def FullAttention

bendtorch@1.3.0.0 · main.bend:3088

import Base
import bend-ml-tensor-array@0.1.18.0/main.bend as TA

3088# the same layer as Attention with every key visible; Attention itself is untouched, so the causal laws and the bits
3089# of every existing model stay as they are
3090def FullAttention(+n: Nat, +w: Nat, +kvw: Nat, +b: Nat, +t: Nat, +kv: Nat, +g: Nat, +hd: Nat, +sc: F32, -e1: {n == Nat.mul(b, t) : Nat}, -e2: {w == Nat.mul(Nat.mul(kv, g), hd) : Nat}, -e3: {kvw == Nat.mul(kv, hd) : Nat}) -> Layer<Unit, Two<TA.Mat<n, w>, Two<TA.Mat<n, kvw>, TA.Mat<n, kvw>>>, TA.Mat<n, w>, AttnS>:
3091  Layer{p => x => attn_fwd_full(n, w, kvw, AttnDims{b, t, kv, g, hd, sc}, x), p => s => dy => g2 => attn_bwd_full(n, w, kvw, AttnDims{b, t, kv, g, hd, sc}, s, dy), op => p => g2 => st => no_step(Unit, op, p, g2, st), u => Unit{}, p => PL{p, Nil{}}, p => xs => off => PA{p, xs, off}, p => xs => off => PA{p, xs, off}, u => 0}