import numpy as np import ml_dtypes BF16 = ml_dtypes.bfloat16 def bf(x): return np.ascontiguousarray(np.asarray(x, np.float32).astype(BF16)) import math SHAPES = {"m8192": 8192, "m32768": 32768} def make_inputs(seed, shape_id): M = SHAPES[shape_id] rng = np.random.default_rng(seed) f = lambda a: np.ascontiguousarray(np.asarray(a, np.float32)) lr, b1, b2, eps, wd, step = 1e-3, 0.9, 0.999, 1e-8, 0.05, 10 bc1, bc2 = 1 - b1 ** step, 1 - b2 ** step wdf = 1 - lr * wd hp = f([[b1, 1 - b1, b2, math.sqrt(1 - b2), 1 / bc2, eps, -lr / bc1 / wdf, wdf]]) p = f(rng.standard_normal((128, M)) * 0.02) m = f(rng.standard_normal((128, M)) * 1e-3) v = f(np.abs(rng.standard_normal((128, M))) * 1e-6) g = bf(rng.standard_normal((128, M)) * 1e-3) return (p, m, v, g, np.zeros((128, M), BF16), hp), {}