Neeraj Pradhan

> @neerajprad Sorry for not replying earlier! What do you mean by initializing across all platforms? (Specifically, what is a platform in this case?) Thanks for looking into this, @skye....

I think we should do this by default. I have found NUTS on many models to be 2-4 times faster with JIT, but due to some complexity that has been...

@fehiepsi - Do you mean setting it to 4 in brmp or in NumPyro? I think setting it to 4 for brmp makes sense (for CPU). In any case, I...

Looking at http://xarray.pydata.org/en/stable/pandas.html, I think we should be able to swap pandas for xarray internally in the future.

> (in cases where the data shape doesn't change between calls) Even if the data shape changes, but we can commit to a certain max size, it might be possible...

From #54: -------- Currently a formula like `'y ~ a * b'` will throw an assertion error during code generation. Maybe we can recognize `*` as a valid Op but...

Using double precision will cause perf issues with #14, but we haven't found it to significantly impact performance on the CPU. Also, NumPyro (JAX) I think will cast everything to...

@null-a : We were running some benchmarks here - https://github.com/pyro-ppl/numpyro/pull/470, and are mostly concluding that right now. From what we can see, it seems very likely that you will observe...

While we don't automatically reparametrize a model, a user can easily do so in NumPyro (not Pyro though) using TransformedDistributions. See this [example](https://github.com/pyro-ppl/numpyro/blob/master/examples/funnel.py) on Neal's funnel. i.e. instead of using...

@null-a - I was looking at this. Currently there is one issue - we need to set the `device` before jax is loaded so doing this on a per-model basis...