您的位置:首页 > 科技 > 能源 > 网络编程基础_163建筑网站_百度助手手机下载_附近电脑培训班零基础

网络编程基础_163建筑网站_百度助手手机下载_附近电脑培训班零基础

2025/4/7 18:48:38 来源:https://blog.csdn.net/CITY_OF_MO_GY/article/details/147013697  浏览:    关键词:网络编程基础_163建筑网站_百度助手手机下载_附近电脑培训班零基础
网络编程基础_163建筑网站_百度助手手机下载_附近电脑培训班零基础

       2025年,大模型爆发元年,各种各样的大模型、框架、工具层出不穷,不断刷新人们应用大模型的门槛,短短10行代码,就能完成“加载模型+加载数据集+推理+强化学习”的全流程训练,但其底层的运行机制也被高度抽象的接口给掩盖,大模型那些超参是如何在代码中起作用的我们不得而知,而这种似见非见的感觉总让人感觉惴惴不安;

       而最近,我发现了一个宝藏项目,叫MiniMind,作者通过Pytorch自己构建了大模型框架,自己实现的模型的预训练、SFT、Lora、RLHF、蒸馏等一系列的操作,并且开源了相关的训练集数据,同时上传了自己的各种预训练模型,感兴趣的同学可以自己到Github尝试,这里也致敬该作者的开源精神;那接下来我就以此项目为基础,分享一些我在该项目中学习到的知识,通过我的经验,将代码中的‘故事’汇集成一个系列讲述给屏幕前的你;

       关于项目的环境搭建和安装运行在README中以及说的非常清楚了,这里就不再过多的去浪费时间在这些上面,今天就以项目的推理代码‘eval_model.py’为起点,手撕模型的推理代码,探索大模型推理的全流程;

一、俯视整个推理框架

       我们先从eval_model.py代码看起,拆解eval_model.py,整体来看一下该项目的推理过程都包含了哪些模块;

# eval_model.pyimport argparse
import random
import time
import numpy as np
import torch
import warnings
from transformers import AutoTokenizer, AutoModelForCausalLM
from model.model import MiniMindLM
from model.LMConfig import LMConfig
from model.model_lora import *warnings.filterwarnings('ignore')def init_model(args):tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')if args.load == 0:moe_path = '_moe' if args.use_moe else ''modes = {0: 'pretrain', 1: 'full_sft', 2: 'rlhf', 3: 'reason'}# ckp = f'./{args.out_dir}/{modes[args.model_mode]}_{args.dim}{moe_path}.pth'ckp = f'../Weights/MiniMind2-PyTorch/{modes[args.model_mode]}_{args.dim}{moe_path}.pth'model = MiniMindLM(LMConfig(dim=args.dim,n_layers=args.n_layers,max_seq_len=args.max_seq_len,use_moe=args.use_moe))state_dict = torch.load(ckp, map_location=args.device)model.load_state_dict({k: v for k, v in state_dict.items() if 'mask' not in k}, strict=True)if args.lora_name != 'None':apply_lora(model)load_lora(model, f'./{args.out_dir}/lora/{args.lora_name}_{args.dim}.pth')else:transformers_model_path = '../Weights/MiniMind2'tokenizer = AutoTokenizer.from_pretrained(transformers_model_path)model = AutoModelForCausalLM.from_pretrained(transformers_model_path, trust_remote_code=True)print(f'MiniMind模型参数量: {sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e6:.2f}M(illion)')return model.eval().to(args.device), tokenizerdef get_prompt_datas(args):if args.model_mode == 0:# pretrain模型的接龙能力(无法对话)prompt_datas = ['马克思主义基本原理','人类大脑的主要功能','万有引力原理是','世界上最高的山峰是','二氧化碳在空气中','地球上最大的动物有','杭州市的美食有']else:if args.lora_name == 'None':# 通用对话问题prompt_datas = ['请介绍一下自己。','你更擅长哪一个学科?','鲁迅的《狂人日记》是如何批判封建礼教的?','我咳嗽已经持续了两周,需要去医院检查吗?','详细的介绍光速的物理概念。','推荐一些杭州的特色美食吧。','请为我讲解“大语言模型”这个概念。','如何理解ChatGPT?','Introduce the history of the United States, please.']else:# 特定领域问题lora_prompt_datas = {'lora_identity': ["你是ChatGPT吧。","你叫什么名字?","你和openai是什么关系?"],'lora_medical': ['我最近经常感到头晕,可能是什么原因?','我咳嗽已经持续了两周,需要去医院检查吗?','服用抗生素时需要注意哪些事项?','体检报告中显示胆固醇偏高,我该怎么办?','孕妇在饮食上需要注意什么?','老年人如何预防骨质疏松?','我最近总是感到焦虑,应该怎么缓解?','如果有人突然晕倒,应该如何急救?'],}prompt_datas = lora_prompt_datas[args.lora_name]return prompt_datas# 设置可复现的随机种子
def setup_seed(seed):random.seed(seed)np.random.seed(seed)torch.manual_seed(seed)torch.cuda.manual_seed(seed)torch.cuda.manual_seed_all(seed)torch.backends.cudnn.deterministic = Truetorch.backends.cudnn.benchmark = Falsedef main():parser = argparse.ArgumentParser(description="Chat with MiniMind")parser.add_argument('--lora_name', default='None', type=str)parser.add_argument('--out_dir', default='out', type=str)parser.add_argument('--temperature', default=0.85, type=float)parser.add_argument('--top_p', default=0.85, type=float)parser.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu', type=str)# 此处max_seq_len(最大允许输入长度)并不意味模型具有对应的长文本的性能,仅防止QA出现被截断的问题# MiniMind2-moe (145M):(dim=640, n_layers=8, use_moe=True)# MiniMind2-Small (26M):(dim=512, n_layers=8)# MiniMind2 (104M):(dim=768, n_layers=16)parser.add_argument('--dim', default=512, type=int)parser.add_argument('--n_layers', default=8, type=int)parser.add_argument('--max_seq_len', default=8192, type=int)parser.add_argument('--use_moe', default=False, type=bool)# 携带历史对话上下文条数# history_cnt需要设为偶数,即【用户问题, 模型回答】为1组;设置为0时,即当前query不携带历史上文# 模型未经过外推微调时,在更长的上下文的chat_template时难免出现性能的明显退化,因此需要注意此处设置parser.add_argument('--history_cnt', default=0, type=int)parser.add_argument('--stream', default=True, type=bool)parser.add_argument('--load', default=0, type=int, help="0: 原生torch权重,1: transformers加载")parser.add_argument('--model_mode', default=1, type=int,help="0: 预训练模型,1: SFT-Chat模型,2: RLHF-Chat模型,3: Reason模型")args = parser.parse_args()model, tokenizer = init_model(args)prompts = get_prompt_datas(args)test_mode = int(input('[0] 自动测试\n[1] 手动输入\n'))messages = []for idx, prompt in enumerate(prompts if test_mode == 0 else iter(lambda: input('👶: '), '')):setup_seed(random.randint(0, 2048))# setup_seed(2025)  # 如需固定每次输出则换成【固定】的随机种子if test_mode == 0: print(f'👶: {prompt}')messages = messages[-args.history_cnt:] if args.history_cnt else []messages.append({"role": "user", "content": prompt})new_prompt = tokenizer.apply_chat_template(messages,tokenize=False,add_generation_prompt=True)[-args.max_seq_len + 1:] if args.model_mode != 0 else (tokenizer.bos_token + prompt)answer = new_promptwith torch.no_grad():x = torch.tensor(tokenizer(new_prompt)['input_ids'], device=args.device).unsqueeze(0)outputs = model.generate(x,eos_token_id=tokenizer.eos_token_id,max_new_tokens=args.max_seq_len,temperature=args.temperature,top_p=args.top_p,stream=True,pad_token_id=tokenizer.pad_token_id)print('🤖️: ', end='')try:if not args.stream:print(tokenizer.decode(outputs.squeeze()[x.shape[1]:].tolist(), skip_special_tokens=True), end='')else:history_idx = 0for y in outputs:answer = tokenizer.decode(y[0].tolist(), skip_special_tokens=True)if (answer and answer[-1] == '�') or not answer:continueprint(answer[history_idx:], end='', flush=True)history_idx = len(answer)except StopIteration:print("No answer")print('\n')messages.append({"role": "assistant", "content": answer})if __name__ == "__main__":main()

