#6387 [TUI] recheck saved tuning when its launch conditions change
open · @Held0fTheWelt · 0 comments · View on GitHub
Description
## Problem and result A saved tuning can look current after an engine rebuild or a change to the RAM or conversation cache policy used for the measurement. Rechecking can also reset an explicitly chosen benchmark goal. Affected presets are now marked for recheck, and the recheck keeps the selected goal. ## Related discussion The broader tuning and cache discussion is #6342. This PR addresses the accuracy of the saved tuning and its recheck. ## Approach Compare the engine identity and effective launch memory policy, including inherited RAM settings, with the saved preset. Keep the existing stale path for old presets and require a measurable baseline before updating the recorded identity. Use the explicit goal when building the recheck request, with the existing fallback when no goal is supplied. ## Validation **Passed:** - Windows, Python 3.14.4, commit `75eb812fd`. With `PYTHONDONTWRITEBYTECODE=1`, `python -m pytest -q -p no:cacheprovider terminal/tests/tuning` passed all 348 tests. - `git diff --check upstream/main...HEAD` returned 0. **Failed:** - None observed in the checks above. **Not run:** - A live benchmark with a loaded GPU model and the full Terminal test suite. ## Limits Preset checks can ask for a new measurement when older presets have no saved memory policy. This PR makes no claim that changing a preset improves speed.
Related on this site
Quick links to related pages.