git clone https://github.com/intel/auto-round
cd auto-round
pip install -r requirements.txt
cd examples/language-modeling
pip install -r requirements.txt
python3 main.py \
--model_name  Intel/neural-chat-7b-v3-3 \
--device 0 \
--group_size 128 \
--bits 4 \
--iters 1000 \
--enable_minmax_tuning \
--minmax_lr 0.002 \
--deployment_device 'gpu' \
--scale_dtype 'fp32' \
--disable_quanted_input \
--eval_bs 32 \
--output_dir "./tmp_autoround" \
--amp 

Use the model

INT4 Inference with ITREX on CPU

Install the latest intel-extension-for-transformers

python

from intel_extension_for_transformers.transformers import AutoModelForCausalLM
from transformers import AutoTokenizer
quantized_model_dir = "Intel/neural-chat-7b-v3-3-int4-inc"
model = AutoModelForCausalLM.from_pretrained(quantized_model_dir,
                                             device_map="auto",
                                             trust_remote_code=False,
                                             use_neural_speed=False,
                                             )
tokenizer = AutoTokenizer.from_pretrained(quantized_model_dir, use_fast=True)
print(tokenizer.decode(model.generate(**tokenizer("There is a girl who likes adventure,", return_tensors="pt").to(model.device),max_new_tokens=50)[0]))
"""
 <s> There is a girl who likes adventure, and she is a bit of a daredevil. She loves to travel and explore new places. She is always looking for the next thrill, whether it be skydiving, bungee jumping, or even just hiking up a mountain
"""

INT4 Inference with AutoGPTQ

Install AutoGPTQ from source first

python

from transformers import AutoModelForCausalLM, AutoTokenizer
quantized_model_dir = "Intel/neural-chat-7b-v3-3-int4-inc"
model = AutoModelForCausalLM.from_pretrained(quantized_model_dir,
                                             device_map="auto",
                                             trust_remote_code=False,
                                             )
tokenizer = AutoTokenizer.from_pretrained(quantized_model_dir, use_fast=True)
print(tokenizer.decode(model.generate(**tokenizer("There is a girl who likes adventure,", return_tensors="pt").to(model.device),max_new_tokens=50)[0]))

Evaluate the model

Install lm-eval-harness from source, we used the git id f3b7917091afba325af3980a35d8a6dcba03dc3f

lm_eval  --model hf --model_args pretrained="Intel/neural-chat-7b-v3-3-int4-inc",autogptq=True,gptq_use_triton=True --device cuda:0 --tasks lambada_openai,hellaswag,piqa,winogrande,truthfulqa_mc1,openbookqa,boolq,rte,arc_easy,arc_challenge,mmlu  --batch_size 128

Metric	FP16	INT4
Avg.	0.6778	0.6748
mmlu	0.5993	0.5926
lambada_openai	0.7303	0.7370
hellaswag	0.6639	0.6559
winogrande	0.7632	0.7735
piqa	0.8101	0.8074
truthfulqa_mc1	0.4737	0.4737
openbookqa	0.3880	0.3680
boolq	0.8694	0.8694
rte	0.7581	0.7509
arc_easy	0.8266	0.8249
arc_challenge	0.5734	0.5691

Ethical Considerations and Limitations

The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs.

Therefore, before deploying any applications of the model, developers should perform safety testing.

Caveats and Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.

Here are a couple of useful links to learn more about Intel's AI software:

Intel Neural Compressor link
Intel Extension for Transformers link

Disclaimer

The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.

Cite

@article{cheng2023optimize, title={Optimize weight rounding via signed gradient descent for the quantization of llms}, author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao}, journal={arXiv preprint arXiv:2309.05516}, year={2023} }

arxiv github