How is Co-dfns not what you want? It is by far the best attempt I know of for bringing the full power of APL to accelerators like GPUs[0].
...however, it is not easy. Full APL has a lot of features (like nested arrays) that are incredibly difficult to map to GPU execution. And in general, APL is not easier to map to GPUs than your bog standard Numpy-style matrix library (in fact it's harder, since Numpy only has regular arrays and simple operations).
[0]: I was involved in a project that targeted only a small well-behaved subset of APL, which was hard enough: https://github.com/henrikurms/tail2futhark
It would be a great dsl for tensor comprehensions. Also it's a better notation for much (most?) of the math of data science. I wish papers used it. Then the math would be unambiguous and testable and maybe even correct.
No experience, and not APL; but [0] pops up on my radar now and then.
There is this:
https://github.com/Co-dfns/Co-dfns
But it's not quite it.
As for the word 'deserve': computer languages don't really deserve anything, they get adopted, or not. APL is a beautiful language and it gives me immense pleasure when I 'get' a program but the barrier between 'the problem' and 'the solution' hinges on libraries, eco systems and communities. I suspect that APL's window of opportunity has passed and even if Clojure became 'a modern lisp' and Elixir became 'a modern Erlang' I don't really see a big driver behind 'A modern APL'. What would it do that can't be done in other ways?