Kai Xu

That might be updated to https://turing.ml/dev/docs/using-turing/guide#sampling-multiple-chains. But this link is no longer informative to AHMC users, as many of the patterns have been built into AbstractMCMC now and users doesn't...

> Is the vectorized sampling already available for Turing and does AbstractMCMC have an interface for it? No. IIRC that part was written by @Vaibhavdixit02. Can you remind us what...

Currently the codes are at https://github.com/TuringLang/AdvancedHMC.jl/blob/master/src/contrib/ad.jl but we may want to make them synchronized with Turing's somehow.

Just sync the understanding. The paper has NUTS on GPUs but not the batch-mode multiple chain stuff in #117 right?

Related DynamicHMC.jl issue: https://github.com/tpapp/DynamicHMC.jl/issues/110 We could also check the difference between AHMC and DHMC on GPU using the example there.

I agree we should keep both for a while. As you said, doing consistency check on more targets can give us confidence on both implementations.

> I did need to make one functional change to recursive NUTS for consistency with iterative. Namely, if a proposed subtree has terminated, it still draws a random number for...

> Thanks @xukai92 for the review. I got busy and have had to put this on hold. I also think the PR might need to be broken into a few...

It looks like if we refactor our metrics using [Manifolds.jl](https://github.com/JuliaManifolds/Manifolds.jl), we could get this for free.

@torfjelde Just realise that we have an issue for this thing.