Solving sparse linear systems on the GPU

I believe the sparse matrix features in pytorch aren’t as developed as the dense features yet, but at least we can create sparse matrices with torch.sparse.Tensor. If I wanted to do a sparse linear solve on the GPU by giving it a sparse matrix and a dense vector, what would be the best way to do it? Should I look into external packages like cupy? Does anybody have experience calling packages like suitesparse from python? I’d still prefer to have my tensors in pytorch since I’m using many pytorch features, but at some point I need to do the sparse solve so even if I’m using a external library I’d like it to have decent interop with pytorch at the GPU level.

1 Like

Hi I am also insterested in this! Have you found any solution?