Xianda Sun

That is what confuses me, because in this PR, the order can't be reversed. (I think it will fail if we use the old velocity.)

Tag @yebai for knowledge and review. I think this is ready (certainty of correctness is quite high but not 100 percent).

@ErikQQY version bump maybe? (We would need to bump minor version because `SGHMC` will be exported. Probably lump several changes together or do something like https://github.com/TuringLang/Turing.jl/pull/2517)

the failed test is not related to this PR, @yebai maybe you already have an answer to why it fails?

let's delay a bit to see through https://github.com/TuringLang/AbstractMCMC.jl/pull/150

### Some notes for completion of the second item One can already use `AdvancedHMC` and `AdvancedMH` though the [`ExternalSampler` interface](https://github.com/TuringLang/Turing.jl/blob/5b24cebe773922e0f3d5c4cb7f53162eb758b04d/src/mcmc/Inference.jl#L97-L165) (more details to be find at https://turinglang.org/docs/tutorials/docs-16-using-turing-external-samplers/) ```julia julia> using...

This is pretty good. Couple of things, but just my opinion: * it might make more sense to define [AGENTS.md](https://agents.md/) instead of [copilot-instructions.md](https://copilot-instructions.md/), the former would work with more coding...

It might also be useful to check for if `logp` ever become `Inf` at any point

I don't know of a reason for it either. (JuliaBUGS always stores the untransformed values, but JuliaBUGS is simpler than DynamicPPL too.)