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 918–922

streamingAttentionForwardSharedPathSM86

Full file
What SM86.Scoreboard can follow of it: the straight-line pass reads one body after the other, and the softmax's in-place EX2 (no read barrier) leaves each score register's read live, so the second body's first write of the scores is refused there -- the register kernel has the same pair across its back edge, which the pass never follows. Each path from the prologue through one body to the epilogue is checked whole.
918def streamingAttentionForwardSharedPathSM86 =
919  (lambda unrestricted body : (family SM86Program) . (lambda unrestricted seq : Nat .
920    (lambda unrestricted heads : Nat . (lambda unrestricted keyValueHeads : Nat .
921      (sfPrologue 1 seq heads keyValueHeads
922        (sm86ProgramAppend body (sfEpilogue seq heads)))))))

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.