Key tiles already run longest first (tile j reads the queries from j on);
order 1 only puts the heads fastest. The constant still selects the range.
3121def streamingAttentionKeyHeadsFirstByGridAndTileSM86 =
3122 (lambda unrestricted seq : Nat . (lambda unrestricted heads : Nat .
3123 (lambda unrestricted keyValueHeads : Nat .
3124 (lambda erased admitted : (equal Nat (streamingAttentionGroupedSM86Admitted seq heads keyValueHeads) 1) .
3125 (sm86GuardLateReads
3126 (streamingAttentionKeyGroupedOrderedUncheckedSM86 1 seq heads keyValueHeads 0 1))))))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.