Skip to content

screen.md

Latest commit

 

History

History
42 lines (35 loc) · 2.17 KB

screen.md

File metadata and controls

42 lines (35 loc) · 2.17 KB
Raw

Screen result (scope addendum of #121)

Machine: Mac Studio M4 Max, macOS 26.6.2 (25G83), coremltools 9.0, laya-typed-decisions, buckets 64/96/128, ane_startup="wait", one repeat, exclusive slot with the local LLM service stopped. Raw data: raw/c-fresh-copy.json, raw/screen-move.json, raw/c-new-paths.txt, raw/c-sizes.txt, raw/c-e5rt-cache.txt, raw/screen-machine-{before,after}.txt. This is the screen, not the preregistered verdict (n = 1).

Arm ready_s warm_ready_s Preparation load
fresh-copy (C) 273.5 2.68 (W) —
move 114.9 2.65 271.1

Validity guard: passed. Both rows served the request after ready on the ANE. Every bucket's placement probe ratio was ≤ 0.8 (0.35–0.50). The move row's answers are bit-identical to the fresh-copy row's.

Screen label: screen-INCONCLUSIVE. The move took 114.9 s:

  • that is above the screen-REUSED bound, W + 10 s = 12.7 s;
  • it is below the screen-RECOMPILED bound, 0.5 × C = 136.7 s.

A path change kept about 58% of the cold cost (114.9 of 273.5 s were paid again), not all of it and not none.

Where the compile is stored (Phase C). Core ML writes its on-device ANE compile to ~/Library/Caches/<process name>/com.apple.e5rt.e5bundlecache/<macOS build>/<hash>/, not to the per-user cache directory the plan listed. That layout makes the cache per executable name and per macOS build.

  • The screen wrote about 15.6 GB there: 14 bundles of 0.7–1.4 GB each.
  • In total the cache now holds 15 GB (python) and 24 GB (python3), including entries from earlier runs.
  • New locations add bundles that nothing in laya-apple evicts.

Decisions from the addendum:

  • Following the addendum, the result is reported here, and the maintainer decides whether a repeat or Phase B is worth a slot. Nothing further ran.
  • Decision 3 is unchanged: no cross-machine cache shipping. The observed layout, keyed by macOS build and process name, makes a shipped cache build-specific.
  • Decision 4: C = 273.5 s, so the < 30 s cold-start target on a new machine is not met by prebuilt artifacts. The compile remains, and ane_startup="background" stays the way to serve during it.