log2 e, ln 2 and 1 / 256 as binary32. BigramLossGradientSM86 converts
the log2 reduction back to natural logarithms for each row's cross entropy;
inverse rows scales the gradients, not the stored row losses. The row
block must be reduced separately before reporting a mean objective. Each
constant is held to its exact value (Float32Exact).
226def bobInverseRows : Nat = 0x3b800000The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.