feat(SAMPLE-CORE): the checkpoint's sampling defaults, and the Gumbel draw that scanned 248,320 tokens on one thread (#1984, #1985) - #1997

Merged

localai-bot merged 11 commits into

mainfrom

row/SAMPLE-CORE-GENCFG-GUMBEL

Aug 27, 2026

Commits on Aug 27, 2026

spec(SAMPLE-CORE): the checkpoint's sampling defaults, and the Gumbel draw that scans 248,320 tokens on one thread (#1984, #1985)
mudler
committed
fd284ab
feat(SAMPLE-CORE): the checkpoint's own sampling defaults reach an OpenAI request that omits them (#1985)
mudler
committed
a7bf176
perf(SAMPLE-CORE): the Gumbel draw reuses the parallel argmax instead of walking 248,320 tokens on one thread (#1984)
mudler
committed
1a4b139
docs(SAMPLE-CORE): --generation-config in the flag reference and the server's own usage line (#1985)
mudler
committed
701a78d
fix(SAMPLE-CORE): name the scratch each allocation is for, and re-anchor the spec on the renamed kernel (#1984)
mudler
committed
b76aafb
spec(SAMPLE-CORE): say why the measurement must NOT pass --temperature 0 (#1984)
mudler
committed
aa9cebe
feat(SAMPLE-CORE): the C ABI's chat entry point resolves the same checkpoint defaults the server does (#1985)
mudler
committed
bbb5c97
fix(SAMPLE-CORE): allowlist VT_FAST_RANDOM_SAMPLE beside VT_FAST_ARGMAX (#1984)
mudler
committed
94152f2
record(SAMPLE-CORE): the acceptance measurement must drop --speculative-config, because drafts bypass the sampler entirely (#2002)
mudler
committed
6c37ab6
spec(SAMPLE-CORE): drop --speculative-config from every measurement arm, and record why (#2002)
mudler
committed
d0b3f7c
record(SAMPLE-CORE): the benchmark checkpoint ships the sampling defaults, so landing this re-baselines our arm (#1985, #1984)
mudler
committed
ab4c382