AlphaGitHub
Source/Packages

Realization.Nvidia.SM86.StreamingAttentionSM86

packages/realizations/cooperative/nvidia-sm86/src/Realization/Nvidia/SM86/StreamingAttentionSM86.alpha

3,230 lines493 declarations193.3 KiBSHA-256 cc0366b3d857
Download exact fileView at revisionbcc11a4e7647d31628

def · lines 3158–3158

streamingAttentionKeyContiguousBlocksPerLaunch

Full file
Every key tile in one launch. On 2026-09-28 the 3090 full-grid launch did not retire and half the key CTAs a launch did; with the QMD register allocation and late-read fixes since, the full grid retires bit for bit (2026-10-03) and saves a launch and a wave tail a layer.
3158def streamingAttentionKeyContiguousBlocksPerLaunch : Nat = 128

The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.