The 3090 Baguette full-grid launch did not retire, while the same update
with half as many key CTAs per launch retired all its update fences on
2026-09-28. This bound is provisional until the complete two-part update
and speed comparison qualify; it belongs to this realization, not to the
architecture-neutral grouped decoder.
1133def streamingAttentionKeyContiguousBlocksPerLaunch : Nat = 64The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.