fn
rms_norm_backward_f64
void rms_norm_backward_f64(const double * x, const double * gamma, const double * saved_rstd, const double * g, double * dx, double * dgamma, int outer, int N)Double-precision counterpart to rms_norm_backward_f32.
Parameters
xconst double*Original forward input of shape
(outer, N).gammaconst double*Affine scale of shape
(N,).saved_rstdconst double*Per-row reciprocal RMS from the matching forward call.
gconst double*Upstream gradient of shape
(outer, N).dxdouble*Output gradient w.r.t.
x; overwritten.dgammadouble*Output gradient w.r.t.
gamma; accumulated — caller must zero before the first invocation.outer, Nstd::size_tLayout extents.