[Python] Why does generating the Sobol Sequence in SciPy hang after completing 25% of the...

I'm doing a heavy Quasi Monte Carlo simulation using the Sobol sequence, I need the generator to be able to create over 2 ** 30 samples. SciPy says that the generator is capped at 2 ** 30. It says also that by modifying the argument bits it should be possible to lift the limit to a maximium of 2 ** 64 (That is enough for me).

sampler = qmc.Sobol(d=2, scramble=True, bits=64)

However I still encounter the same bug, at exactly 25 per cent of the function, everything stops without error message. I am sure the problem is in the generator because i have tried moving it inside the for loop and it worked, however regenerating the Sobol sequence each time i think might cause math trouble with QMC properties.

Here it is the code in a minimal reproducible version (You must have CUDA installed)

import cupy as cp
from scipy.stats import qmc
from tqdm import tqdm
import os


def run_simulation_batch_cupy_QMC(samples, a):
batch_size = 2 ** 27
n_batches = samples // batch_size
sampler = qmc.Sobol(d=2, scramble=True, bits=64)

for i in tqdm(range(n_batches)):

qmc_samples = sampler.random(n=batch_size)

#Here there is some computation I cut off

cuda_bin_path = r"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\bin"
os.add_dll_directory(cuda_bin_path)

if __name__ == "__main__":

samples = 2 ** 34 #Samples must be, to get the bug, greater than 10 ** 10 and a power of 2 so 2**34
a_values = [1.1, 1.2, 1.3] #Example

for a in tqdm(a_values):
run_simulation_batch_cupy_QMC(samples, a)



I was expecting the function to complete the generation. Monitoring the GPU and other I can say I do not get an memory overflow. The GPU just stop calculating exactly at 25 %.

Continue reading...