SMC-based LLM provider with user-supplied potential functions (LLaMPPL-style steering)

[datvo06]

Currently, LiteLLMProvider generates each Template result from a single greedycall_assistant trajectory.
With the latest development in LLM and probabilistic programming for more semantically oriented generation beyond prompting:

1. Sequential Monte Carlo steering (LLamPPL), so we can sample from $p_{\text{LM}}(x),\varphi(x)$ where $\varphi$ is a potential the caller supplies.
2. Alternatively, we could apply this as per-call/per-template level, like ModelSMC, or use sampling for parallelized sequence generation

A few design choices to make:

1. What is the granularity? per-token, per-template, etc.?
2. How much of numpyro/dynestyx can we reuse?
3. Interface: Do we want to extend the existing interface to return multiple results instead of one, along with problogs?
4. Interfacing with tools. If part of the context is generated by tools, do we still get those tokens' problog?

[jfeser]

I'm not sure we want to get into the business of implementing low-level inference methods in effectful; I think any method that can't be built on top of a provider API is out of scope for this library. That doesn't mean we should make it hard to use a steered LLM as a template provider, but we shouldn't build token-level steering in this repository.

I think per-template or per-response sampling is in scope. This is a plausible use case for the weighted module.

[datvo06]

Right. Litellm provides log probs for whole sequence so we could still implement ModelSMC on top of Effectful/Weighted.

[datvo06]

But token-level steering is out of scope unless we can import it from another library and provide functionality on top of it, say, an LLamPPL API/provider. But there might be a huge gap for LLamPPL to provide the functionality that LiteLLM provides.