spent way too long doing this properly so you don't have to: specdec benchmark of nemotron 3.5 lightning 30b on one gx10
no-spec vs mtp vs dspark vs dflash. preregistered, frozen protocol, quality gates, bootstrap cis, the whole thing
dspark won by far: +38.9% median e2e vs no-spec, and +47% at short context single stream
push context to 128k or concurrency to 16 and everything starts to collapse
repo w/ everything: