Diagonal CMA error with high number of parameters and multiple workers

[N1K1T4]

The error is the same as in multiple issues (#1508, #1593, #1609):
RuntimeError: Mean shift samples are expected but missing.

I stumbled upon it with NGOpt -> NGOpt16 -> CMA -> CMAbounded.

I did not find how it was fixed in the other issues so I am not sure if it is caused by nevergrad use of cma or cma itself.

Reproduction steps

With this environement:

• Python 3.11.2
• cma==4.0.0
• nevergrad==1.0.8

Run the following script:

#!/usr/bin/env python3

import nevergrad as ng
import random


def _loss(*args, **kwargs):
    return random.random()


def main():
    optimizer = ng.optimizers.CMAbounded(
        parametrization=300,
        budget=1000,
        num_workers=2,
    )
    optimizer.minimize(_loss)


if __name__ == "__main__":
    main()

Notes

• It seems to work fine with a dimension of 299.
• Same error with LargeDiagCMA and CMApara while OldCMA, LargeCMA, TinyCMA, CMAsmall, CMAstd and CMAtuning work fine. The common denomminator is diagonal=True.