catid/cat-llama-3-8b-awq-q128-w4-gemm

AI Model Name: Llama 3 8B "Built with Meta Llama 3" https://llama.meta.com/llama3/license/

This is the result of running AutoAWQ to quantize the LLaMA-3 8B model to ~4 bits/parameter.

To launch an OpenAI-compatible API endpoint on your Linux server:

git lfs install
git clone https://huggingface.co/catid/cat-llama-3-8b-awq-q128-w4-gemm

conda create -n vllm8 python=3.10 -y && conda activate vllm8

pip install -U git+https://github.com/vllm-project/vllm.git@a134ef6

python -m vllm.entrypoints.openai.api_server --model cat-llama-3-8b-awq-q128-w4-gemm

To use 2 GPUs add --tensor-parallel-size 2 --gpu-memory-utilization 0.95:

python -m vllm.entrypoints.openai.api_server --model cat-llama-3-8b-awq-q128-w4-gemm --tensor-parallel-size 2 --gpu-memory-utilization 0.95

My personal TextWorld common-sense reasoning benchmark ( https://github.com/catid/textworld_llm_benchmark ) results for this model:

cat-llama-3-8b-awq-q128-w4-gemm : Average Score: 2.02 ± 0.29
Mixtral 8x7B : Average Score: 2.22 ± 0.33
GPT 3.5 : Average Score: 2.8 ± 1.69

This is very respectable for a relatively small model!