Why does FFI Calls require a cuda implementation

[pseudo-rnd-thoughts]

I'm adding XLA support to a python/cpp project so have been reading the FFI documentation.

The cpp code all runs on the CPU, I just want to be able to JIT with the python/cpp being unaffected.
I've successfully compiled the code and got it working, however, as soon as I have jax[cuda12] installed, then I get the following error UNIMPLEMENTED: No registered implementation for custom call to "func_name" for platform CUDA

From an uninformed perspective, this doesn't make any sense to me.

• Why do I need a cuda implementation for a function that needs to run on cpu?
• Is there way to effective implementation this? It looks like I copy and paste the functions with the addition of cudaStream_t at the beginning for the cuda implementation

I hope someone understands and can help

Code for reference: https://github.com/pseudo-rnd-thoughts/Arcade-Learning-Environment/tree/xla-support