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

def · lines 1619–1619

sqsOffset

Full file
=========================================================================== The query pass with K and V staged in shared memory, as the forward above. dQ's B fragments (K^T's rows in the register kernel) are the shared K tile read transposed (LDSM.T: the same keys and columns in the same lanes), so a buffer holds K's and V's rows only and two fit two blocks. Registers: the loop no longer names R6, R7, R10..R13, R180..R183 or R194.
1619def sqsOffset = (lambda unrestricted i : Nat . (naturalAdd 10 i))

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.