       该脚本中一共定义了4个函数API,分别是:

函数API作用
init_model()加载模型和分词器
get_prompt_datas()定义自动测试的提示词
setup_seed()设定可复现的随机种子
main()指定超参、推理、后处理

1.1 main() 

def main():parser = argparse.ArgumentParser(description="Chat with MiniMind")parser.add_argument('--lora_name', default='None', type=str)parser.add_argument('--out_dir', default='out', type=str)parser.add_argument('--temperature', default=0.85, type=float)parser.add_argument('--top_p', default=0.85, type=float)parser.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu', type=str)# 此处max_seq_len(最大允许输入长度)并不意味模型具有对应的长文本的性能,仅防止QA出现被截断的问题# MiniMind2-moe (145M):(dim=640, n_layers=8, use_moe=True)# MiniMind2-Small (26M):(dim=512, n_layers=8)# MiniMind2 (104M):(dim=768, n_layers=16)parser.add_argument('--dim', default=512, type=int)parser.add_argument('--n_layers', default=8, type=int)parser.add_argument('--max_seq_len', default=8192, type=int)parser.add_argument('--use_moe', default=False, type=bool)# 携带历史对话上下文条数# history_cnt需要设为偶数,即【用户问题, 模型回答】为1组;设置为0时,即当前query不携带历史上文# 模型未经过外推微调时,在更长的上下文的chat_template时难免出现性能的明显退化,因此需要注意此处设置parser.add_argument('--history_cnt', default=0, type=int)parser.add_argument('--stream', default=True, type=bool)parser.add_argument('--load', default=0, type=int, help="0: 原生torch权重,1: transformers加载")parser.add_argument('--model_mode', default=1, type=int,help="0: 预训练模型,1: SFT-Chat模型,2: RLHF-Chat模型,3: Reason模型")args = parser.parse_args()model, tokenizer = init_model(args)prompts = get_prompt_datas(args)test_mode = int(input('[0] 自动测试\n[1] 手动输入\n'))messages = []for idx, prompt in enumerate(prompts if test_mode == 0 else iter(lambda: input('👶: '), '')):setup_seed(random.randint(0, 2048))# setup_seed(2025)  # 如需固定每次输出则换成【固定】的随机种子if test_mode == 0: print(f'👶: {prompt}')messages = messages[-args.history_cnt:] if args.history_cnt else []messages.append({"role": "user", "content": prompt})new_prompt = tokenizer.apply_chat_template(messages,tokenize=False,add_generation_prompt=True)[-args.max_seq_len + 1:] if args.model_mode != 0 else (tokenizer.bos_token + prompt)answer = new_promptwith torch.no_grad():x = torch.tensor(tokenizer(new_prompt)['input_ids'], device=args.device).unsqueeze(0)outputs = model.generate(x,eos_token_id=tokenizer.eos_token_id,max_new_tokens=args.max_seq_len,temperature=args.temperature,top_p=args.top_p,stream=True,pad_token_id=tokenizer.pad_token_id)print('🤖️: ', end='')try:if not args.stream:print(tokenizer.decode(outputs.squeeze()[x.shape[1]:].tolist(), skip_special_tokens=True), end='')else:history_idx = 0for y in outputs:answer = tokenizer.decode(y[0].tolist(), skip_special_tokens=True)if (answer and answer[-1] == '�') or not answer:continueprint(answer[history_idx:], end='', flush=True)history_idx = len(answer)except StopIteration:print("No answer")print('\n')messages.append({"role": "assistant", "content": answer})

1.1.1 超参

参数含义
lora_name指定Lora模型(本章未用到)
out_dir指定输出路径(本章未用到)
temperature用于控制生成文本随机性的超参数
top_p用于控制生成文本随机性的超参数
device指定算力设备
dim指定大模型的宽度,即token转化为向量的维度
n_layers指定transformer模块的个数
max_seq_len指定最大允许输入长度
use_moe使用MOE专家模型
history_cnt指定携带历史对话上下文条数
stream指定是否流式输出
load指定加载模型的方式,0: 原生torch权重,1: transformers加载
model_mode指定加载模型的类型,0: 预训练模型,1: SFT-Chat模型,2: RLHF-Chat模型,3: Reason模型

