dalai VS minimal-llama

I agree, I've definitely seen way more information about running image synthesis models like Stable Diffusion locally than I have LLMs. It's counterintuitive to me that Stable Diffusion takes less RAM than an LLM, especially considering it still needs the word vectors. Goes to show I know nothing.

I guess it comes down to the requirement of a very high end (or multiple) GPU that makes it impractical for most vs just running it in Colab or something.

Tho there are some efforts:

https://github.com/cocktailpeanut/dalai

If you're just looking to play with something locally for the first time, this is the simplest project I've found and has a simple web UI: https://github.com/cocktailpeanut/dalai

It works for 7B/13B/30B/65B LLaMA and Alpaca (fine-tuned LLaMA which definitely works better). The smaller models at least should run on pretty much any computer.

I am not against easy mode options dude, for example I used to run GANs through command line. I replaced them with Upscayl when I found it. Convenience is king after all. Something about this one isn't right though. They are advertising it as a model they built meanwhile their own github show it to be a frontend of LLAMA. Why aren't they honest about it? Why use bots to spam about it? This causes me to not trust the executable they share to 1 to 1 compliation of the source code neither. I would still recommend looking for more decent alternatives. Btw, running it directly isn't that complicated

https://github.com/ggerganov/llama.cpp https://github.com/oobabooga/text-generation-webui https://github.com/mlc-ai/mlc-llm https://github.com/cocktailpeanut/dalai https://github.com/ido-pluto/catai (this is super easy to install but it doesnt provide an api or have integration with langchain)

This was easy to get running: https://github.com/cocktailpeanut/dalai with alpaca 13B (on my 16GB or ram)

I had it running before with Dalai (https://github.com/cocktailpeanut/dalai) but have since moved to using the browser based WebGPU method (https://mlc.ai/web-llm/) which uses Vicuna 7B and is quite good.

https://github.com/nsarrazin/serge let's you pick which model and runs in a container. For API https://github.com/cocktailpeanut/dalai looks super promising.

I can't edit my comment now, but it's 30B that needs 18GB of VRAM.

LLaMA-13B, GPT-3 175B level, only needs 10GB of VRAM with the GPTQ 4bit quantization.

>do you think there's anything left to trim? like weight pruning, or LoRA, or I dunno, some kind of Huffman coding scheme that lets you mix 4-bit, 2-bit and 1-bit quantizations?

Absolutely. The GPTQ paper claims negligible output quality loss with 3-bit quantization. The GPTQ-for-LLaMA repo supports 3-bit quantization and inference. So this extra 25% savings is already possible.

As of right GPTQ-for-LLaMA is using a VRAM hungry attention method. Flash attention will reduce the requirements for 7B to 4GB and possibly fit 30B with a 2048 context window into 16GB, all before stacking 3-bit.

Pruning is a possibility but I'm not aware of anyone working on it yet.

LoRa has already been implemented. See https://github.com/zphang/minimal-llama#peft-fine-tuning-wit...