---- the prologue: T, the constants, the scalars, the count ----
`scaled` 1: the gradient factor (the first scalar, `one`) is instead the
word at argument 10's pointer -- a device-computed scale, such as the
clip factor -- loaded once (data under SB4, retired with the specials).
g = g s is then exactly the separate scaling pass's product, and that
pass over the bank is not needed.
77def adamWHalfSM86ScaleArgument : Nat = 10The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.