fast::hadamard_transform over the last axis. scale defaults to 1/sqrt(dim) (mlx's own default when unset) — pass it explicitly to match a reference that always states its normalization, per packages/inference/src/kernels/turboquant/ops.ts's fwht().
scale
fast::hadamard_transform over the last axis.
scaledefaults to 1/sqrt(dim) (mlx's own default when unset) — pass it explicitly to match a reference that always states its normalization, per packages/inference/src/kernels/turboquant/ops.ts's fwht().