Functions from random module used only in preliminary stage for placement tame and wild herds. So, there is almost no influence on speed of the algorithm. Only if we try to select a huge number of kangaroos. Also, when the problem is very simple (16 bit, for example) the time of the placement may be significant in comparison with the time of solution. But total time still very small for human.
For pseudorandom hops the coordinate X itself is used.
Main goal of the code presented is a template with minimal dependencies. Numpy is external module which requires installation.
But I can't eliminate gmpy2 at all, it is very efficient.
Anyway, the table with time-to-solve estimations posted by j2002ba2 means GPUs are 3..4-order more efficient than the code presented.
j2002ba2, if you are here, how many hops/sec you have reached on single Tesla V100? Is my estimation of 2 GHops/s correct? At least 1 GHops/s required if authomorphisms taken into account, but I can't beat even half of this value.