the streaming attention's: each row's base-2
log-sum-exp and the backward's row terms ([head][T] binary32), and V in
its heads' layout for the backward ([head][T][64] half)
468def cgLogSumExpBytes : Nat =
469 (naturalMultiply 4 (naturalMultiply cgHeads cgSeq))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.