fix(serve): reply-content indent loss and float16 sampling/softcap hardening - #63

Merged

asher merged 5 commits into

mainfrom

fp16-logprob-pins

Aug 21, 2026

Commits on Aug 21, 2026

fix(dtype): widen sampled logprob math to float32; keep MSA indexer F32 under float16
asher
committed
d680ee0
tests: parametrize long-context suite over float16 and bfloat16 activations
asher
committed
c1bc8fa
fix(gemma4): compute the final logit softcap in float32 and emit float32 logits
asher
committed
0fad8ec
fix(reasoning): keep first-line content indent in non-streamed harmony/ATEM replies
asher
committed
d98b661
fix(serve): keep first-line content indent in think-tag non-streamed replies too
asher
committed
901eaf0