Nightly 2026-08-06: flash attention measured and rejected, fallback chat renderer, and the f16 numbers were f32 - #17

Merged

physics515 merged 5 commits into

mainfrom

mummu-nightly-2026-08-06

Aug 6, 2026

Commits on Aug 6, 2026

deps: cargo update (8 crates) - budgets and gates hold

Justin Icenhour
and
claude
committed
35da9ea
bench: flash attention measured and rejected; long-prefill row lands

Justin Icenhour
and
claude
committed
7d06519
template: render a checkpoint's own chat_template (feature jinja-template)

Justin Icenhour
and
claude
committed
f75c2ec
docs: fold 2026-08-06 research - graph capture is the dispatch lever

Justin Icenhour
and
claude
committed
28c5d7e
bench: the old f16 rows were f32 - bisected, retracted, gated

Justin Icenhour
and
claude
committed
aeaee7a