Support for speculative decoding, draft models in llamafile?

[dagbdagb]

I have been playing with tabbyAPI and its support for draft models. In short, the performance benefit is very, very obvious. And I wonder what it could mean for inference on the CPU. Or even a mixed CPU/NPU/GPU setup.

Intuitively (which may be very wrong, of course) I think this could allow for 14B models being more available to the GPU poor. In a practical sense.

See some ballpark numbers without speculative decoding here:

[richieledude]

YASSSSS! vllm does this easily. would be awesome to see support in llamafile and llama-cpp!!!