I would understand if numpy was smart enough to see that array % 1 does not do anything, and that runtime is independent of the input. I would also understand if runtime was longer for larger numbers, maybe because dividing a large number takes longer (compare Why is floating-point division in Python faster with smaller numbers?). But why does array % 1 run longer for smaller numbers?

from timeit import timeit

import numpy as np

a1 = np.random.randint(500,        size=20_000_000, dtype=np.int32) - 250
a2 = np.random.randint(20_000_000, size=20_000_000, dtype=np.int32) - 250

print(timeit(lambda: a1 % 1, number=5))
print(timeit(lambda: a2 % 1, number=5))
print(timeit(lambda: a1 % 1, number=5))
print(timeit(lambda: a2 % 1, number=5))

Output:

0.4755298000000039
0.19294559999980265
0.460197700000208
0.19560679999995045