1.1.2 推理前

# 加载模型和分词器
model, tokenizer = init_model(args)# 加载/遍历提示词
prompts = get_prompt_datas(args)# 设置随机种子
setup_seed(random.randint(0, 2048))# 将提示词以字典的形式添加到messages
if test_mode == 0: print(f'👶: {prompt}')
messages = messages[-args.history_cnt:] if args.history_cnt else []
messages.append({"role": "user", "content": prompt})# 将messages通过对话模版输出
new_prompt = tokenizer.apply_chat_template(messages,tokenize=False,add_generation_prompt=True)[-args.max_seq_len + 1:] if args.model_mode != 0 else (tokenizer.bos_token + prompt)# 通过分词器将文本转换成对应的token_ID
x = torch.tensor(tokenizer(new_prompt)['input_ids'], device=args.device).unsqueeze(0)

1.1.3 推理

# 将转换为token_ID的数据送入大模型推理,得到对应输出
outputs = model.generate(x,eos_token_id=tokenizer.eos_token_id,max_new_tokens=args.max_seq_len,temperature=args.temperature,top_p=args.top_p,stream=True,pad_token_id=tokenizer.pad_token_id)

1.1.4 推理后 

# 解码模型的输出,转换为文本数据
tokenizer.decode(outputs.squeeze()[x.shape[1]:].tolist(), skip_special_tokens=True)

1.2  init_model()

def init_model(args):tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')if args.load == 0:moe_path = '_moe' if args.use_moe else ''modes = {0: 'pretrain', 1: 'full_sft', 2: 'rlhf', 3: 'reason'}# ckp = f'./{args.out_dir}/{modes[args.model_mode]}_{args.dim}{moe_path}.pth'ckp = f'../Weights/MiniMind2-PyTorch/{modes[args.model_mode]}_{args.dim}{moe_path}.pth'model = MiniMindLM(LMConfig(dim=args.dim,n_layers=args.n_layers,max_seq_len=args.max_seq_len,use_moe=args.use_moe))state_dict = torch.load(ckp, map_location=args.device)model.load_state_dict({k: v for k, v in state_dict.items() if 'mask' not in k}, strict=True)if args.lora_name != 'None':apply_lora(model)load_lora(model, f'./{args.out_dir}/lora/{args.lora_name}_{args.dim}.pth')else:transformers_model_path = '../Weights/MiniMind2'tokenizer = AutoTokenizer.from_pretrained(transformers_model_path)model = AutoModelForCausalLM.from_pretrained(transformers_model_path, trust_remote_code=True)print(f'MiniMind模型参数量: {sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e6:.2f}M(illion)')return model.eval().to(args.device), tokenizer

1.2.1 加载分词器

tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')

1.2.2 初始化模型

model = MiniMindLM(LMConfig(dim=args.dim,n_layers=args.n_layers,max_seq_len=args.max_seq_len,use_moe=args.use_moe))

1.2.3 读取模型权重并赋值给模型

        这里我们选择通过Pytorch加载,最后返回模型和分词器;

moe_path = '_moe' if args.use_moe else ''
modes = {0: 'pretrain', 1: 'full_sft', 2: 'rlhf', 3: 'reason'}
# ckp = f'./{args.out_dir}/{modes[args.model_mode]}_{args.dim}{moe_path}.pth'
ckp = f'../Weights/MiniMind2-PyTorch/{modes[args.model_mode]}_{args.dim}{moe_path}.pth'
state_dict = torch.load(ckp, map_location=args.device)
model.load_state_dict({k: v for k, v in state_dict.items() if 'mask' not in k}, strict=True)

 1.3 get_prompt_datas()

       作者定义了很多不同类型模型对应的提示词;

def get_prompt_datas(args):if args.model_mode == 0:# pretrain模型的接龙能力(无法对话)prompt_datas = ['马克思主义基本原理','人类大脑的主要功能','万有引力原理是','世界上最高的山峰是','二氧化碳在空气中','地球上最大的动物有','杭州市的美食有']else:if args.lora_name == 'None':# 通用对话问题prompt_datas = ['请介绍一下自己。','你更擅长哪一个学科?','鲁迅的《狂人日记》是如何批判封建礼教的?','我咳嗽已经持续了两周,需要去医院检查吗?','详细的介绍光速的物理概念。','推荐一些杭州的特色美食吧。','请为我讲解“大语言模型”这个概念。','如何理解ChatGPT?','Introduce the history of the United States, please.']else:# 特定领域问题lora_prompt_datas = {'lora_identity': ["你是ChatGPT吧。","你叫什么名字?","你和openai是什么关系?"],'lora_medical': ['我最近经常感到头晕,可能是什么原因?','我咳嗽已经持续了两周,需要去医院检查吗?','服用抗生素时需要注意哪些事项?','体检报告中显示胆固醇偏高,我该怎么办?','孕妇在饮食上需要注意什么?','老年人如何预防骨质疏松?','我最近总是感到焦虑,应该怎么缓解?','如果有人突然晕倒,应该如何急救?'],}prompt_datas = lora_prompt_datas[args.lora_name]return prompt_datas

1.4 整体流程

二、模型初始化及推理详情

       根据‘from model.model import MiniMindLM’我们可以找到定义模型结构的模型类;

