Adrian Seyboldt

Thanks for working through all those examples, this is really nice! I only had a quick look at the code so far, but I think there are two issues with...

I don't think looking at the jaxpr can tell us if the jax/xla will constant fold the function. jaxpr is just the input to the optimization pipeline, not the result...

That you for the bug report. I had seen something possibly related recently, but didn't manage to find an example in a smaller model. This example should make it much...

Well, that was easier than I though, and won't be hard to fix. The problem is that the argument name for the point in parameter space is `x` and if...

Yes, definitely needs a fix, I'll push one soon. Out of curiosity (I don't have a apple), could you do me a small favor and run this with jax and...

Thanks for the numbers :-) First, I think it is important to distinguish compile time and sampling time. The numbers you just gave me show that the numba backend samples...

For sampling in separate processes: ```python # At the very start... import os os.environ["JOBLIB_START_METHOD"] = "forkserver" import joblib from joblib import parallel_config, Parallel, delayed import arviz def run_chain(data, idx, seed):...

I just did a little experiment with strace, to see what zarr does when we write a value to a chunk. I wrote a little script that creates the array,...

Is the cost of checking the input length really something that matters in any circumstances? That should be many orders of magnitude smaller than any logp evaluation anyway?