mirror of
https://github.com/hanzoai/laya.git
synced 2026-10-30 19:33:31 +00:00
On a shared CI runner one descheduled run pushed the 4x-input ratio to 19 on Node 22 (it passed on the rerun and on Node 20). One untimed call per size keeps JIT tiering out of the measurement and five samples keep a single slow run from deciding it. The absolute ceilings above still carry the quadratic guarantee. Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>