Automatically optimize transformer loglikelihood scoring - #142

Merged

Qubitium merged 11 commits into

mainfrom

codex/buffer-loglikelihood-d2h

Aug 31, 2026

Commits on Aug 31, 2026

Buffer transformer loglikelihood D2H reductions
Qubitium
committed
df07188
Bump Evalution version to 0.0.17
Qubitium
committed
d966bee
Deduplicate repeated one-token loglikelihood prefixes
Qubitium
committed
bfda845
Add opt-in loglikelihood prefix KV cache
Qubitium
committed
065679d
Prewarm shared loglikelihood prefixes
Qubitium
committed
c52a5cf
Enable prefix prewarm by default
Qubitium
committed
2cbda35
Keep Evalution one patch above latest PyPI
Qubitium
committed
042b1ef
Batch prefix prewarm and release completed groups
Qubitium
committed
0c74736
Enable prefix caching across transformer loglikelihood
Qubitium
committed
bd8ba27
Harden automatic prefix cache fallback
Qubitium
committed
d13c302
Remove obsolete cache batch-size helper
Qubitium
committed
3219460