit log / constructor
lmms_eval/evaluator.py Import + pass eval_version to ResponseCache
.claude-plugin/marketplace.json Deleted (orphaned marketplace config)

Why

After upgrading lmms-eval, cached responses from a previous version could silently be reused even if the evaluation logic changed. Including the eval version in the cache key ensures automatic invalidation — old entries simply won't match, with zero risk of stale results.

Luodian and others added 2 commits March 2, 2026 20:25
…n isolation

- Add get_lmms_eval_cache_version() utility: git commit hash for dev
  installs, PyPI version for pip installs
- Include eval_version in cache key payload (schema v2 -> v3)
- Store eval_version in SQLite meta table, warn on version mismatch
- Add eval_version to JSONL audit log records
- Remove unused .claude-plugin/marketplace.json
@Luodian Luodian merged commit f45e7ef into main Mar 2, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant