Qwen量化脚本run_gptq.py解析

代码路径 https://github.com/QwenLM/Qwen/
run_gptq.py路径 https://github.com/QwenLM/Qwen/blob/main/run_gptq.py

代码解析：

import argparse
import json
from typing import Dict
import loggingimport 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#其中json文件格式如下
# [
#   {
#     "id": "identity_0",
#     "conversations": [
#       {
#         "from": "user",
#         "value": "xxxx"
#       },
#       {
#         "from": "assistant",
#         "value": "xxx"
#       }
#     ]
#   },
#   {
#     "id": "identity_1",
#     "conversations": [
#       {
#         "from": "user",
#         "value": "xxx"
#       },
#       {
#         "from": "assistant",
#         "value": "xxx"
#       }
#     ]
#   },
# ]def preprocess(sources,tokenizer: transformers.PreTrainedTokenizer,max_len: int,system_message: str = "You are a helpful assistant."
) -> Dict:"""preprocess函数接收一个包含对话数据的json列表作为输入,\n通过调用transformers库中的tokenizer对数据进行编码,\n并按照特定格式构建输入ID序列和目标ID序列.\n返回一个包含预处理数据的列表,这些数据已转换为PyTorch张量,适合于后续模型训练或推断"""#roles字典：为对话中的角色（"user"和"assistant"）分配特殊的前缀标签，用于区分对话双方roles = {"user": "<|im_start|>user", "assistant": "<|im_start|>assistant"}#im_start和im_end：指定tokenizer中im_start_id和im_end_id对应的整数ID。im_start = tokenizer.im_start_idim_end = tokenizer.im_end_id#nl_tokens：存储tokenizer处理换行符\n得到的输入ID序列。nl_tokens = tokenizer('\n').input_ids#_system、_user和_assistant：分别存储经过tokenizer处理后的"system"、"user"和"assistant"标签及其后的换行符对应的输入ID序列。_system = tokenizer('system').input_ids + nl_tokens_user = tokenizer('user').input_ids + nl_tokens_assistant = tokenizer('assistant').input_ids + nl_tokens# Apply prompt templates 定义空列表data，用于存放预处理后的数据样本data = []# input_ids, targets = [], []#遍历输入数据sources中的每个样本（source）for i, source in enumerate(sources):source = source["conversations"]#检查首个对话是否由用户发起（即source[0]["from"]是否为"user"），如果不是，则从源数据中移除首个对话。#过滤无效的identityif roles[source[0]["from"]] != roles["user"]:source = source[1:]#初始化空列表input_id和target，分别用于存储当前样本的输入ID序列和目标ID序列input_id, target = [], []#添加系统消息：将系统消息（包含system_message内容）转换为ID序列，添加到input_id和target中。system = [im_start] + _system + tokenizer(system_message).input_ids + [im_end] + nl_tokensinput_id += system#target中的非关键部分（如系统标签和消息内容）用IGNORE_TOKEN_ID填充。target += [im_start] + [IGNORE_TOKEN_ID] * (len(system)-3) + [im_end] + nl_tokensassert len(input_id) == len(target)#遍历源数据中的每个对话（sentence）for j, sentence in enumerate(source):# 提取角色和消息内容，并转换为ID序列role = roles[sentence["from"]]_input_id = tokenizer(role).input_ids + nl_tokens + \tokenizer(sentence["value"]).input_ids + [im_end] + nl_tokens#添加到input_id中input_id += _input_id#根据角色类型，生成对应_target的目标ID序列，_target只提取assistant的对话内容，忽略user的对话内容。if role == '<|im_start|>user':#若角色为"user"，则目标ID序列仅包含开始标签和结束标签，用忽略ID填充对话内容。_target = [im_start] + [IGNORE_TOKEN_ID] * (len(_input_id)-3) + [im_end] + nl_tokens#若角色为"assistant"，则目标ID序列包含开始标签、忽略ID填充（仅对角色标签）、对话内容（不包括角色标签和结束标签）、结束标签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_tokenselse:raise NotImplementedErrortarget += _targetassert len(input_id) == len(target)#截取并转换为张量：#截取input_id和target至最大长度max_leninput_id = torch.tensor(input_id[:max_len], dtype=torch.int)target = torch.tensor(target[:max_len], dtype=torch.int)#创建一个字典，包含键input_ids（存储输入张量）和attention_mask（等于输入张量，用于指示非填充位置）。将该字典添加到data列表中data.append(dict(input_ids=input_id, attention_mask=input_id.ne(tokenizer.pad_token_id)))return dataif __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 badstatic_groups=False,sym=True,true_sequential=True,model_name_or_path=None,model_file_base_name="model")#使用AutoTokenizer类从给定路径args.model_name_or_path加载预训练的tokenizertokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, trust_remote_code=True)tokenizer.pad_token_id = tokenizer.eod_id#加载json数据文件，调用process函数预处理数据，返回处理后的数据data = preprocess(json.load(open(args.data_path)), tokenizer, args.max_len)#加载预训练的模型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")#对模型进行量化,不在GPU上缓存示例数据model.quantize(data, cache_examples_on_gpu=False)#保存量化后的模型model.save_quantized(args.out_path, use_safetensors=True)#将tokenizer保存到输出路径tokenizer.save_pretrained(args.out_path)