Model Recipe Details
This is an int4 model recipe with group_size 128 of facebook/opt-1.3b generated by intel/auto-round. Inference of this model is compatible with AutoGPTQ's Kernel.
Quantize the model
Here is the sample command to reproduce the model
pip install auto-round
auto-round
--model facebook/opt-1.3b \
--device 0 \
--group_size 128 \
--bits 4 \
--iters 1000 \
--nsamples 512 \
--format 'auto_gptq' \
--minmax_lr 2e-3 \
--disable_quanted_input \
--output_dir "./tmp_autoround" \
How to use
INT4 Inference with IPEX on Intel CPU
Install the latest Intel Extension for Pytorch and Intel Neural Compressor
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install intel_extension_for_pytorch
pip install neural_compressor_pt
from transformers import AutoTokenizer
from neural_compressor.transformers import AutoModelForCausalLM
## note: use quantized model directory name below
model_name_or_path="./tmp_autoround/<model directory name>"
q_model = AutoModelForCausalLM.from_pretrained(model_name_or_path)
prompt = "Once upon a time, a little girl"
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=True)
print(tokenizer.decode(q_model.generate(**tokenizer(prompt, return_tensors="pt").to(q_model.device),max_new_tokens=50)[0]))
## Once upon a time, a little girl was born. She was a beautiful little girl, with a beautiful smile. She was a little girl who loved to play. She was a little girl who loved to sing.She was a little girl who loved to dance.
INT4 Inference on Intel Gaudi Accelerator
docker image with Gaudi Software Stack is recommended. More details can be found in Gaudi Guide.
import habana_frameworks.torch.core as htcore
from neural_compressor.torch.quantization import load
from transformers import AutoTokenizer, AutoModelForCausalLM
## note: use quantized model directory name below
model_name_or_path="./tmp_autoround/<model directory name>"
model = load(
model_name_or_path=model_name_or_path,
format="huggingface",
device="hpu"
)
prompt = "Once upon a time, a little girl"
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
print(tokenizer.decode(model.generate(**tokenizer(prompt, return_tensors="pt").to("hpu"),max_new_tokens=50)[0]))
Accuracy Result
| Metric |
FP16 |
INT4 |
|---|---|---|
| Avg. | 0.4405 | 0.4315 |
| mmlu | 0.2502 | 0.2510 |
| lambada_openai | 0.5789 | 0.5389 |
| hellaswag | 0.4153 | 0.4076 |
| winogrande | 0.5927 | 0.5848 |
| piqa | 0.7165 | 0.7067 |
| truthfulqa_mc1 | 0.2362 | 0.2338 |
| openbookqa | 0.2320 | 0.2140 |
| boolq | 0.5765 | 0.5786 |
| arc_easy | 0.5724 | 0.5673 |
| arc_challenge | 0.2346 | 0.2321 |
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:
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} }