Honestly it is mostly caused by people from the machine learning end of the ecosystem not understanding how cooperative multitasking works and trying to bolt a web framework to their model.

That coupled with the relative immaturity of the python async ecosystem leads to lots of rough edges. Especially when they deploy these things into heavily abstracted cloud ecosystems like Kubernetes.

FastAPI also trys to help by making it "easy" to run sync code but that too is an abstraction that is not majorly documented and has limitations.