AI Model Name: Llama 3 70B "Built with Meta Llama 3" https://llama.meta.com/llama3/license/
How to quantize 70B model so it will fit on 2x4090 GPUs:
I tried EXL2, AutoAWQ, and SqueezeLLM and they all failed for different reasons (issues opened).
HQQ worked:
I rented a 4x GPU 1TB RAM ($19/hr) instance on runpod with 1024GB container and 1024GB workspace disk space. I think you only need 2x GPU with 80GB VRAM and 512GB+ system RAM so probably overpaid.
Note you need to fill in the form to get access to the 70B Meta weights.
You can copy/paste this on the console and it will just set up everything automatically:
apt update
apt install git-lfs vim -y
mkdir -p ~/miniconda3
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
~/miniconda3/bin/conda init bash
source ~/.bashrc
conda create -n hqq python=3.10 -y && conda activate hqq
git lfs install
git clone https://github.com/mobiusml/hqq.git
cd hqq
pip install torch
pip install .
pip install huggingface_hub[hf_transfer]
export HF_HUB_ENABLE_HF_TRANSFER=1
huggingface-cli login
Create quantize.py file by copy/pasting this into console:
echo "
import torch
model_id = 'meta-llama/Meta-Llama-3-70B-Instruct'
save_dir = 'cat-llama-3-70b-hqq'
compute_dtype = torch.bfloat16
from hqq.core.quantize import *
quant_config = BaseQuantizeConfig(nbits=4, group_size=64, offload_meta=True)
zero_scale_group_size = 128
quant_config['scale_quant_params']['group_size'] = zero_scale_group_size
quant_config['zero_quant_params']['group_size'] = zero_scale_group_size
from hqq.engine.hf import HQQModelForCausalLM, AutoTokenizer
model = HQQModelForCausalLM.from_pretrained(model_id)
from hqq.models.hf.base import AutoHQQHFModel
AutoHQQHFModel.quantize_model(model, quant_config=quant_config,
compute_dtype=compute_dtype)
AutoHQQHFModel.save_quantized(model, save_dir)
model = AutoHQQHFModel.from_quantized(save_dir)
model.eval()
" > quantize.py
Run script:
python quantize.py