`JointDistribution`s yield different PRNG behaviour on GPU and CPU

[chrism0dwk]

Hi All,

I noticed recently that stateless PRNG behaviour for JointDistribution* objects appear to do different things depending on whether random numbers are generated on GPU or CPU. This is demonstrated this Colab.

This appears to go against the assertions in the tf.random.stateless_* documentation that suggests that PRNG behaviour should be consistent across CPU and GPU devices. I guess that there may be reordering of operations within the JointDistribution* logic, but perhaps this should be documented?

Regards,

Chris

[edit] interestingly, this seems relevant only to Eager and Graph mode. XLA mode appears consistent.