The suspicon was that it was doing a loop and calling the sylvester solver inplace, but since the solver works "inplace" it might require adding new buffers or zeroing out buffers between calls in the iteration.
But the issue may have been that that it was changing something inplace. Maybe the buff.E or something like that?
After checking which of those values gets modified inplace, you will need to make a copy of them before the loop there and then copy it back into the buff.E or whatever before calling the buffers.
Again, if any of the arguments are modified inplace then you will need to make a copy of it before the loop over n_p and the copy it back over top prior to the call to the generalized_sylvester_solver
[ ] Make sure the unit tests are going to catch gradients on multiple parameters, and with a reused cache between the two calls.
There were previous attempts to integrate the https://github.com/DynareJulia/GeneralizedSylvesterSolver.jl but there was a bug in its usage.
The suspicon was that it was doing a loop and calling the sylvester solver inplace, but since the solver works "inplace" it might require adding new buffers or zeroing out buffers between calls in the iteration.
To do this:
buff.Eor something like that?buff.Eor whatever before calling the buffers.n_pand the copy it back over top prior to the call to thegeneralized_sylvester_solverIf it is working well then
:MatrixEquationsto:GeneralizedSylvesterSolverin https://github.com/HighDimensionalEconLab/DifferentiableStateSpaceModels.jl/blob/v0.4.19/src/types.jl#L372If it seems to win in the benchmarks version
MatrixEquationsand unit tetss pass, then leave that change in the toggle.