# model.pyimport math
import struct
import inspect
import timefrom .LMConfig import LMConfig
from typing import Any, Optional, Tuple, List
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from transformers import PreTrainedModel
from transformers.modeling_outputs import CausalLMOutputWithPastclass RMSNorm(torch.nn.Module):def __init__(self, dim: int, eps: float = 1e-6):super().__init__()self.eps = epsself.weight = nn.Parameter(torch.ones(dim))def _norm(self, x):return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)def forward(self, x):return self.weight * self._norm(x.float()).type_as(x)def precompute_pos_cis(dim: int, end: int = int(32 * 1024), theta: float = 1e6):freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))t = torch.arange(end, device=freqs.device)  # type: ignorefreqs = torch.outer(t, freqs).float()  # type: ignorepos_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64return pos_cisdef apply_rotary_emb(xq, xk, pos_cis):def unite_shape(pos_cis, x):ndim = x.ndimassert 0 <= 1 < ndimassert pos_cis.shape == (x.shape[1], x.shape[-1])shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]return pos_cis.view(*shape)xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))pos_cis = unite_shape(pos_cis, xq_)xq_out = torch.view_as_real(xq_ * pos_cis).flatten(3)xk_out = torch.view_as_real(xk_ * pos_cis).flatten(3)return xq_out.type_as(xq), xk_out.type_as(xk)def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:"""torch.repeat_interleave(x, dim=2, repeats=n_rep)"""bs, slen, n_kv_heads, head_dim = x.shapeif n_rep == 1:return xreturn (x[:, :, :, None, :].expand(bs, slen, n_kv_heads, n_rep, head_dim).reshape(bs, slen, n_kv_heads * n_rep, head_dim))class Attention(nn.Module):def __init__(self, args: LMConfig):super().__init__()self.n_kv_heads = args.n_heads if args.n_kv_heads is None else args.n_kv_headsassert args.n_heads % self.n_kv_heads == 0self.n_local_heads = args.n_headsself.n_local_kv_heads = self.n_kv_headsself.n_rep = self.n_local_heads // self.n_local_kv_headsself.head_dim = args.dim // args.n_headsself.wq = nn.Linear(args.dim, args.n_heads * self.head_dim, bias=False)self.wk = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False)self.wv = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False)self.wo = nn.Linear(args.n_heads * self.head_dim, args.dim, bias=False)self.attn_dropout = nn.Dropout(args.dropout)self.resid_dropout = nn.Dropout(args.dropout)self.dropout = args.dropoutself.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention') and args.flash_attn# print("WARNING: using slow attention. Flash Attention requires PyTorch >= 2.0")mask = torch.full((1, 1, args.max_seq_len, args.max_seq_len), float("-inf"))mask = torch.triu(mask, diagonal=1)self.register_buffer("mask", mask, persistent=False)def forward(self,x: torch.Tensor,pos_cis: torch.Tensor,past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,use_cache=False):bsz, seq_len, _ = x.shapexq, xk, xv = self.wq(x), self.wk(x), self.wv(x)xq = xq.view(bsz, seq_len, self.n_local_heads, self.head_dim)xk = xk.view(bsz, seq_len, self.n_local_kv_heads, self.head_dim)xv = xv.view(bsz, seq_len, self.n_local_kv_heads, self.head_dim)xq, xk = apply_rotary_emb(xq, xk, pos_cis)# kv_cache实现if past_key_value is not None:xk = torch.cat([past_key_value[0], xk], dim=1)xv = torch.cat([past_key_value[1], xv], dim=1)past_kv = (xk, xv) if use_cache else Nonexq, xk, xv = (xq.transpose(1, 2),repeat_kv(xk, self.n_rep).transpose(1, 2),repeat_kv(xv, self.n_rep).transpose(1, 2))if self.flash and seq_len != 1:dropout_p = self.dropout if self.training else 0.0output = F.scaled_dot_product_attention(xq, xk, xv,attn_mask=None,dropout_p=dropout_p,is_causal=True)else:scores = (xq @ xk.transpose(-2, -1)) / math.sqrt(self.head_dim)scores += self.mask[:, :, :seq_len, :seq_len]scores = F.softmax(scores.float(), dim=-1).type_as(xq)scores = self.attn_dropout(scores)output = scores @ xvoutput = output.transpose(1, 2).reshape(bsz, seq_len, -1)output = self.resid_dropout(self.wo(output))return output, past_kvclass FeedForward(nn.Module):def __init__(self, config: LMConfig):super().__init__()if config.hidden_dim is None:hidden_dim = 4 * config.dimhidden_dim = int(2 * hidden_dim / 3)config.hidden_dim = config.multiple_of * ((hidden_dim + config.multiple_of - 1) // config.multiple_of)self.w1 = nn.Linear(config.dim, config.hidden_dim, bias=False)self.w2 = nn.Linear(config.hidden_dim, config.dim, bias=False)self.w3 = nn.Linear(config.dim, config.hidden_dim, bias=False)self.dropout = nn.Dropout(config.dropout)def forward(self, x):return self.dropout(self.w2(F.silu(self.w1(x)) * self.w3(x)))class MoEGate(nn.Module):def __init__(self, config: LMConfig):super().__init__()self.config = configself.top_k = config.num_experts_per_tokself.n_routed_experts = config.n_routed_expertsself.scoring_func = config.scoring_funcself.alpha = config.aux_loss_alphaself.seq_aux = config.seq_auxself.norm_topk_prob = config.norm_topk_probself.gating_dim = config.dimself.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim)))self.reset_parameters()def reset_parameters(self) -> None:import torch.nn.init as initinit.kaiming_uniform_(self.weight, a=math.sqrt(5))def forward(self, hidden_states):bsz, seq_len, h = hidden_states.shapehidden_states = hidden_states.view(-1, h)logits = F.linear(hidden_states, self.weight, None)if self.scoring_func == 'softmax':scores = logits.softmax(dim=-1)else:raise NotImplementedError(f'insupportable scoring function for MoE gating: {self.scoring_func}')topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False)if self.top_k > 1 and self.norm_topk_prob:denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20topk_weight = topk_weight / denominatorif self.training and self.alpha > 0.0:scores_for_aux = scoresaux_topk = self.top_ktopk_idx_for_aux_loss = topk_idx.view(bsz, -1)if self.seq_aux:scores_for_seq_aux = scores_for_aux.view(bsz, seq_len, -1)ce = torch.zeros(bsz, self.n_routed_experts, device=hidden_states.device)ce.scatter_add_(1, topk_idx_for_aux_loss,torch.ones(bsz, seq_len * aux_topk, device=hidden_states.device)).div_(seq_len * aux_topk / self.n_routed_experts)aux_loss = (ce * scores_for_seq_aux.mean(dim=1)).sum(dim=1).mean() * self.alphaelse:mask_ce = F.one_hot(topk_idx_for_aux_loss.view(-1), num_classes=self.n_routed_experts)ce = mask_ce.float().mean(0)Pi = scores_for_aux.mean(0)fi = ce * self.n_routed_expertsaux_loss = (Pi * fi).sum() * self.alphaelse:aux_loss = 0return topk_idx, topk_weight, aux_lossclass MOEFeedForward(nn.Module):def __init__(self, config: LMConfig):super().