You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository has been archived by the owner on Apr 18, 2023. It is now read-only.
Failing with MethodError: no method matching cholesky(::Branch{Matrix{Float64}, Nothing})
08b0d31
Generalize diagonal cholesky to Hermitian
Failing with MethodError: no method matching UpperTriangular(::ZeroTangent)
ebe03e0
Create tangent with factors
Same error, but only for last two tests in test/sensitivities/linalg/factorization/cholesky.jl
cce75a0
Work with factors directly
Exactly the same behaviour as before
b543b0e
Rewrite getproperty rule to store factors
Same errors above, but with some mul! method errors
So far I can kinda follow the ZeroTangent errors from JuliaDiff/ChainRules.jl#630, but I still need to run another bisect to figure out when we started getting the Branch method errors?
The text was updated successfully, but these errors were encountered:
I talked to Rory about this last week in person.
Concluded that we should put in the [compat] section: ChainRules="=1.35.2"
and just pin it to an exactly specified known working version.
Nabla is fragile to ChainRules version changes in subtle ways.
So any changes should be done with care.
As we are not using this in prod (still using the old version of Nabla)
this version of Nabla is more a proof of concept of ChainRules integration than actually needing to be fully robust.
If we did move it into production we should think about how to robustify it more.
(which might be just solving a lot of these case)
Looks like how cholesky factors are being handled by ChainRules is breaking our CI now.
https://github.com/invenia/Nabla.jl/runs/6959631070?check_suite_focus=true#step:6:390
Bisect results:
MethodError: no method matching cholesky(::Branch{Matrix{Float64}, Nothing})
MethodError: no method matching UpperTriangular(::ZeroTangent)
mul!
method errorsSo far I can kinda follow the ZeroTangent errors from JuliaDiff/ChainRules.jl#630, but I still need to run another bisect to figure out when we started getting the
Branch
method errors?The text was updated successfully, but these errors were encountered: