OPEA/Falcon3-3B-Base-int4-sym-awq-inc

Model Details

This model is an int4 model with group_size 128 and symmetric quantization of tiiuae/Falcon3-3B-Base generated by intel/auto-round.

How To Use

INT4 Inference(CPU/HPU/CUDA)

from auto_round import AutoRoundConfig  ##must import for auto_round format
from transformers import AutoModelForCausalLM, AutoTokenizer

quantized_model_dir = "OPEA/Falcon3-3B-Base-int4-sym-awq-inc"
tokenizer = AutoTokenizer.from_pretrained(quantized_model_dir)
model = AutoModelForCausalLM.from_pretrained(
    quantized_model_dir,
    device_map="auto"
)
text = "How many r in strawberry? The answer is "
inputs = tokenizer(text, return_tensors="pt", return_token_type_ids=False).to(model.device)
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=50)[0]))
##INT4:
## How many r in strawberry? The answer is 10000000000000000000000000000000000000000000000000

##BF16:
## How many r in strawberry? The answer is 1.
## Q: How many r in strawberry?
## A: 1
## Q: How many r in strawberry?
## A: 1
## Q: How many r in strawberry?
## A: 1
## Q: How many r

text = "Which number is bigger, 9.8 or 9.11? The answer is"
##INT4:
## Which number is bigger, 9.8 or 9.11? The answer is 9.8.
## Q: Which number is bigger, 9.8 or 9.11?
## A: 9.8
## Q: Which number is bigger, 9.8 or 9.11?

##BF16:
## Which number is bigger, 9.8 or 9.11? The answer is 9.8.
## Q: Which number is bigger, 9.8 or 9.11?
## Options:
## - 9.8
## - 9.11
## Let's think step by step.

text = "Once upon a time,"
##INT4:
## Once upon a time, there was a little girl named Lily who loved to play in the garden. She would spend hours picking flowers, digging in the dirt, and watching the bees buzz around the colorful blooms. One day, while she was picking a bouquet of daisies, she

##BF16:
## Once upon a time, there was a little girl named Lily who loved to play with her toys. One day, she found a special box in her attic. The box was old and dusty, but when she opened it, she found a beautiful, shiny object inside. It

text = "There is a girl who likes adventure,"
##INT4:
## There is a girl who likes adventure, she likes to go on adventures with her friends. She likes to go on adventures with her friends because she likes to go on adventures with her friends. She likes to go on adventures with her friends because she likes to go on adventures with her friends.

##BF16:
## There is a girl who likes adventure, she likes to go on adventures with her friends. She likes to go on adventures with her friends because she likes to go on adventures with her friends. She likes to go on adventures with her friends because she likes to go on adventures with her friends.

Evaluate the model

pip3 install lm-eval==0.4.5

auto-round --model "OPEA/Falcon3-3B-Base-int4-sym-awq-inc" --eval --eval_bs 16  --tasks lambada_openai,hellaswag,piqa,winogrande,truthfulqa_mc1,openbookqa,boolq,arc_easy,arc_challenge,mmlu

Metric	BF16	INT4
Avg	0.5816	0.5751
mmlu	0.5192	0.5127
lambada_opeai	0.6109	0.6103
hellaswage	0.5357	0.5285
winogrande	0.6614	0.6638
piqa	0.7753	0.7715
truthfulqa_mc1	0.2803	0.2742
openbookqa	0.3560	0.3360
boolq	0.8040	0.7884
arc_easy	0.7896	0.7845
arc_challenge	0.4838	0.4812

Generate the model

Here is the sample command to generate the model.

auto-round  \
--model tiiuae/Falcon3-3B-Base \
--device 0 \
--group_size 128 \
--nsamples 512 \
--bits 4 \
--iter 2000 \
--disable_eval \
--model_dtype float16 \
--format 'auto_awq' \
--output_dir "./tmp_autoround" 

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

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 and Liu, Yi}, journal={arXiv preprint arXiv:2309.05516}, year={2023} }

arxiv github