Revert "Fix ordering of the promotion [Take II]" #1075
Conversation
| sol, pb_f = DiffEqBase._solve_adjoint( | ||
| prob, sensealg, Tracker.data(u0), Tracker.data(p), | ||
| out = DiffEqBase._solve_adjoint(prob, sensealg, Tracker.data(u0), Tracker.data(p), | ||
| SciMLBase.TrackerOriginator(), args...; kwargs...) | ||
|
|
||
| if sol isa AbstractArray | ||
| !hasfield(typeof(sol), :u) && return sol, pb_f # being safe here | ||
| return sol.u, pb_f # AbstractNoTimeSolution isa AbstractArray | ||
| end | ||
| return convert(AbstractArray, sol), pb_f |
There was a problem hiding this comment.
probably reverting just this will do
|
Can you make that PR? I'm out and about right now. |
|
Source of the errors doesn't seem to be that PR. tests are still failing https://github.com/SciML/DiffEqBase.jl/actions/runs/10584877403/job/29330163818?pr=1075#step:6:1828 |
|
@ChrisRackauckas based on the stacktrace I have a feeling this is an OrdinaryDiffEq issue. I would expect the integrator cache to hold TrackedArrays and not regular arrays when using discrete adjoints. |
|
I was starting to fear that it might be related to the OrdinaryDiffEq split PRs. But I don't see how? https://github.com/search?q=repo%3ASciML%2FOrdinaryDiffEq.jl%20value&type=code Everything is directly using DiffEqBase.value, which is the thing that should be de-tracking as appropriate. |
Reverts #1032. Sees if that fixes some of the Tracker tests that started failing.