tune-ci-thresholds

To calibrate CI thresholds and distinguish model regressions from host-level noise, this skill performs repeated, statistically validated observations using MAD and gap analysis to automatically reject destructive rounds contaminated by cold caches or contention.

To stabilize flaky CI performance tests across heterogeneous GPU environments, this skill generates a worst-of-N threshold report that accounts for hardware variability, ensuring your CI assertions remain stable and meaningful.

You need statistical validation for CI threshold calibration when a new model version causes CI tests to flake, ensuring that reported metrics represent genuine operational reliability rather than transient host-level noise.

Yes, strict environment provenance tracking is required, as this skill ensures every calibration run is tied to a specific git commit and environment fingerprint to prevent the use of stale or incompatible data.

CI threshold calibration with sglang currently supports ASR, TTS, and Qwen3-Omni models, validating sample-scope metrics across assigned GPU groups to satisfy operational reliability requirements.

Yes, you need uv and torch installed along with sglang, as these dependencies manage the calibration lifecycle, destructive round rejection, and strict sample-scope validation across heterogeneous GPU environments.