Hi all! I know I’m reviving quite an old thread; so much has happened over the last five years, but I eventually developed a new theory of tensor compilation and wrote a tensor compiler: Finch.jl. It supports much more than just sparse tensors, including run-length-encoding, triangular or padded arrays, block-sparsity, and more! Finch doesn’t use Einstein summation as input, rather it uses full loop nests with statements and if-conditions, so you can do convolutions, scatter/gather, fuse or reorder loops, mask computation, and so on. I remembered this thread and thought you might find it interesting. Let me know what you think!