~/jendhub

def lf_tkv

bendtorch-gpu@1.3.0.0 · main.bend:736

import bendtorch-gpu@1.3.0.0/main.bend as BendtorchGpu
import Base
import bend-ml-tensor-gpu@0.1.5.0/main.bend as TG

736# x1 = k, x2 = v; outputs kᵀ and vᵀ (written), dQ (made here, written by backward 3)
737def lf_tkv(+t: U32, +l: U32, +c: EAC, x1: Array<F32>, x2: Array<F32>, x3: Array<F32>, x4: Array<F32>, o: TG.O3) -> TG.O3:
738  match c o:
739    case EAC{+n, +tt, +t32, +h, +g, +gn, +hd, +hd32, +w, +kvw, +sc, +nw, +nk, +tr} TG.O3{kt, vt, dq}: tkv_o(dq, tkv_loop(TG.share(t, n, l), t, l, nk, t32, U32.div(h, g), hd32, kvw, S4{x1, x2, kt, vt}))

Uses (10)

Used by (2)