def bu4

bendtorch@1.3.0.0 · main.bend:2819

import Base

2819# the backward pass's unit: one sequence's group j, as fslices' (Q, K, V, and dO's columns of the g heads), with
2820# dQ, dK and dV of its own
2821def bu4(+t: Nat, +gw: Nat, +hd: Nat, qq: Array<F32>, kk: Array<F32>, vv: Array<F32>, xs: List<&1, BG>, q: Array<F32>, k: Array<F32>, v: Array<F32>, r: A2) -> BL:
2822  match r:
2823    case A2{dd, d}:
2824      BL{qq, kk, vv, dd, BG{q, k, v, d, zeros_a(Nat.mul(t, gw)), zeros_a(Nat.mul(t, hd)), zeros_a(Nat.mul(t, hd))} <> xs}