speculative decoding in llama.cpp

[thetagrid15]

I'd approach it in a boring order. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at reranking. Not glamorous advice, but repeatable baselines save a lot of time.