Reproducer:

import numpy as np
import cupy as cp


cp.cuda.cub_enabled = False  # fall back to cupy's reduction routines
a = [0.5 + 3.7j, np.complex(0.7, np.nan), np.complex(np.nan, -3.9),  np.complex(np.nan, np.nan)]

a_np = np.asarray(a)
print(a_np.max(), a_np.min())

a_cp = cp.asarray(a_np)
print(a_cp.max(), a_cp.min())

Outputs from NumPy and CuPy do not match:

(0.7+nanj) (0.7+nanj)
(nan-3.9j) (nan-3.9j)

NumPy seems to propagate the first encountered NaN, but I can't verified this rigorously. BTW, CUB handles this correctly (set cp.cuda.cub_enabled = True).

I can't figure out what's wrong by inspecting the _amax and _amin implementations in cupy/core/_routines_statistics.pyx. Seems to be an issue of how pairwise operations are actually done in the reduction kernel.