I Am facing a similar issue. with the exact same code and same dataset and even after fixing the seed and such, my notebook result differ completely when i train on my MacBook (on which I have tensorflow metal) and when i train on my school GPU.

I have been working on these experiments for 6 months and just recently moved to GPU but now I can’t reproduce any of my result that I have when I run on my MacBook.

For a bit of context I am working on Bayesian deep learning with variational inference. The deterministic model I train gives similar result on both machines. But the different Bayesian networks give extremely different results.

I am looking for any help or any advice regarding this

I have been trying to fix this for the past 3 days and I am very anxious about my results not being reproducible and thus not publishable (PhD student)