Is there a better solution? Should we treat model weights like a reference genome in the bioinformatics world and arrange to always pass it as an explicit module input, moving the responsibility for downloading and caching the weights into the pipeline?