you either use like a game engine that generates all the platform specific gpu code and shaders for you, or you can take advantage of writing all the different gpu specific and shader code yourself (bc you need specific features or can make it faster or some such), i can’t really imagine something in between tbh
and a generalised tool like this is destined to not fit platform specific features that well; for example, i couldn’t find any mention of an interface to use tensor cores, and automatic differentiation w/o a relatively complex expression optimiser is kinda useless for machine learning, where tensor wrt tensor derivatives' dimensionality explodes w/o using an expression optimiser…
it’s kind of the same story as w/ cross platform gpu frameworks: they’re destined to just kind of do everything but badly bc different platforms have different architectures that you can’t map to w/ a single api… imo the best solution remains to have domain specific solutions that generate separate cross platform code solely in that domain...