Hey (maybe @sakrejda, @syclik, whoever else might know?),
Ben just upgraded the Jenkins Linux node and now I’m seeing the gamma tests fail with numerical differences that look kind of suspicious to me (like a newer compiler exposing some use of undefined behavior or memory?):
http://d1m1s1b1.stat.columbia.edu:8080/blue/organizations/jenkins/Math%20Pipeline/detail/PR-1047/3/pipeline
Anyone have ideas?
That does look weird. I just upgraded a box to 18 so I can check.
bgoodri
3
I think it is the g++-7
thing
I don’t recall what the “9++-7
thing” is but I’ll check for solutions on Monday if nobody else gets to it.
Ah yeah, that sounds right. @sakrejda I think it was just that our tests don’t pass under g++-7
- not sure if anyone found a cause or cure yet.
It looks like they’re up to g++ 7.3. Is that what’s being tested here? We should get Stan math working on recent g++ compilers.
There are unexpected serious numerical differences then NaN differences in:
[ FAILED ] mathMixScalFunGammaP.GoodGammaP (14 ms)
We need to get to the bottom of this. Is there an issue in the math lib for it?
1 Like