Always expect a Libtask trace to be present

[penelopeysm]

Closes #2754: with this patch, I get the following results

julia> while true
           score = AdvancedPS.advance!(trace)
           score === nothing && break
           acc_val = trace.model.f.varinfo.accs[:LogLikelihood].logp

           print("score: $(round(score, digits=3))")
           println("\taccumulator: $(round(acc_val, digits=3))")
       end
score: -3.418	accumulator: 0.0
score: -3.417	accumulator: -3.418
score: -3.417	accumulator: -6.835
score: -3.417	accumulator: -10.252
score: -3.418	accumulator: -13.669
score: -3.417	accumulator: -17.087
score: -3.419	accumulator: -20.505
score: -3.418	accumulator: -23.923
score: -3.417	accumulator: -27.341
score: -3.417	accumulator: -30.758

[github-actions]

Turing.jl documentation for PR #2757 is available at:
https://TuringLang.github.io/Turing.jl/previews/PR2757/

[penelopeysm]

I think some explanation is necessary, mostly for the unknown future person (which might well be me) who comes across this, and wonders why I got rid of these checks.

Generally, these functions used to attempt to check whether the function was being executed from inside a Libtask taped task. If not, then we attempted to use a "normal" execution path, i.e., using the VarInfo that is passed to tilde_assume!!, etc. -- but if we were, then we had to use the VarInfo that was stored as a 'taped global' in Libtask.

The problem with this is that the check objectid(vi) == arg_vi_id used to sometimes return true, even if there was a trace present, i.e., even if get_trace_local_varinfo_maybe returned the VarInfo stored in the taped globals. I actually completely don't know why. The fact that this was true meant that several lines below this, we never set the VarInfo back in the task's taped globals, which was responsible for the behaviour in #2754.

Why can we remove these checks...? Well, there is no way to actually use ParticleMCMCContext without going through sample(model, ::Union{SMC,PG}, ...). That function (technically, the functions that it calls) is what sets the leaf context of a model to be ParticleMCMCContext:

Turing.jl/src/mcmc/particle_mcmc.jl

Lines 31 to 38 in 026d07c

	function TracedModel(
	model::Model, varinfo::AbstractVarInfo, rng::Random.AbstractRNG, resample::Bool
	)
	model = DynamicPPL.setleafcontext(model, ParticleMCMCContext(rng))
	args, kwargs = DynamicPPL.make_evaluate_args_and_kwargs(model, varinfo)
	fargs = (model.f, args...)
	return TracedModel(model, varinfo, resample, fargs, kwargs)
	end

It is also what wraps the model in a Libtask trace. (That code is in AdvancedPSLibtaskExt.)

Furthermore, it is meaningless to use ParticleMCMCContext outside of a Libtask trace (it is pretty much impossible to run SMC on a Turing model without the functionality that Libtask provides -- trust me, I've tried). So there is a one-to-one correspondence: if you are using ParticleMCMCContext, you are within a Libtask trace, and vice versa.

So there's really no point in checking whether it is local or not. My current hypothesis is that the check is a vestigial bit of code from the old implementation, which used SamplingContext{<:SMC}, and in those cases maybe it was easier to accidentally run into this situation where the model was being executed with that context, but outside of a Libtask trace. Personally, I still don't know how that could ever happen; but anyway, here we are.

Technically, the code that I'm changing here was written in #2585; but the whole logic of 'checking whether we're in a trace or not' predates that PR, so goes back quite a while.

[codecov]

Codecov Report

❌ Patch coverage is 70.00000% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.05%. Comparing base (751ae1d) to head (026d07c).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/mcmc/particle_mcmc.jl	70.00%	6 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2757      +/-   ##
==========================================
+ Coverage   84.96%   85.05%   +0.08%     
==========================================
  Files          20       20              
  Lines        1264     1258       -6     
==========================================
- Hits         1074     1070       -4     
+ Misses        190      188       -2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[penelopeysm]

CI failures are in the optimisation interface, which is completely unexpected, but also unrelated to this PR, and will be fixed in #2747. It appears to be random since not all runners are failing.