basaran VS AutoGPTQ

What are the current best "no reinventing the wheel" approaches to have Langchain use an LLM through a locally hosted REST API, the likes of Oobabooga or hyperonym/basaran with streaming support for 4-bit GPTQ?

Disclaimer: I am curating LLM-tools on github [1]

A few thoughts:

* allow for custom endpoint URLs, this way people can use open source LLMs with a fake openAI API backend like basaran[2] or llama-api-server[3]

* look into better embedding methods for info-retrieval like InstructorEmbeddings or Document Summary Index

* Don't use a single embedding per content item, use multiple to increase retrieval quality

1 https://github.com/underlines/awesome-marketing-datascience/...

2 https://github.com/hyperonym/basaran

3 https://github.com/iaalm/llama-api-server

open-source alternative to the OpenAI text completion API (https://github.com/hyperonym/basaran)

Guanaco-65B[0] using Basaran[1] for your OpenAI compatible API. You can use any ChatGPT front-end which lets you change the OpenAI endpoint URL.

[0] An fp4 finetune of LLaMA-30B by Tim Dettmers

[1] https://github.com/hyperonym/basaran

For lm-eval, I think you'd either need to take GPTQ's inference script and shim it into a model: https://github.com/EleutherAI/lm-evaluation-harness/tree/master/lm_eval/models or you might be able to use a project like https://github.com/hyperonym/basaran and then you could use the gpt3 model...

There are also: - Basaran repo: "Basaran is an open-source alternative to the OpenAI text completion API. It provides a compatible streaming API for your Hugging Face Transformers-based text generation models". "...Compatibility with OpenAI API and client libraries..."; - llama-cpp-python repo: "Simple Python bindings for @ggerganov's llama.cpp library...". "...OpenAI-like API...".

I want to start with a totally freely available model, so again, that excludes things like LLaMA where the weights are only available through a wait list. The two models that most get my attention and (I think, and hope) fit my criteria of open availability are Cerebras-GPT (13b) and Bloomz (7b). The tools to process and fine-tune that seem most feasible to me, from my limit knowledge, are xturing and basaran.

Deploying the 60B version is a challenge though and you might need to apply 4-bit quantization with something like https://github.com/PanQiWei/AutoGPTQ or https://github.com/qwopqwop200/GPTQ-for-LLaMa . Then you can improve the inference speed by using https://github.com/turboderp/exllama .

If you prefer to use an "instruct" model à la ChatGPT (i.e. that does not need few-shot learning to output good results) you can use something like this: https://huggingface.co/TheBloke/Wizard-Vicuna-30B-Uncensored...

AutoGPTQ: an attempt at standardizing GPTQ-for-LLaMa and turning it into a library that is easier to install and use, and that supports more models. https://github.com/PanQiWei/AutoGPTQ

Just use the script: https://github.com/PanQiWei/AutoGPTQ/blob/main/examples/quantization/quant_with_alpaca.py

In the wild, people tend to use GTPQ quantization for pure GPU inference: https://github.com/PanQiWei/AutoGPTQ

And ggml's quant for CPU inference with some offload, which just got updated to a more GPTQ-like method days ago: https://github.com/ggerganov/llama.cpp/pull/1684

Some other runtimes like Apache TVM also have their own quant implementations: https://github.com/mlc-ai/mlc-llm

For training, 4-bit bitsandbytes is SOTA, as far as I know.

TBH I'm not sure why this November paper is being linked. Few are running 8 bit models when they could fit a better 3-5 bit model in the same memory pool.

Instead of integrating GPTQ-for-Lllama, use AutoGPTQ instead.