the sub-tile's 32 x 32 words of dS^T (P^T (dP^T - D) packed as dK's A
fragments: the same words the query pass packs for dQ), then the next
sub-tile's offset, 32 queries down
2756def sksStoreScoreGradients = (lambda unrestricted seq : Nat . (lambda unrestricted tail : (family SM86Program) .
2757 (sfWide sksScoreGradientPointer sksScoreGradientOffset (saArgument streamingAttentionKeyScoreGradientArgument) saWait5
2758 (saFor 8 (lambda unrestricted i : Nat . (saOp (sksScoreGradientWord seq sksPackedScoreGradient i)))
2759 (saAddImm sksScoreGradientOffset sksScoreGradientOffset
2760 (naturalSaturatingSubtract 4294967296 (sksScoreGradientSubTileBytes seq))
2761 tail)))))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.