Specifically, calling `sample()` seems inordinately slow on 1.10

[penelopeysm]

The testsets that involve actual AD evaluation (i.e. DynamicPPL.TestUtils.AD.run_ad) have quite similar runtimes, but the moment we call sample, all hell breaks loose (and we are barely even sampling!):

@testset verbose = true "AD / Gibbs sampling" begin
    # Make sure that Gibbs sampling doesn't fall over when using AD.
    @testset "adtype=$adtype" for adtype in ADTYPES
        spl = Gibbs(
            @varname(s) => HMC(0.1, 10; adtype=adtype),
            @varname(m) => HMC(0.1, 10; adtype=adtype),
        )
        @testset "model=$(model.f)" for model in DEMO_MODELS
            @test sample(model, spl, 2) isa Any
        end
    end
end

This testset on 1.10:

    AD / Gibbs sampling                             |   39     39   93m39.7s
      adtype=ADTypes.AutoForwardDiff()              |   13     13   13m57.6s
      adtype=ADTypes.AutoReverseDiff()              |   13     13   36m59.2s
      adtype=ADTypes.AutoMooncake{Nothing}(nothing) |   13     13   42m42.9s

This testset on 1.11:

     AD / Gibbs sampling                             |   39     39  12m58.3s
      adtype=ADTypes.AutoForwardDiff()              |   13     13   2m33.9s
      adtype=ADTypes.AutoReverseDiff()              |   13     13   2m13.8s
      adtype=ADTypes.AutoMooncake{Nothing}(nothing) |   13     13   8m10.6s

It's only two iterations --- is it compilation time that's the problem??

[mhauru]

Ewww... Do we know if the impact is uniform across all models? Hard to imagine it not being compilation if we are only taking two samples.

[devmotion]

Maybe JuliaLang/julia#53371 - a fix was backported to 1.10.2 but the tests were run with 1.10.0 so could be affected by it. Or it could be some other issue that was fixed in 1.10.x so it would be interesting to rerun tests with lts (latest 1.10 release).

[penelopeysm]

It is indeed fixed on 1.10.2 (and slow on 1.10.1) so might very well be that bug specifically.

We could either specifically run CI on >= 1.10.2, or we could change the minimum Julia version to 1.10.2 -- thoughts? I have a mild preference for the latter because it's nice to be able to specify min in CI.

[mhauru]

Assuming Julia is doing semantic versioning well, I see little downside to bumping our min to 1.10.2. Upgrading from 1.10.0 to 1.10.2 shouldn't be a problem for anyone.

[penelopeysm]

thanks @devmotion @mhauru I'll lump that into #2569 🙂