loss scaling: the backward runs on the loss times this, and AdamW's
epsilon is scaled by the same factor, so the update is the unscaled
loss's: S m / (S sqrt v + S eps). bfloat16's range needs none (1): the
factor binary16 needed (8,192) is what its range lacked.
66def coppeliusLossScale : Nat = 1The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.