__init__()self.config = configself.experts = nn.ModuleList([FeedForward(config)for _ in range(config.n_routed_experts)])self.gate = MoEGate(config)if config.n_shared_experts is not None:self.shared_experts = FeedForward(config)def forward(self, x):identity = xorig_shape = x.shapebsz, seq_len, _ = x.shape# 使用门控机制选择专家topk_idx, topk_weight, aux_loss = self.gate(x)x = x.view(-1, x.shape[-1])flat_topk_idx = topk_idx.view(-1)if self.training:# 训练模式下,重复输入数据x = x.repeat_interleave(self.config.num_experts_per_tok, dim=0)y = torch.empty_like(x, dtype=torch.float16)for i, expert in enumerate(self.experts):y[flat_topk_idx == i] = expert(x[flat_topk_idx == i]).to(y.dtype)  # 确保类型一致y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)y = y.view(*orig_shape)else:# 推理模式下,只选择最优专家y = self.moe_infer(x, flat_topk_idx, topk_weight.view(-1, 1)).view(*orig_shape)if self.config.n_shared_experts is not None:y = y + self.shared_experts(identity)self.aux_loss = aux_lossreturn y@torch.no_grad()def moe_infer(self, x, flat_expert_indices, flat_expert_weights):expert_cache = torch.zeros_like(x)idxs = flat_expert_indices.argsort()tokens_per_expert = flat_expert_indices.bincount().cpu().numpy().cumsum(0)token_idxs = idxs // self.config.num_experts_per_tok# 例如当tokens_per_expert=[6, 15, 20, 26, 33, 38, 46, 52]# 当token_idxs=[3, 7, 19, 21, 24, 25,  4,  5,  6, 10, 11, 12...]# 意味着当token_idxs[:6] -> [3,  7, 19, 21, 24, 25,  4]位置的token都由专家0处理,token_idxs[6:15]位置的token都由专家1处理......for i, end_idx in enumerate(tokens_per_expert):start_idx = 0 if i == 0 else tokens_per_expert[i - 1]if start_idx == end_idx:continueexpert = self.experts[i]exp_token_idx = token_idxs[start_idx:end_idx]expert_tokens = x[exp_token_idx]expert_out = expert(expert_tokens).to(expert_cache.dtype)expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]])# 使用 scatter_add_ 进行 sum 操作expert_cache.scatter_add_(0, exp_token_idx.view(-1, 1).repeat(1, x.shape[-1]), expert_out)return expert_cacheclass MiniMindBlock(nn.Module):def __init__(self, layer_id: int, config: LMConfig):super().__init__()self.n_heads = config.n_headsself.dim = config.dimself.head_dim = config.dim // config.n_headsself.attention = Attention(config)self.layer_id = layer_idself.attention_norm = RMSNorm(config.dim, eps=config.norm_eps)self.ffn_norm = RMSNorm(config.dim, eps=config.norm_eps)self.feed_forward = FeedForward(config) if not config.use_moe else MOEFeedForward(config)def forward(self, x, pos_cis, past_key_value=None, use_cache=False):h_attn, past_kv = self.attention(self.attention_norm(x),pos_cis,past_key_value=past_key_value,use_cache=use_cache)h = x + h_attnout = h + self.feed_forward(self.ffn_norm(h))return out, past_kvclass MiniMindLM(PreTrainedModel):config_class = LMConfigdef __init__(self, params: LMConfig = None):self.params = params or LMConfig()super().__init__(self.params)self.vocab_size, self.n_layers = params.vocab_size, params.n_layersself.tok_embeddings = nn.Embedding(params.vocab_size, params.dim)self.dropout = nn.Dropout(params.dropout)self.layers = nn.ModuleList([MiniMindBlock(l, params) for l in range(self.n_layers)])self.norm = RMSNorm(params.dim, eps=params.norm_eps)self.output = nn.Linear(params.dim, params.vocab_size, bias=False)self.tok_embeddings.weight = self.output.weightself.register_buffer("pos_cis",precompute_pos_cis(dim=params.dim // params.n_heads, theta=params.rope_theta),persistent=False)self.OUT = CausalLMOutputWithPast()def forward(self,input_ids: Optional[torch.Tensor] = None,past_key_values: Optional[List[Tuple[torch.Tensor, torch.Tensor]]] = None,use_cache: bool = False,**args):past_key_values = past_key_values or [None] * len(self.layers)start_pos = args.get('start_pos', 0)h = self.dropout(self.tok_embeddings(input_ids))# print(h.shape)pos_cis = self.pos_cis[start_pos:start_pos + input_ids.size(1)]past_kvs = []for l, layer in enumerate(self.layers):h, past_kv = layer(h, pos_cis,past_key_value=past_key_values[l],use_cache=use_cache)past_kvs.append(past_kv)logits = self.output(self.norm(h))aux_loss = sum(l.feed_forward.aux_loss for l in self.layers if isinstance(l.feed_forward, MOEFeedForward))self.OUT.__setitem__('logits', logits)self.OUT.__setitem__('aux_loss', aux_loss)self.OUT.__setitem__('past_key_values', past_kvs)return self.OUT@torch.inference_mode()def generate(self, input_ids, eos_token_id=2, max_new_tokens=1024, temperature=0.75, top_p=0.90,stream=False, rp=1., use_cache=True, pad_token_id=0, **args):# 流式生成if stream:return self._stream(input_ids, eos_token_id, max_new_tokens, temperature, top_p, rp, use_cache, **args)# 直接生成generated = []for i in range(input_ids.size(0)):non_pad = input_ids[i][input_ids[i] != pad_token_id].unsqueeze(0)out = self._stream(non_pad, eos_token_id, max_new_tokens, temperature, top_p, rp, use_cache, **args)tokens_list = [tokens[:, -1:] for tokens in out]gen = torch.cat(tokens_list, dim=-1) if tokens_list else non_padfull_sequence = torch.cat([non_pad, gen], dim=-1)generated.append(full_sequence)max_length = max(seq.size(1) for seq in generated)generated = [torch.cat([seq, torch.full((1, max_length - seq.size(1)), pad_token_id, dtype=seq.dtype, device=seq.device)],dim=-1)for seq in generated]return torch.cat(generated, dim=0)def _stream(self, input_ids, eos_token_id, max_new_tokens, temperature, top_p, rp, use_cache, **args):start, first_seq, past_kvs = input_ids.shape[1], True, Nonewhile input_ids.shape[1] < max_new_tokens - 1:if first_seq or not use_cache:out, first_seq = self(input_ids, past_key_values=past_kvs, use_cache=use_cache, **args), Falseelse:out = self(input_ids[:, -1:], past_key_values=past_kvs, use_cache=use_cache,start_pos=input_ids.shape[1] - 1, **args)logits, past_kvs = out.logits[:, -1, :], out.past_key_valueslogits[:, list(set(input_ids.tolist()[0]))] /= rplogits /= (temperature + 1e-9)if top_p is not None and top_p < 1.0:sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)sorted_probs = F.softmax(sorted_logits, dim=-1)cumulative_probs = torch.cumsum(sorted_probs, dim=-1)sorted_indices_to_remove = cumulative_probs > top_psorted_indices_to_remove[:, 1:] = sorted_indices_to_remove[:, :-1].clone()sorted_indices_to_remove[:, 0] = Falseindices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)logits[indices_to_remove] = -float('Inf')input_ids_next = torch.multinomial(F.softmax(logits, dim=-1), num_samples=1)input_ids = torch.cat((input_ids, input_ids_next), dim=1)yield input_ids[:, start:]if input_ids_next.item() == eos_token_id:break

