Thanks Andreas. 

I found the answer in an old post after all: http://osdir.com/ml/python-cuda/2011-12/msg00014.html

On 21 September 2016 at 17:16, Andreas Kloeckner <lists@informa.tiker.net> wrote:
Hi Peter,

Please send messages like this to the mailing list in the
future. There's some drama going on with Gmane, but it should be back up
at some point. In the meantime, Mailman and the conventional archives
are still available.

Peter Walsh <pwalsh0@gmail.com> writes:
> I have a question regarding cuda random number generation. This isn't
> appropriate for github I think. Also, FYI I tried searching the mailing
> list for details, but the searchable archive seems to be down:
> http://news.gmane.org/gmane.comp.python.cuda
>
> My question is thus:
> How do I generate random numbers in my own kernel (called from pycuda)? I
> am trying to build a montecarlo simulation - and passing in an "entropy
> pool" from the host is not acceptable.

Random123 is available in PyOpenCL. To my mind, that's the right
approach to parallel RNG. Documentation here:

https://documen.tician.de/pyopencl/array.html#module-pyopencl.clrandom

Hope that helps,
Andreas