It’s a bit o a wild guess, but a quick google suggests some MKL floating-point operations are not necessarily deterministic, and I see you have a dot operation there… numpy - Same Python code, same data, different results on different machines - Stack Overflow
Can you try on a model with just just a simple unobserved normal?
with pm.Model() as m:
x = pm.Normal("x")
pm.sample(..., random_seed=self.seed)
Also do the final results look very different across the different machine?