SBI Dependency Arviz Introduced Breaking Changes

[florianehmann]

🐛 Bug Description

When you currently install the sbi package through one of the recommended ways, it will also install the dependency arviz. They recently had their 1.0.0. release and that seems to have introduced breaking changes. Notably, the arviz.data module can no longer be found.

🔄 Steps to Reproduce

1. Follow instructions to install sbi with uv and Python3.12
2. Minimal code example that triggers the bug: from sbi.inference import NPE
3. Error message:

   Traceback (most recent call last):
     File "<stdin>", line 1, in <module>
   ModuleNotFoundError: No module named 'arviz.data'
   

✅ Expected Behavior

Would have expected import to work without error.

📌 Additional Context

...

[florianehmann]

Workaround is to

pip uninstall arviz
pip install "arviz<1.0"

[janfb]

Hi @florianehmann, thanks for reporting this and posting a workaround 🙏

In a recent PR we dropped the explicit dependency on arviz (now only indirectly via pymc), but we will not have any direct use of arviz anymore. Thus, in the upcoming sbi release this issue will likely be solved.

Let's keep the issue here open until then.

[patelshivani2283-lab]

Hi, I took a quick look into this and noticed that sbi is still importing InferenceData from arviz.data in a couple of places:

• sbi/samplers/mcmc/pymc_wrapper.py
• sbi/inference/posteriors/mcmc_posterior.py

Since arviz.data was removed in ArviZ ≥1.0, this seems to be the direct cause of the error.

I’m thinking of updating these imports to use from arviz import InferenceData (with a small fallback for older versions). Does that sound like a reasonable approach, or would you prefer handling this differently given the upcoming release?

[manuelgloeckler]

Hey @patelshivani2283-lab ,

Thanks for checking, but I cannot find any imports in

• sbi/samplers/mcmc/pymc_wrapper.py
• sbi/inference/posteriors/mcmc_posterior.py

Might have been merged in/you considered a earlier version. Given that we no longer depend on arviz there should not be any import from arviz.

[patelshivani2283-lab]

Thanks for the clarification!

Got it — that makes sense. It seems like the issue mainly affects the current released version, where pymc still brings in arviz>=1.0.

Maybe it could help to mention the workaround (pinning arviz<1.0) in the docs until the next release is out? Happy to help with that if needed.

[manuelgloeckler]

This should be resolved with the new release which is almost done i.e. see #1718 . Will keep this issue open which contains the workaround.