基于hugging face的autogptq量化实践

1.量化并保存到本地的

python 复制代码

#导入库：
from transformers import AutoModelForCausalLM, AutoTokenizer, GPTQConfig
model_id = "facebook/opt-125m"

quantization_config = GPTQConfig(
     bits=4,
     group_size=128,
     dataset="c4",
     desc_act=False,
)

tokenizer = AutoTokenizer.from_pretrained(model_id)
quant_model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=quantization_config, device_map='auto')

出现问题：

1.关于hugging face无法下载模型、数据的问题

OSError: We couldn't connect to 'https://huggingface.co' to load this file, couldn't find it in the cached files and it looks like facebook/opt-125m is not the path to a directory containing a file named config.json.

以及ConnectionError: Couldn't reach 'allenai/c4' on the Hub (ConnectTimeout)

采用方法：在官网下载到本地。

模型：https://huggingface.co/facebook/opt-125m/tree/main

数据集：https://huggingface.co/datasets

完整代码：

python 复制代码

####实现基于hugging face的模型量化及保存
from transformers import AutoModelForCausalLM, AutoTokenizer, GPTQConfig

model_id = "/pytorch/opt-125m"
#可选择公开数据集量化
tokenizer = AutoTokenizer.from_pretrained(model_id)
gptq_config = GPTQConfig(bits=4, dataset = "c4", tokenizer=tokenizer)
#或者采用自定义数据集量化
dataset = ["auto-gptq 是一个基于 GPTQ 算法的易于使用的模型量化库，具有用户友好的 api。"]
quantization = GPTQConfig(bits=4, dataset = dataset, tokenizer=tokenizer)

#注意，quantization_config用于选择数据集，输出量化后的模型
quant_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto",quantization_config=quantization)
#输出量化后权重，验证是否量化了
# print(quant_model.model.decoder.layers[0].self_attn.q_proj.__dict__)
#测试量化后的模型
text = "My name is"
inputs = tokenizer(text, return_tensors="pt").to(0)

out = quant_model.generate(**inputs)
print(tokenizer.decode(out[0], skip_special_tokens=True))

examples = [
    tokenizer(
        "auto-gptq is an easy-to-use model quantization library with user-friendly apis, based on GPTQ algorithm."
    )
]
#保存量化模型：

quant_model.save_pretrained("/pytorch/AutoGPTQ-main/demo/opt-125m-gptq")
tokenizer.save_pretrained("/pytorch/AutoGPTQ-main/demo/opt-125m-gptq")

从hugging face已经量化好的模型加载到本地

python 复制代码

###加载hugging face Hub中已量化好的模型到本地，并测试其效果
from transformers import AutoTokenizer, AutoModelForCausalLM

# model_id = "TheBloke/Llama-2-7b-Chat-GPTQ"
model_id = "/pytorch/llama"
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)
print(model)
print(model.config.quantization_config.to_dict())
text = "Hello my name is"
inputs = tokenizer(text, return_tensors="pt").to(0)
out = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(out[0], skip_special_tokens=True))

参考：
colab文档关于autogptq量化模型实践
 hugging face官网
 github快速实践
 github高阶实践
 transformer bitsandbytes通过int4量化LLM
其他