eb8680

Hi @learning-chip, thanks for the suggestion. This seems like an issue best resolved at the level of PyTorch, JAX, and TensorFlow, none of which are fully compatible with NEP-47 yet...

@learning-chip to be clear, Pyro and NumPyro rely crucially on advanced features of PyTorch and JAX (especially higher-order forward- and reverse-mode AD and a large library of probability distributions with...

Hi @JIAQING-XIE, Pyro currently contains two structured Gaussian autoguides, [`AutoStructured`](https://docs.pyro.ai/en/stable/infer.autoguide.html#autostructured) and [`AutoGaussian`](https://docs.pyro.ai/en/stable/infer.autoguide.html#autogaussian). You could subclass or fork either of these if they don't quite do what you want or you...

@JIAQING-XIE I'm afraid I don't quite understand your question. Maybe it would help to clarify some terminology. The term "guide" is [Pyro shorthand for a tractable parametric approximation to a...

@gioelelm good observation! I think we could address this by applying `pyro.subsample` to observed values inside of `ConditionMessenger._pyro_sample` and `pyro.sample` (before any handlers are applied).

For context, here is a condensed training loop from the tutorial notebook that I was trying to enable: ```python # new: keep PyroParams out of the global parameter store pyro.enable_module_local_param(True)...

@jmuchovej sorry, I'll try to look at it this week.

I can take this issue, since I'll need to update `numpyro.contrib.funsor` in the process.

@kamathhrishi is this ready for review? No rush, just don't want you blocked - if so, can you run `make lint` locally and fix the linter errors?

@arijc76 sorry I am so slow this week, great work! Here are some minor comments: **Notebook** - Could you use the same model function for SVI and MCMC, instead of...