David Widmann
David Widmann
What I'm arguing for, actually is not a new or updated policy but the current behaviour already: We already define right multiplication of `PDMat`s with generic vectors and matrices -...
Closing as I don't view the original issue as an actual issue (probably should have been closed when #206 was opened > 1 year ago...).
No, NLSolversBase is not ready yet.
I ran into this issue of AbstractFFTs multiple times, and IMO some way to switch between backends similar to DifferentiationInterface is very much needed at this point, as nowadays there...
> I've tried out a variant of the active backend based on a ScopedValue, which would additionally offer the following interface: I'm not convinced yet that this should be done....
> FFT is here only an implementation detail. Should all functions now get an FFTbackend as karg? If you want to make it configurable, you could make it a keyword...
Maybe instead of a global setting we could copy another design choice from the AD ecosystem, in addition to the unified interface based on backend instances: Every FFT package defines...
With TransformVariables#master and [email protected] I get: ```julia julia> Zygote.gradient(f ∘ transform(t), zeros(2)) ([12.5, 1.0],) julia> FiniteDifferences.grad(central_fdm(5, 1), f ∘ transform(t), zeros(2)) ([12.499999999993662, 0.9999999999999681],) julia> (f∘transform(t))(zeros(2)) 1.0 ```
I was curious and checked it with the latest Enzyme release. The example seems to work now.
buildkite only works from time to time. It doesn't really reliably support the conda installation which is enforced currently.