You don’t need to set the pytensor compile mode, that’s done automatically for you when you ask for a non-standard sampler. There might be something funny going on because you’re asking for numba mode, but then using numpyro which is jax-based. Actually different optimizations are applied to your model based on the backend, so this might end up mattering.
Also LLVM is the JIT compiler (transpiler?) that numba uses (JAX uses XLA) so it makes me suspect this line.