Gemma 4 26B-A4B inference in ~2 GB of RAM on any M-series MacBook
Swift apple-silicongemmagemma4gemma4-26b-a4bgpgpullmllm-inferencelocal-aimacosmetalon-device-aion-device-llmswift
import Foundation
public enum WhtRef {
public static func apply(_ x: [Float]) -> [Float] {
let d = x.count
precondition(d > 0 && (d & (d - 1)) == 0, "D must be a power of two")
if d == 1 { return x }
return walsh(x)
}
private static func walsh(_ x: [Float]) -> [Float] {
let n = x.count
if n == 1 { return x }
let half = n / 2
let lo = walsh(Array(x[0..<half]))
let hi = walsh(Array(x[half..<n]))
let invSqrt2: Float = 1.0 / Float(2.0).squareRoot()
var y = [Float](repeating: 0, count: n)
for i in 0..<half {
y[i] = (lo[i] + hi[i]) * invSqrt2
y[half + i] = (lo[i] - hi[i]) * invSqrt2
}
return y
}
}