Valentin Churavy

StatsFun.jl does have a ChainRules extension, but I see nothing for logsumexp https://github.com/JuliaStats/StatsFuns.jl/blob/master/ext/StatsFunsChainRulesCoreExt.jl

Ah it might be this: https://github.com/JuliaStats/LogExpFunctions.jl/blob/77c5cf030b58b14f118f237b6b518005499a7f40/ext/LogExpFunctionsChainRulesCoreExt.jl#L136

> However I am curious if there would be a way to avoid silent failure here and instead have the forward AD print out an error. This is tricky, Enzyme...

> What's the error you're seeing? If `prev` is an immutable the line ``` setfield!(prev, i, make_zero_immutable!(xi, seen)) ``` in the generic handler will fail.

We definilty need one of the PRs that @danielwe is working on.

Please post `cat /proc/cpuinfo` On Mon, May 23, 2022, 13:56 Oscar Smith ***@***.***> wrote: > How did you install Julia? Many OSes distribute broken versions of Julia. > > —...

Depressingly enough, on 1.11 when it doesn't crash, it calculates the wrong thing. ``` # 377.761956 seconds (794.20 M allocations: 37.793 GiB, 2.14% gc time, 99.78% compilation time) # dA...

> this is the wrong fix overall getRenamedPerCallingConv("PMPI_...", "MPI_x") should give PMPI_x, which ought resolve? No, that's the crux of the issue. A module may contain a mix of PMPI...

> since I assume julia never gives us a mix. So we should never generate a mix? Enzyme.jl is to blame, it tries to invert a pointer from Julia and...

Yeah it makes me wonder if https://github.com/EnzymeAD/Enzyme.jl/blob/43654f904beeeed866305bd5b9e282818a098e50/src/rules/jitrules.jl#L187 should wrap the shadow in a `Ref` but I don't know that code well enough.