Feature request: Multiple server runtimes

I have multiple GPUs of different architectures in the same machine (for example, a laptop with an intel iGPU + nvidia dGPU), where I want to run Vulkan based llama.cpp in iGPU (where I have 96 Gb of shared memory, but it is awfully slow) for one model, and cuda based llama in dGPU (where I have only 4 GB of memory, and I can only feasibly run nemotron-4b for light tasks)
I would be very glad if I could do this within the lms.
Maybe the selection of runtime can be carried to model load, and can be selected per model instead of a global setting, sparkrun style?

The issue identifies no files, tests, or entry points, so first map where server runtime selection and per-model loading are documented or implemented. Confirm the supported runtime behavior and define completion as a documented, working way to choose different runtimes for different models.

Written by the indexing model from the issue text.