Christopher Rackauckas

It would definitely be good to document these, just haven't gotten around to it yet.

The number of allocations seems like a regression somewhere, some kind of inference issue.

We can bump the cutoff to using Krylov methods to be higher. Since it's not preconditioned that would help in some cases.

yeah bump it up. > Also @TorkelE, could Catalyst start providing jacobian sparsity? It should just always do that IMO

at least for anything sufficiently large.

Oh interesting, we should make it so if you switch but never actually step that you can keep history alive

Well sparse matrices will be slow for things like 100x100 lu factorization. So you don't want to do it if too small. But yes, if in that large enough range,...

That's fine. I know some cases where that will be slower, so I think overall it's a wash.

Only if the size is sufficiently large. Since a smaller sparse matrix would be slower.

And it needs to be "sparse enough"