Nightly 2026-08-09: f16 becomes a verified path - and immediately kills flash-attention prefill - #18

Merged

physics515 merged 7 commits into

mainfrom

mummu-nightly-2026-08-09

Aug 9, 2026

Commits on Aug 9, 2026

deps: refresh the lockfile - 24 patch/minor bumps, every gate green

Justin Icenhour
and
claude
committed
f0ed82e
perf: the f16 cold-start tax is one 32-token burst - measured, then paid off the critical path

Justin Icenhour
and
claude
committed
08b9ed8
parity: f16 is now a verified path, not just a live one - both legs pass first run

Justin Icenhour
and
claude
committed
4c671f1
docs: flash attention for f16 prefill is rejected on CORRECTNESS - the new parity leg caught it

Justin Icenhour
and
claude
committed
d8d6cab
tune: give consumers a way to throw away a bad autotune cache

Justin Icenhour
and
claude
committed
c92b75f
docs: fold 2026-08-09 research - upstream is already answering the autotune-cache hazard

Justin Icenhour
and
claude
committed
71e58d8
plan: pick the highest float precision that actually fits one adapter

Justin Icenhour
and
claude
committed
8b36856