I'm building a system which has many different passes (~15 so far). Almost every pass is a LLM invocation, which takes time. My original idea was to use a smaller model, such as Phi-2, as a gateway in front of all those passes: I'd describe which pass does what, and then ask Phi-2 to list the passes which are relevant for the user query (I called it "pass masking"). That would save a lot of time and collapse 15 steps to 2-3 steps on average. In fact, my Solar 10.7B model does it pretty well, but it takes 7 seconds for the masking pass to work on my GPU. Phi-2 would finish in ~1 second. However, I'm really struggling with Phi-2: it fails to reason (what's relevant and what's not), unlike Solar, and it also refuses to follow the output format (so that I could parse the output programmatically and disable the irrelevant passes). Again, my proof of concept works with Solar, and fails spectacularly with Phi-2.
> You are a helpful assistant to 'User'. You do not respond as 'User' or pretend to be 'User'. You only respond once as 'Assistant'. 'System' will give you data. Do not respond as 'System'. Allow yourself inner thoughts as 'Thoughts'.
and then I constrain its answers to Thoughts: [^\n]* and Assistant: <JSON schema>, and I have two shots included in the prompt.
I haven't been able to get anything useful out of Phi-2 in llama.cpp (but I only tried quantized models). I use python/huggingface's transformers lib instead.