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?
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 now 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...
LLaMA-65B 4bit needs 36 GB of VRAM, but far exceeds GPT-3's capabilities and even takes on PaLM 540B.
See: https://github.com/oobabooga/text-generation-webui/wiki/LLaM... for 4bit setup instructions
See also: The case for 4-bit precision, which shows effectively no output quality reduction for these 4bit quantization methods (and considerable speedup) https://arxiv.org/abs/2212.09720