Closed superckl closed 1 year ago
There is a big notice in the README about this kind of thing: https://github.com/JuliaDiff/SparseDiffTools.jl#note-about-sparse-differentiation-of-gpuarrays-bandedmatrices-and-blockbandedmatrices , though I am surprised that this case hits some kind of bad fallback for the matrix decompression instead of just erroring.
This is fixed by ArrayInterface v7
When using a
BandedMatrix
as the sparsity pattern for a Jacobian, only the diagonal elements are computed:These issues do not occur if
using ArrayInterfaceBandedMatrices
is included.