Defining the Rust 2024 edition

*CUDA* is the issue (as icc/MKL would be for any libraries depending on it), not PyTorch. Assuming you sorted out licensing so that your RPM(s) were fine to distribute (which I suspect would be non-trivial, given CUDA likely has a runtime that you'd need to share with users as well), then having a repository of reproducibly-builtable-by-anyone-with-CUDA PyTorch RPMs would not be a great challenge.

I suspect though because many of the machine learning libraries produce different results with different versions, users will want to freeze exact versions of everything, and only move when they tested with their particular code/dataset pair a new set of versions (which likely means freezing the versions forever because of a lack of time/funds), which isn't exactly amenable to using a distro-provided package.