llm-mlc VS ad-llama

I'm still iterating on that. Plugins get complete control over the prompts, so they can handle the various weirdnesses of them. Here's some relevant code:

https://github.com/simonw/llm-gpt4all/blob/0046e2bf5d0a9c369...

https://github.com/simonw/llm-mlc/blob/b05eec9ba008e700ecc42...

https://github.com/simonw/llm-llama-cpp/blob/29ee8d239f5cfbf...

I'm not completely happy with this yet. Part of the problem is that different models on the same architecture may have completely different prompting styles.

I expect I'll eventually evolve the plugins to allow them to be configured in an easier and more flexible way. Ideally I'd like you to be able to run new models on existing architectures using an existing plugin.

What is the advantage of this versus running something like https://github.com/simonw/llm , which also gives you options to e.g. use https://github.com/simonw/llm-mlc for accelerated inference?

I'm quite impressed with Llama 2 13B - the more time I spend with it the more I think it might be genuinely useful for more than just playing around with local LLMs.

I'm using the MLC version (since that works with a GPU on my M2 Mac) via my https://github.com/simonw/llm-mlc plugin.

I took a stab at making something[1] like guidance - I'm not sure exactly how guidance does it (and I'm also really curious how it would work with chat api's) but here's how my solution works.

Each expression becomes a new inference request, so it's not a single inference pass. Because each subsequent pass includes the previously inferenced text, the LLM ends up doing a lot of prefill and less decode. You only decode as much as you actually inference, the repeated passes only end up costing more in prefill (which tend to be much faster tok/s).

To work with chat tuned instruction models, you can basically still treat it as a completion model. I provide the previously completed inference text as a partially completed assistant response, e.g. with llama 2 it goes after [/INST]. You can add a bit of instruction for each inference expression which gets added to the [INST]. This approach lets you start off the inference with `{ "someField": "` for example to guarantee (at least the start of) a json response and allow you to add a little bit of instruction or context just for that field.

I didn't even try with openai api's since afaict you can't provide a partial assistant response for it to continue from. Even if you were to request a single token at a time and use logit_bias for biased sampling, I don't see how you can get it to continue a partially completed inference.

[1] https://github.com/gsuuon/ad-llama

Can you point me to some text-adventure engines? I'm hacking on an in-browser local llm structured inference library[1] and am trying to put together a text game demo[2] for it. It didn't even occur to me that text-adventure game engines exist, I was apparently re-inventing the wheel.

[1] https://github.com/gsuuon/ad-llama

[2] https://ad-llama.vercel.app/murder/

Yup, I'm building a library that runs LLM's in browser with tagged template literals: https://github.com/gsuuon/ad-llama

I think it has fundamental DX benefits over python for complex prompt chaining (or I wouldn't be building it!) Even still -- if their focus is purely on AI, python is still the better choice starting from scratch. The python AI ecosystem has many more libraries, stack overflow answers, tutorials, etc available.

Generating an FSM over the vocabulary is a really interesting approach to guided sampling! I'm hacking on a structured inference library (https://github.com/gsuuon/ad-llama) - I also tried to add a vocab preprocessing step to generate a valid tokens mask (just with regex or static strings initially) but discovered that doing so would cause unlikely / unnatural tokens to be masked rather than the token which represents the natural encoding given the existing sampled tokens.

Given the stateful nature of tokenizers, I decided that trying to preprocess the individual token ids was a losing battle. Even in the simple case of whitespace - tokenizer merges can really screw up generating a static mask, e.g. we expect a space next, but a token decodes to 'foo', but is actually a '_foo' and would've decoded with a whitespace if it were following a valid pair. When I go to construct the static vocab mask, it would then end up matching against 'foo' instead of ' foo'.

How did you work around this for the FSM approach? Does it somehow include information about merges / whitespace / tokenizer statefulness?