2.1 模型初始化 

class MiniMindLM(PreTrainedModel):config_class = LMConfigdef __init__(self, params: LMConfig = None):self.params = params or LMConfig()super().__init__(self.params)self.vocab_size, self.n_layers = params.vocab_size, params.n_layersself.tok_embeddings = nn.Embedding(params.vocab_size, params.dim)self.dropout = nn.Dropout(params.dropout)self.layers = nn.ModuleList([MiniMindBlock(l, params) for l in range(self.n_layers)])self.norm = RMSNorm(params.dim, eps=params.norm_eps)self.output = nn.Linear(params.dim, params.vocab_size, bias=False)self.tok_embeddings.weight = self.output.weightself.register_buffer("pos_cis",precompute_pos_cis(dim=params.dim // params.n_heads, theta=params.rope_theta),persistent=False)self.OUT = CausalLMOutputWithPast()

2.1.1 超参 

# LMConfig.pyfrom transformers import PretrainedConfig
from typing import Listclass LMConfig(PretrainedConfig):model_type = "minimind"def __init__(self,dim: int = 512,     # 模型宽度n_layers: int = 8,    # 模型深度n_heads: int = 8,    # 自注意力Q的头数n_kv_heads: int = 2,    # K、V的头数vocab_size: int = 6400,   # 词汇表数量hidden_dim: int = None,   multiple_of: int = 64,norm_eps: float = 1e-5,    #这是一个非常小的数值,用于确保数值稳定性,避免除以接近零的数导致的数值问题max_seq_len: int = 8192,rope_theta: int = 1e6,dropout: float = 0.0,flash_attn: bool = True,##################################################### Here are the specific configurations of MOE# When use_moe is false, the following is invalid####################################################use_moe: bool = False,####################################################num_experts_per_tok: int = 2,n_routed_experts: int = 4,n_shared_experts: bool = True,scoring_func: str = 'softmax',aux_loss_alpha: float = 0.1,seq_aux: bool = True,norm_topk_prob: bool = True,**kwargs,):self.dim = dimself.n_layers = n_layersself.n_heads = n_headsself.n_kv_heads = n_kv_headsself.vocab_size = vocab_sizeself.hidden_dim = hidden_dimself.multiple_of = multiple_ofself.norm_eps = norm_epsself.max_seq_len = max_seq_lenself.rope_theta = rope_thetaself.dropout = dropoutself.flash_attn = flash_attn##################################################### Here are the specific configurations of MOE# When use_moe is false, the following is invalid####################################################self.use_moe = use_moeself.num_experts_per_tok = num_experts_per_tok  # 每个token选择的专家数量self.n_routed_experts = n_routed_experts  # 总的专家数量self.n_shared_experts = n_shared_experts  # 共享专家self.scoring_func = scoring_func  # 评分函数,默认为'softmax'self.aux_loss_alpha = aux_loss_alpha  # 辅助损失的alpha参数self.seq_aux = seq_aux  # 是否在序列级别上计算辅助损失self.norm_topk_prob = norm_topk_prob  # 是否标准化top-k概率super().__init__(**kwargs)

2.1.2 各层介绍 

2.1.2.1 embedding词嵌入层

       用于将tokenID转换为指定的dim维度的张量,该层的可训练参数维度(6400,512),在这里会将其转换为(1,512);

self.tok_embeddings = nn.Embedding(params.vocab_size, params.dim)
2.1.2.2  Dropout层

       torch的一个标准层,在训练过程中会随机的不更新一定比例的权重的梯度,防止过拟合;

self.dropout = nn.Dropout(params.dropout)
2.1.2.3  MiniMindBlock核心层

       定义transformer模块的具体结构,像这样的模块该模型一共叠加了8个;

class MiniMindBlock(nn.Module):def __init__(self, layer_id: int, config: LMConfig):super().__init__()self.n_heads = config.n_headsself.dim = config.dimself.head_dim = config.dim // config.n_headsself.attention = Attention(config)self.layer_id = layer_idself.attention_norm = RMSNorm(config.dim, eps=config.norm_eps)self.ffn_norm = RMSNorm(config.dim, eps=config.norm_eps)self.feed_forward = FeedForward(config) if not config.use_moe else MOEFeedForward(config)def forward(self, x, pos_cis, past_key_value=None, use_cache=False):h_attn, past_kv = self.attention(self.attention_norm(x),pos_cis,past_key_value=past_key_value,use_cache=use_cache)h = x + h_attnout = h + self.feed_forward(self.ffn_norm(h))return out, past_kv
2.1.2.4 归一化层

       RMSNorm 是一种替代传统 LayerNorm(层归一化)的技术,旨在减少计算开销和内存占用,同时保持或改善模型的性能;

self.norm = RMSNorm(params.dim, eps=params.norm_eps)
2.1.2.5 输出层 

       输出层将512维的张量通过线性变换转换为6400维与词汇表数量对齐;

self.output = nn.Linear(params.dim, params.vocab_size, bias=False)

2.2 推理 

       大家安装好环境,在eval_model.py同级目录创建一个jujupyterNotebook文件,这里带领大家一起推理一次,感受一下推理全过程,以及中间的数据处理细节;

