Upper-bound DPPL to remove most recent compiler change

[torfjelde]

No description provided.

[codecov]

Codecov Report

Merging #1578 (3047dc4) into master (b9db77c) will increase coverage by 4.71%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #1578      +/-   ##
==========================================
+ Coverage   73.37%   78.08%   +4.71%     
==========================================
  Files          24       23       -1     
  Lines        1292     1424     +132     
==========================================
+ Hits          948     1112     +164     
+ Misses        344      312      -32     

Impacted Files	Coverage Δ
src/variational/VariationalInference.jl	53.84% <0.00%> (-3.30%)	⬇️
src/Turing.jl	100.00% <0.00%> (ø)
src/core/Core.jl	100.00% <0.00%> (ø)
src/contrib/inference/dynamichmc.jl	100.00% <0.00%> (ø)
src/utilities/Utilities.jl
src/core/ad.jl	84.81% <0.00%> (+0.08%)	⬆️
src/contrib/inference/sghmc.jl	98.41% <0.00%> (+0.19%)	⬆️
src/inference/gibbs.jl	97.33% <0.00%> (+0.23%)	⬆️
src/inference/AdvancedSMC.jl	96.52% <0.00%> (+0.56%)	⬆️
src/inference/ess.jl	93.75% <0.00%> (+0.89%)	⬆️
... and 13 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update bffc8be...3047dc4. Read the comment docs.

[torfjelde]

@devmotion Tests are running like 10mins faster here compared to #1577 , so we're looking at a ~15% slowdown?

[devmotion]

Is this relative increase consistent with the timings in your experiments?

[devmotion]

Is this relative increase consistent with the timings in your experiments?

[torfjelde]

Was difficult to see a significant slowdown in those experiments. We observed a ~15% slowdown for model2 (bernoulli coin-flip), so maybe?

[devmotion]

I am a bit surprised though since my impression was that it was mainly compile time regressions. So I assumed it would have a smaller impact in the tests since we use the same model repeatedly.

[torfjelde]

I guess we didn't check AD-related stuff :/ E.g. maybe Zygote is significantly slower?
We need to recompile the model for different types, so maybe that's why we're seeing such an increase in testing time too?

Anyways, it's probably a good idea to run the benchmarks present in https://github.com/TuringLang/TuringExamples for the different versions and see if we can find something clearcut.
I'll get on it, but uncertain if I'll have it available today.

In the meantime, can we try to get the outstanding PRs merged? E.g. bump AbstractMCMC versions across the board and merge #1567 ?

[devmotion]

In the meantime, can we try to get the outstanding PRs merged? E.g. bump AbstractMCMC versions across the board and merge #1567 ?

Sure, let's finish these other PRs.