Significantly improve the Libtask library using ideas from Mooncake / ReverseDiff

[mhauru]

TODO

Edit this issue with more details about what needs to be done.

Known issues

• Examples of a segfault with current Libtask: here and test/mcmc/Inference.jl segfaults on GHA Windows runner (sometimes) Turing.jl#2379 (comment)
• Performance degrades on Julia 1.11: Replace old Gibbs sampler with the experimental one. Turing.jl#2328 (comment)
• TracedRNG isn't deterministic even on the same architecture Particle Gibbs reproducibly gives different results on x64 vs x86 AdvancedPS.jl#108 -- this has been resolved independently of this issue.

[yebai]

@willtebbutt and I have discussed this a few times during meetings. Will, can you add your plan here?

[willtebbutt]

Regarding the issues above linked above.

Examples of a segfault with current Libtask: here and TuringLang/Turing.jl#2379 (comment)

The first example isn't a segfault, it's an OOM error on x86. Additionally, TuringLang/Turing.jl#2379 (comment) appears to be a Windows specific issue.

In both cases, it's quite unclear to me what relationship we believe that these have to the introduction of ideas from Mooncake / ReverseDiff . @yebai could you explain how you think these are related?

[yebai]

In both cases, it's quite unclear to me what relationship we believe that these have to the introduction of ideas from Mooncake / ReverseDiff . @yebai could you explain how you think these are related?

No, they are not directly related -- ideas from Mooncake and ReverseDiff are mainly to improve the robustness and maintainability of the Libtask implementation.

[willtebbutt]

Re TuringLang/Turing.jl#2379 -- I've just spoken with @mhauru and @penelopeysm and apparently we've not seen it happen again since it was initially reported. If it pops up again we can take a look, but otherwise I'm going to assume that I'm unlikely to see it.

[penelopeysm]

Would you assume that you don’t observe it? ;)