Llama-3.1-8B-Instruct_gemlite-ao_a16w4_gs_128_pack_16bit

This is an <a href="https://github.com/mobiusml/hqq/">HQQ</a> all 4-bit (group-size=128) quantized <a href="https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct"> Meta-Llama-3-8B-Instruct</a> model, via <a href="https://github.com/pytorch/ao/">TorchAO</a> and <a href="https://github.com/mobiusml/gemlite/">GemLite</a> as a backend.

Usage

First, install the dependecies:

pip install torchao;
pip install git+https://github.com/mobiusml/gemlite.git;

Then you can use the sample code below:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, HqqConfig 

model_id = "mobiuslabsgmbh/Llama-3.1-8B-Instruct_gemlite-ao_a16w4_gs_128_pack_16bit"
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
    device_map='cuda', 
)

tokenizer = AutoTokenizer.from_pretrained(model_id)

Use in <a href="https://github.com/vllm-project/vllm/">vLLM</a>:

from vllm import LLM
from vllm.sampling_params import SamplingParams

model_id = "mobiuslabsgmbh/Llama-3.1-8B-Instruct_gemlite-ao_a16w4_gs_128_pack_16bit"

llm = LLM(model=model_id, max_model_len=4096)
sampling_params = SamplingParams(temperature=0.8, top_p=0.95, max_tokens=1024)
outputs = llm.generate(["What is the capital of Germany?"], sampling_params)
print(outputs[0].outputs[0].text)