the tied embedding's binary16 copy transposed (width x vocab): the final
norm's gradient is dlogits E, and the product takes its right operand as
B^T -- the vocab x width copy itself would be read as E^T
576def cgEmbeddingHalfTBytes : Nat =
577 (naturalMultiply 2 (naturalMultiply cgVocab cgWidth))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.