The very broad answer is that Fedora packages tend to be built for one combination of release streams or for a very small number of combinations of release streams, and those aren’t necessarily the right combination for every application.
So, vLLM might need PyTorch, but it might need PyTorch for Python 3.12, while Fedora is building PyTorch for Python 3.14.
In my opinion, if there is a group of developers who want to build vLLM in or for Fedora, we should empower them to do that, and we shouldn’t put more maintenance burden on the people maintaining PyTorch. I think this is one of the class of problems that Modularity intended to solve, but now we need some other approach.