2.2.1 导包

代码:

import argparse
import random
import time
import numpy as np
import torch
import warnings
from transformers import AutoTokenizer, AutoModelForCausalLM
from model.model import MiniMindLM
from model.LMConfig import LMConfig
from model.model_lora import *warnings.filterwarnings('ignore')

2.2.2 定义超参

代码 :

class ARG():def __init__(self):self.lora_name = 'None'self.out_dir = 'out'self.temperature = 0.85self.top_p = 0.85self.device = 'cpu'self.dim = 512self.n_layers = 8self.max_seq_len = 8192self.use_moe = Falseself.history_cnt = 0self.stream = Trueself.load = 0self.model_mode = 1

2.2.3 定义模型、分词器初始化API

代码:

def init_model(args):tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')if args.load == 0:moe_path = '_moe' if args.use_moe else ''modes = {0: 'pretrain', 1: 'full_sft', 2: 'rlhf', 3: 'reason'}# ckp = f'./{args.out_dir}/{modes[args.model_mode]}_{args.dim}{moe_path}.pth'ckp = f'../Weights/MiniMind2-PyTorch/{modes[args.model_mode]}_{args.dim}{moe_path}.pth'model = MiniMindLM(LMConfig(dim=args.dim,n_layers=args.n_layers,max_seq_len=args.max_seq_len,use_moe=args.use_moe))state_dict = torch.load(ckp, map_location=args.device)model.load_state_dict({k: v for k, v in state_dict.items() if 'mask' not in k}, strict=True)if args.lora_name != 'None':apply_lora(model)load_lora(model, f'./{args.out_dir}/lora/{args.lora_name}_{args.dim}.pth')else:transformers_model_path = '../Weights/MiniMind2'tokenizer = AutoTokenizer.from_pretrained(transformers_model_path)model = AutoModelForCausalLM.from_pretrained(transformers_model_path, trust_remote_code=True)print(f'MiniMind模型参数量: {sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e6:.2f}M(illion)')return model.eval().to(args.device), tokenizer

2.2.4 创建提示词API

代码:

def get_prompt_datas(args):if args.model_mode == 0:# pretrain模型的接龙能力(无法对话)prompt_datas = ['马克思主义基本原理','人类大脑的主要功能','万有引力原理是','世界上最高的山峰是','二氧化碳在空气中','地球上最大的动物有','杭州市的美食有']else:if args.lora_name == 'None':# 通用对话问题prompt_datas = ['请介绍一下自己。','你更擅长哪一个学科?','鲁迅的《狂人日记》是如何批判封建礼教的?','我咳嗽已经持续了两周,需要去医院检查吗?','详细的介绍光速的物理概念。','推荐一些杭州的特色美食吧。','请为我讲解“大语言模型”这个概念。','如何理解ChatGPT?','Introduce the history of the United States, please.']else:# 特定领域问题lora_prompt_datas = {'lora_identity': ["你是ChatGPT吧。","你叫什么名字?","你和openai是什么关系?"],'lora_medical': ['我最近经常感到头晕,可能是什么原因?','我咳嗽已经持续了两周,需要去医院检查吗?','服用抗生素时需要注意哪些事项?','体检报告中显示胆固醇偏高,我该怎么办?','孕妇在饮食上需要注意什么?','老年人如何预防骨质疏松?','我最近总是感到焦虑,应该怎么缓解?','如果有人突然晕倒,应该如何急救?'],}prompt_datas = lora_prompt_datas[args.lora_name]return prompt_datas

2.2.5  加载超参、加载模型、加载分词器、选择输入提示词

代码:

args = ARG()
model, tokenizer = init_model(args)
prompts = get_prompt_datas(args)
prompt = prompts[3]
print(prompt)

输出结果:

MiniMind模型参数量: 25.83M(illion)
我咳嗽已经持续了两周,需要去医院检查吗?

2.2.6  格式化信息

代码:

messages = []
messages.append({"role": "user", "content": prompt})
print(messages)

输出结果:

[{'role': 'user', 'content': '我咳嗽已经持续了两周,需要去医院检查吗?'}]

2.2.7 套用对话模版

代码:

template = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
print(template)

输出结果:

<s>system
你是 MiniMind,是一个有用的人工智能助手。</s>
<s>user
我咳嗽已经持续了两周,需要去医院检查吗?</s>
<s>assistant

2.2.8 截取最大文本长度

       这里因为没有超出最大输入长度,所以没有任何变化;

代码:

new_prompt = template[-args.max_seq_len + 1:] if args.model_mode != 0 else (tokenizer.bos_token + prompt)
print(new_prompt)

输出结果:

<s>system
你是 MiniMind,是一个有用的人工智能助手。</s>
<s>user
我咳嗽已经持续了两周,需要去医院检查吗?</s>
<s>assistant

2.2.9  文本2tokenID

       我们可以看到分词器将上述的提示词分成了52个token,并且生成对应的tokenID;

代码:

embedding_tensor = tokenizer(new_prompt)['input_ids']
print(len(embedding_tensor))
print(embedding_tensor)

输出结果:

52
[1, 85, 736, 201, 608, 345, 562, 261, 75, 47, 807, 270, 1589, 400, 411, 1946, 740, 1728, 945, 1184, 286, 2, 201, 1, 320, 275, 201, 397, 312, 114, 6339, 124, 2434, 3011, 446, 1346, 2055, 270, 590, 1473, 2037, 4238, 3351, 2235, 814, 2, 201, 1, 1078, 538, 501, 201]

2.2.10 剔除填充token并升维

代码:

x = torch.tensor(embedding_tensor, device=args.device).unsqueeze(0)
print(x.shape)
non_pad = x[0][x[0] != tokenizer.pad_token_id]
print(non_pad.shape)
non_pad = non_pad.unsqueeze(0)
print(non_pad.shape)

输出结果:

torch.Size([1, 52])
torch.Size([52])
torch.Size([1, 52])

2.2.11  推理

       这里的out包括三部分的内容,分别是logits、aux_loss、past_key_values;logits是模型的输出张量,past_key_values是缓存的K、V张量,提高推理速度,因为有8个transformer模块,所以past_key_values长度为8;

