【深度学习】Qwen大模型量化

1. 前言 

        在微调Qwen大模型之后，需要对模型进行后量化。在Qwen预构建的Docker镜像基础上新作镜像作为量化的环境。镜像成为容器之后，将需要的文件使用docker cp命令复制到容器中。整个量化过程在Qwen的github文档中给出了详细的介绍，镜像制作过程可参考【深度学习】微调Qwen1.8B_qwen1.8 fintune-优快云博客

 2.量化过程

         Qwen的介绍中给出了详细的量化教程。可访问千问github中的量化微调后模型章节查看。

2.1 安装 auto_gptq

docker run --gpus all -p 9519:9519-v /ssd/dongzhenheng/Work/Merged_Model/Merged_model_1:/data/shared/Qwen/Qwen-Chat/ -d qwenllm/qwen:cu121_V1

pip install auto-gptq optimum

 注意：使用auto-gptq做量化的时候，会报错。

解决办法如下：

#拉取代码
git clone https://github.com/wangitu/unpadded-AutoGPTQ.git && cd unpadded-AutoGPTQ
#安装
pip install -v -e .

2.2 准备数据

        准备校准集。你可以重用微调你的数据，或者按照微调相同的方式准备其他数据。在实践中使用的是微调的数据集。

[
  {
    "id": "identity_0",
    "conversations": [
      {
        "from": "user",
        "value": "你好"
      },
      {
        "from": "assistant",
        "value": "我是一个语言模型，我叫通义千问。"
      }
    ]
  }
]

2.3 量化

        量化代码在千问github的run_gptq.py中已经提供。

import argparse
import json
from typing import Dict
import logging

import torch
import transformers
from transformers import AutoTokenizer
from transformers.trainer_pt_utils import LabelSmoother
from auto_gptq import AutoGPTQForCausalLM, BaseQuantizeConfig
IGNORE_TOKEN_ID = LabelSmoother.ignore_index

def preprocess(
    sources,
    tokenizer: transformers.PreTrainedTokenizer,
    max_len: int,
    system_message: str = "You are a helpful assistant."
) -> Dict:
    roles = {"user": "<|im_start|>user", "assistant": "<|im_start|>assistant"}

    im_start = tokenizer.im_start_id
    im_end = tokenizer.im_end_id
    nl_tokens = tokenizer('\n').input_ids
    _system = tokenizer('system').input_ids + nl_tokens
    _user = tokenizer('user').input_ids + nl_tokens
    _assistant = tokenizer('assistant').input_ids + nl_tokens

    # Apply prompt templates
    data = []
    # input_ids, targets = [], []
    for i, source in enumerate(sources):
        source = source["conversations"]
        if roles[source[0]["from"]] != roles["user"]:
            source = source[1:]

        input_id, target = [], []
        system = [im_start] + _system + tokenizer(system_message).input_ids + [im_end] + nl_tokens
        input_id += system
        target += [im_start] + [IGNORE_TOKEN_ID] * (len(system)-3) + [im_end] + nl_tokens
        assert len(input_id) == len(target)
        for j, sentence in enumerate(source):
            role = roles[sentence["from"]]
            _input_id = tokenizer(role).input_ids + nl_tokens + \
                tokenizer(sentence["value"]).input_ids + [im_end] + nl_tokens
            input_id += _input_id
            if role == '<|im_start|>user':
                _target = [im_start] + [IGNORE_TOKEN_ID] * (len(_input_id)-3) + [im_end] + nl_tokens
            elif role == '<|im_start|>assistant':
                _target = [im_start] + [IGNORE_TOKEN_ID] * len(tokenizer(role).input_ids) + \
                    _input_id[len(tokenizer(role).input_ids)+1:-2] + [im_end] + nl_tokens
            else:
                raise NotImplementedError
            target += _target
        assert len(input_id) == len(target)
        input_id = torch.tensor(input_id[:max_len], dtype=torch.int)
        target = torch.tensor(target[:max_len], dtype=torch.int)
        data.append(dict(input_ids=input_id, attention_mask=input_id.ne(tokenizer.pad_token_id)))

    return data


if __name__ == "__main__":
    parser = argparse.ArgumentParser("Model Quantization using AutoGPTQ")
    parser.add_argument("--model_name_or_path", type=str, help="model path")
    parser.add_argument("--data_path", type=str, help="calibration data path")
    parser.add_argument("--out_path", type=str, help="output path of the quantized model")
    parser.add_argument("--max_len", type=int, default=8192, help="max length of calibration data")
    parser.add_argument("--bits", type=int, default=4, help="the bits of quantized model. 4 indicates int4 models.")
    parser.add_argument("--group-size", type=int, default=128, help="the group size of quantized model")
    args = parser.parse_args()
    
    quantize_config = BaseQuantizeConfig(
        bits=args.bits,
        group_size=args.group_size,
        damp_percent=0.01,
        desc_act=False,  # set to False can significantly speed up inference but the perplexity may slightly bad
        static_groups=False,
        sym=True,
        true_sequential=True,
        model_name_or_path=None,
        model_file_base_name="model"
    )

    tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, trust_remote_code=True)
    tokenizer.pad_token_id = tokenizer.eod_id
    data = preprocess(json.load(open(args.data_path)), tokenizer, args.max_len)

    model = AutoGPTQForCausalLM.from_pretrained(args.model_name_or_path, quantize_config, device_map="auto", trust_remote_code=True)

    logging.basicConfig(
        format="%(asctime)s %(levelname)s [%(name)s] %(message)s", level=logging.INFO, datefmt="%Y-%m-%d %H:%M:%S"
    )
    model.quantize(data, cache_examples_on_gpu=False)

    model.save_quantized(args.out_path, use_safetensors=True)
    tokenizer.save_pretrained(args.out_path)

        量化脚本如下：

#!/bin/bash

YOUR_LORA_MODEL_PATH="/path/to/your/lora_model"
DATA="/path/to/your/data"
OUTPUT_PATH="/path/to/your/output"
BITS=4  # Set the desired value (4 for int4; 8 for int8)

python run_gptq.py \
    --model_name_or_path "$YOUR_LORA_MODEL_PATH" \
    --data_path "$DATA" \
    --out_path "$OUTPUT_PATH" \
    --bits "$BITS"

量化过程：

 2.4 生成文件

        接下来, 将原模型中所有 *.py, *.cu, *.cpp 文件和 generation_config.json 文件复制到输出模型目录下。同时，使用官方对应版本的量化模型的 config.json 文件覆盖输出模型目录下的文件 (例如, 如果你微调了 Qwen-7B-Chat和--bits 4, 那么你可以从 Qwen-7B-Chat-Int4 仓库中找到对应的config.json )。 并且，你需要将 gptq.safetensors 重命名为 model.safetensors。

2.5 模型测试 

from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers.generation import GenerationConfig

tokenizer = AutoTokenizer.from_pretrained("/path/to/your/model", trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    "/path/to/your/model",
    device_map="auto",
    trust_remote_code=True
).eval()

response, history = model.chat(tokenizer, "你好", history=None)
print(response)