Pull requests / #1687

#1687 docs: STRATA_PREFILL_CPU_SHARE=auto is not repeatable run to run (#1684)

open · @nekomario28 · 0 commentaires · Sur GitHub

Documentation

Description

Part of #1684 — the docs half of the ask ("or could the docs say that a fixed value repeats and `auto` does not?").

DETAILS.md explains that `auto` picks the share from wall-clock timings and that the output differs in the last bits from a run without it — but not that the same request can therefore answer differently **from one run to the next**. The reporter's use is exact-repeat checks and conversation-resume checks; they measured a resumed read differing in 31 of 1,536 teacher-forced argmax tokens, and a fixed `0.4` repeating 4/4 where `auto` repeated 0/4.

One sentence: `auto`'s run-to-run nondeterminism is now stated where the output caveat lives, with the fixed-share / `0` escape named. No code changes; nothing claimed beyond the reporter's measurements (attributed to #1684).

Sur le site

Liens install, modèles, releases.