代码:

start, first_seq, past_kvs = non_pad.shape[1], True, None
print(start, first_seq, past_kvs)
out, first_seq = model(non_pad, past_key_values=past_kvs, use_cache=True, **{}), False
print(out.logits.shape)
print(len(out.past_key_values))

输出结果:

52 True None
torch.Size([1, 52, 6400])
8

2.2.12 获取最后一个token的张量及K、V缓存

       因为后面进行softmax使用的是最后一个token的张量信息,所以获取输出最后一个token的张量;

代码:

logits, past_kvs = out.logits[:, -1, :], out.past_key_values
print(logits.shape)
print(len(past_kvs))

输出结果:

torch.Size([1, 6400])
8

2.2.13 提高输出的随机性

       这里的0.85就是超参中的temperature;

代码:

logits[:, list(set(non_pad.tolist()[0]))] /= 1.0
logits /= (0.85 + 1e-9)

2.2.14  下面进入Top_p随机采样

       这里简单说一下top_p随机采样的流程和原理;首先进行top_p采样的是最后一个token的概率分布张量,维度为(1,6400);然后将其最后一维按照从大到小的概率排序,计算累加概率,比如排序后的概率为[0.5, 0.2,  0.1, 0.1, 0.1]那么他的累积概率就是[0.5, 0.7, 0.8, 0.9, 1];接着找到累加概率大于top_p的位置,保留所有小于top_p的累加概率的索引对应的原始概率值,其余的全部变为负无穷大,再一次计算softmax;然后按照这个概率加权随机选择一个分类对应的tokenID作为下一个预测token;这样做的目的是让输出有一定的随机性,增加了模型的创新力;

代码1:

# 将logits按最后一个维度从大到小排序,返回排序后的张量以及对应张量在原始位置上的索引张量
sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
print(sorted_logits.shape)
print(sorted_indices)

输出结果1:

torch.Size([1, 6400])
tensor([[ 453,  368,  345,  ..., 4166,  252, 5999]])

 代码2:

# 将排序后的张量在最后一个维度上进行softmax计算,得到和logits维度相同(1,6400)的概率值
import torch.nn.functional as F
sorted_probs = F.softmax(sorted_logits, dim=-1)
print(sorted_probs.shape)
print(sorted_probs)

输出结果2:

torch.Size([1, 6400])
tensor([[1.2660e-01, 1.2199e-01, 1.1748e-01,  ..., 3.4979e-13, 3.1038e-13,2.8181e-13]], grad_fn=<SoftmaxBackward0>)

代码3:

# 计算概率张量在最后一个维度上的累加值,返回累加张量
cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
print(cumulative_probs)

输出结果3:

tensor([[0.1266, 0.2486, 0.3661,  ..., 1.0000, 1.0000, 1.0000]],grad_fn=<CumsumBackward0>)

代码4:

# 定义累加张量上概率值大于top_p的mask遮罩
sorted_indices_to_remove = cumulative_probs > 0.85
print(sorted_indices_to_remove.shape)

输出结果4:

torch.Size([1, 6400])

代码5:

# 将上一位的遮罩状态赋值给下一位,因为累积概率是从第一个 token 开始累加的,直接移除会导致可能没有 token 被保留
sorted_indices_to_remove[:, 1:] = sorted_indices_to_remove[:, :-1].clone()
# 将第一位token概率赋值为False,确保至少保留一个token
sorted_indices_to_remove[:, 0] = False
print(sorted_indices_to_remove.shape)

输出结果5:

torch.Size([1, 6400])

代码6:

# 使用 scatter 函数将 sorted_indices_to_remove 映射回原始 logits 的顺序;
# scatter 的作用是根据 sorted_indices 将布尔标记重新分配到原始 logits 的位置
# indices_to_remove 是一个布尔张量,形状与 logits 相同,表示哪些位置的 logits 应该被移除
indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
print(indices_to_remove.shape)

输出结果6:

torch.Size([1, 6400])

代码7:

# 将移除位置的张量赋值为负无穷
logits[indices_to_remove] = -float('Inf')
print(logits.shape)

代码8:

# 对剔除很小概率值的张量再次进行softmax
logits_softmax = F.softmax(logits, dim=-1)
print(logits_softmax.shape)

代码9:

# 通过multinomial函数按照概率进行加权随机采样,num_samples=1表示在对应维度上只采样1个,返回被选中的张量的索引
input_ids_next = torch.multinomial(logits_softmax, num_samples=1)
print(input_ids_next.shape)
print(input_ids_next)

输出结果9:

torch.Size([1, 1])
tensor([[835]])

2.2.15 合并预测tokenID

       合并后的input_ids会作为下一轮的上下文信息进行新一轮的预测,当预测tokenID为停止tokenID时就会停止推理;

代码:

# 将挑选出来的token的索引添加到之前的上下文索引中
print(non_pad.shape)
input_ids = torch.cat((non_pad, input_ids_next), dim=1)
print(input_ids.shape)

 输出结果:

torch.Size([1, 52])
torch.Size([1, 53])

2.2.16  预测tokenID转文本

       这样就完成一轮的推理预测啦;

代码1:

# 查看预测出来的token_ID是835
new_token = input_ids[:, start:]
print(new_token)

输出结果1:

tensor([[835]])

代码2:

# 将这个token_ID放入到分词器中进行文本解码
text = tokenizer.decode([835], skip_special_tokens=True)
print(text)

输出结果2:

三、总结 

       该篇博客介绍了一下大模型的模型架构,以及模型推理的整个流程,包括文本数据的预处理和模型输出结果的后处理等;比较详细的呈现了大模型推理的完整流程,希望对大家理解大模型有一定的帮助;该系列还会以minimind项目为基础,从代码出发,持续更新分享大模型的预训练、微调、蒸馏、量化等一系列的底层过程,如果感觉有帮助到你,可以点一个关注,持续关注该系列的后续,谢谢~

版权声明:

本网仅为发布的内容提供存储空间,不对发表、转载的内容提供任何形式的保证。凡本网注明“来源:XXX网络”的作品,均转载自其它媒体,著作权归作者所有,商业转载请联系作者获得授权,非商业转载请注明出处。

我们尊重并感谢每一位作者,均已注明文章来源和作者。如因作品内容、版权或其它问题,请及时与我们联系,联系邮箱:809451989@qq.com,投稿邮箱:809451989@qq.com