def lf_b3

bendtorch-gpu@1.3.0.0 · main.bend:900

import Base
import bend-ml-tensor-gpu@0.1.5.0/main.bend as TG

900# x1 = q, x2 = k, x3 = dO, x4 = G; P read from the first output (unchanged), dQ, dK and dV (over V) written
901def lf_b3(+t: U32, +l: U32, +c: EAC, x1: Array<F32>, x2: Array<F32>, x3: Array<F32>, x4: Array<F32>, o: TG.O4) -> TG.O4:
902  match o:
903    case TG.O4{p, dq, dk, v}: b3_o(b3_loop(TG.share(t, eac_n(c), l), t, l, c, S8{x1, x2, x3, x4, p, dq, dk, v}))