The craziest thing for me is that they stopped funding it. You would expect them to want a viable drop-in alternative to their biggest rival’s API to ease migration, but no, they let it go.
I get the counterpoint about “supporting CUDA” being detrimental to their own approach, but to be honest, I think they should have kept it going.
At least now it is Open Source, so there might be a fair bit of traction in the HPC and AI/ML communities.