def ab3

bendtorch-gpu@1.3.0.0 · main.bend:943

import Base
import bend-ml-tensor-array@0.1.18.0/main.bend as TA
import bend-ml-tensor-gpu@0.1.5.0/main.bend as TG
import bendtorch@1.3.0.0/main.bend as T

943# backward 3 writes dK over kᵀ and dV over v (neither read any more)
944def ab3(+n: Nat, +w: Nat, +kvw: Nat, +b: Nat, +t: Nat, +kv: Nat, +g: Nat, +hd: Nat, +sc: F32, q: Array<F32>, k: Array<F32>, v: Array<F32>, d: Array<F32>, dq: Array<F32>, kt: Array<F32>, r: TG.R2) -> T.Bwd<Unit, T.Two<TA.Mat<n, w>, T.Two<TA.Mat<n, kvw>, TA.Mat<n, kvw>>>>:
945  match r:
946    case TG.R2{x1, x2, x3, x4, TG.O2{p, gb}}: ab4(n, w, kvw, TG.run4(~EAC, ~lf_b3, TG.depth(Nat.add(Nat.mul(n, w), Nat.mul(2n, Nat.mul(n, kvw)))), eac(Nat.add(Nat.mul(n, w), Nat.mul(2n, Nat.mul(n, kvw))), n, w, kvw, t, kv, g, hd, sc), q, k, d, gb, p, dq, kt, v))