bump Enzyme version in v0.2

[Red-Portal]

This PR bumps Enzyme's compat bound to v0.13 so that AdvancedVI v0.2 doesn't hold the Turing ecosystem behind.

[yebai]

@Red-Portal Enzyme is still fragile, as observed in other parts of the TuringLang ecosystem. I don't think it is worth supporting it officially from AdvancedVI.

Maybe move all Enzyme tests into a separate CI task and allow them to fail during PR checks.

[Red-Portal]

@yebai Yeah I was planning to do that in #129

[Red-Portal]

@mhauru @yebai The Turing integration test is failing, but due issues downstream. Should I just merge anyways?

[mhauru]

Thanks @Red-Portal, this is very helpful!

The tests on Turing's master are indeed broken because of 1.11. @penelopeysm, do you think either TuringLang/Turing.jl#2341 or TuringLang/Turing.jl#2368 is close enough to fixing the test breakage that it makes sense to hold off with this one for a moment?

[penelopeysm]

@mhauru I suspect #2341 might be merged by today or tomorrow!

[wsmoses]

@Red-Portal Enzyme is still fragile, as observed in other parts of the TuringLang ecosystem. I don't think it is worth supporting it officially from AdvancedVI.

Maybe move all Enzyme tests into a separate CI task and allow them to fail during PR checks.

@yebai Do you have a reference to anything which is fragile in other parts of turing? From my understanding TuringLang/Turing.jl#1887 has been passing all tests for the past four months without issue (though certainly there's more to do like the additional ones distributions tests @mhauru opened)

[Red-Portal]

@yebai Do you have a reference to anything which is fragile in other parts of turing? From my understanding TuringLang/Turing.jl#1887 has been passing all tests for the past four months without issue (though certainly there's more to do like the additional ones distributions tests @mhauru opened)

My impression is that Distributions.jl is full of demons and dragons such that someone will need to write a comprehensive set of tests to really see how much works and how much doesn't.

[wsmoses]

So @mhauru has been working on this recently here: EnzymeAD/Enzyme.jl#1819

as far as I can tell, its got a few more failures than the other tools, but they're all determinstic in throwing julia errors for currently unsupported distributions (and tend to be moreso forward mode, with reverse mode generally being more fine)

[Red-Portal]

Oh I wasn't aware of that. I'll keep an eye!

[mhauru]

The code looks good to me, and the Turing integration tests now pass (thanks @penelopeysm!). The buildkite thing is failing (is this GPU tests?) but it seems to have been failing for the previous v2 release too, so maybe that's fine? Happy to merge @Red-Portal?

[yebai]

Merged!

[mhauru]

I see I arrived late by seconds :)