glm-4-9b-chat-int4-sym-lm-head-inc

Model Details

This model is an int4 model with group_size 128 and quantized lm-head of THUDM/glm-4-9b-chat generated by intel/auto-round. For GPTQ format, please load the model with revision d1bb095

How To Use

INT4 Inference on CPU/CUDA/HPU

from auto_round import AutoRoundConfig ## must import for auto-round format
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

quantized_model_dir = "OPEA/glm-4-9b-chat-int4-inc"

model = AutoModelForCausalLM.from_pretrained(quantized_model_dir,
                                             device_map=backend.split(':')[0],                                                                                                torch_dtype=torch.float16,
                                             quantization_config=quantization_config,
                                             trust_remote_code=True 
                                            ).eval()

tokenizer = AutoTokenizer.from_pretrained(quantized_model_dir, trust_remote_code=True)
query = "请介绍一下智谱华章科技有限公司"
inputs = tokenizer.apply_chat_template([{"role": "user", "content": query}],
                                       add_generation_prompt=True,
                                       tokenize=True,
                                       return_tensors="pt",
                                       return_dict=True
                                       )
inputs = inputs.to(model.device)

gen_kwargs = {"max_length": 50, "do_sample": False, "top_k": 1}##change this to follow official usage
with torch.no_grad():
    outputs = model.generate(**inputs, **gen_kwargs)
    outputs = outputs[:, inputs['input_ids'].shape[1]:]
    print(tokenizer.decode(outputs[0], skip_special_tokens=True))
    

Evaluate the model

pip3 install lm-eval==0.4.5

pip3 install langdetect,immutabledict,antlr4-python3-runtime==4.11

auto-round --eval --eval_bs 16 --tasks leaderboard_ifeval,leaderboard_mmlu_pro,gsm8k,lambada_openai,hellaswag,piqa,winogrande,truthfulqa_mc1,openbookqa,boolq,arc_easy,arc_challenge,cmmlu,ceval-valid

Generate the model

Here is the sample command to generate the model.

auto-round \
--model THUDM/glm-4-9b-chat \
--iter 1000 \
--nsamples 512 \
--disable_eval \
--quant_lm_head \
--format "auto_round"
--output_dir "./tmp_autoround" 

copy all the *.py file to the quantized_model folder

For gptq format, need to add "block_name_to_quantize":"transformer.encoder.layers" to config.json, we only tested it on transformers==4.46.1

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