I am looking for a function that makes a SVD on a matrix that does not fit into memory. I am interested in using just one computer, although an answer for a cluster can be useful for other users.

The python dask package is able to do this in two ways (methods):

- divide an conquer (for Tall-and-Skinny matrices)
- compressed svd (for all kinds of matrices)

So far, I have seen the YAXarray.jl library, but it does not implement SVD. It seems to be able to make operations on matrices that do not fit into memory, but those operations seem to be limited to the application of a function on every element of the matrix.

I have also tested the Mmap.jl library with HDF5, and applied the svd() from LinearAlgebra.jl, but this seems to internally create a copy of the matrix in RAM. Even svd!() can internally create matrices in RAM.

I have read a bit of Dagger.jl, but it also seems to not implement this either.

So, can I achieve my purpose by the use and combination of existing Julia libraries, or has this to be implemented?. In the latter case, where should I make the request?