limour
/
limour-blog.github.io
miroir de https://dl.limour.top/cf/https://github.com/limour-blog/limour-blog.github.io.git


			
							123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750
							

<!DOCTYPE html>
<html lang="en" data-default-color-scheme=auto>


<head><!-- hexo injector head_begin start -->
<script defer src="https://api.limour.top/vue/0d2f95c1-755d-436b-adf8-eee12a80ed32/script.js"></script>
<!-- hexo injector head_begin end -->
  <meta charset="UTF-8">
  <link rel="apple-touch-icon" sizes="76x76" href="https://img.limour.top/2023/08/29/64ee07361815a.webp">
  <link rel="icon" href="https://img.limour.top/2023/08/29/64ee07361815a.webp">
  <meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=5.0, shrink-to-fit=no">
  <meta http-equiv="x-ua-compatible" content="ie=edge">
  
  <meta name="theme-color" content="#2f4154">
  <meta name="author" content="Limour">
  <meta name="keywords" content="">
  
    <meta name="description" content="探索整个过程，从在一台搭载1660Ti显卡的笔记本电脑上构建 Tokenizer，定义带有 RoPE 的 Transformer，一直到训练、保存模型和可视化训练过程。沉浸在从零开始训练 GPT 的旅程中，深入了解每一个步骤。跳入深度学习的世界，释放在你的便携1660Ti笔记本上的强大潜能。">
  
  
  <title>【探索】从零开始训练 GPT - Limour&#39;s Blog</title>

  <link  rel="stylesheet" href="https://jscdn.limour.top/npm/bootstrap@4.6.1/dist/css/bootstrap.min.css" />


  <link  rel="stylesheet" href="https://jscdn.limour.top/npm/github-markdown-css@4.0.0/github-markdown.min.css" />

  <link  rel="stylesheet" href="https://jscdn.limour.top/npm/hint.css@2.7.0/hint.min.css" />


<!-- 主题依赖的图标库，不要自行修改 -->
<!-- Do not modify the link that theme dependent icons -->

<link rel="stylesheet" href="//at.alicdn.com/t/font_1749284_hj8rtnfg7um.css">


<link rel="stylesheet" href="//at.alicdn.com/t/font_1736178_lbnruvf0jn.css">


<link  rel="stylesheet" href="/css/main.css" />


  <link id="highlight-css" rel="stylesheet" href="/css/highlight.css" />
  
    <link id="highlight-css-dark" rel="stylesheet" href="/css/highlight-dark.css" />
  

<link rel="stylesheet" href="/theme-inject/custom.css">
<link rel="stylesheet" href="/theme-inject/iconfont.css">


  <script id="fluid-configs">
    var Fluid = window.Fluid || {};
    Fluid.ctx = Object.assign({}, Fluid.ctx)
    var CONFIG = {"hostname":"hexo.limour.top","root":"/","version":"1.9.7","typing":{"enable":false,"typeSpeed":70,"cursorChar":"_","loop":false,"scope":[]},"anchorjs":{"enable":true,"element":"h1,h2,h3,h4,h5,h6","placement":"left","visible":"hover","icon":"§"},"progressbar":{"enable":true,"height_px":3,"color":"#29d","options":{"showSpinner":false,"trickleSpeed":100}},"code_language":{"enable":true,"default":"TEXT"},"copy_btn":true,"image_caption":{"enable":true},"image_zoom":{"enable":false,"img_url_replace":["",""]},"toc":{"enable":true,"placement":"right","headingSelector":"h1,h2,h3,h4,h5,h6","collapseDepth":0},"lazyload":{"enable":true,"loading_img":"https://jscdn.limour.top/gh/Limour-dev/Sakurairo_Vision/load_svg/inload.svg","onlypost":false,"offset_factor":2},"web_analytics":{"enable":false,"follow_dnt":true,"baidu":null,"google":{"measurement_id":null},"tencent":{"sid":null,"cid":null},"woyaola":null,"cnzz":null,"leancloud":{"app_id":null,"app_key":null,"server_url":null,"path":"window.location.pathname","ignore_local":false}},"search_path":"/local-search.xml","include_content_in_search":true};

    if (CONFIG.web_analytics.follow_dnt) {
      var dntVal = navigator.doNotTrack || window.doNotTrack || navigator.msDoNotTrack;
      Fluid.ctx.dnt = dntVal && (dntVal.startsWith('1') || dntVal.startsWith('yes') || dntVal.startsWith('on'));
    }
  </script>
  <script  src="/js/utils.js" ></script>
  <script  src="/js/color-schema.js" ></script>
  

  <link rel="canonical" href="https://hexo.limour.top/training-gpt-from-scratch"/>
<meta name="generator" content="Hexo 7.1.1"><link rel="alternate" href="/atom.xml" title="Limour's Blog" type="application/atom+xml">
<link rel="alternate" href="/rss2.xml" title="Limour's Blog" type="application/rss+xml">
</head>


<body>
  

  <header>
    

<div class="header-inner" style="height: 70vh;">
  <nav id="navbar" class="navbar fixed-top  navbar-expand-lg navbar-dark scrolling-navbar">
  <div class="container">
    <a class="navbar-brand" href="/">
      <strong>Limour&#39;s Blog</strong>
    </a>

    <button id="navbar-toggler-btn" class="navbar-toggler" type="button" data-toggle="collapse"
            data-target="#navbarSupportedContent"
            aria-controls="navbarSupportedContent" aria-expanded="false" aria-label="Toggle navigation">
      <div class="animated-icon"><span></span><span></span><span></span></div>
    </button>

    <!-- Collapsible content -->
    <div class="collapse navbar-collapse" id="navbarSupportedContent">
      <ul class="navbar-nav ml-auto text-center">
        
          
            <li class="nav-item">
              <a class="nav-link" href="https://hexo.limour.top/" target="_self">
                <i class="iconfont icon-home-fill"></i>
                <span>Home</span>
              </a>
            </li>
          
        
            <li class="nav-item">
              <a class="nav-link" href="/archives/" target="_self">
                <i class="iconfont icon-archive-fill"></i>
                <span>Archive1</span>
              </a>
            </li>
          
        
            <li class="nav-item">
              <a class="nav-link" href="https://occdn.limour.top/archives/" target="_self">
                <i class="iconfont icon-archive-fill"></i>
                <span>Archive2</span>
              </a>
            </li>
          
        
            <li class="nav-item">
              <a class="nav-link" href="https://b.limour.top/archives/" target="_self">
                <i class="iconfont icon-archive-fill"></i>
                <span>Archive3</span>
              </a>
            </li>
          
        
            <li class="nav-item">
              <a class="nav-link" href="https://od.limour.top/" target="_self">
                <i class="iconfont icon-onedrive"></i>
                <span>Alist</span>
              </a>
            </li>
          
        
            <li class="nav-item">
              <a class="nav-link" href="https://orcid.org/0000-0001-8897-1685" target="_self">
                <i class="iconfont icon-orcid"></i>
                <span>Orcid</span>
              </a>
            </li>
          
        
            <li class="nav-item">
              <a class="nav-link" href="/links/" target="_self">
                <i class="iconfont icon-link-fill"></i>
                <span>Links</span>
              </a>
            </li>
          
        
            <li class="nav-item">
              <a class="nav-link" href="/atom.xml" target="_self">
                <i class="iconfont icon-rss"></i>
                <span>RSS</span>
              </a>
            </li>
          
        
          <li class="nav-item" id="search-btn">
            <a class="nav-link" target="_self" href="javascript:;" data-toggle="modal" data-target="#modalSearch" aria-label="Search">
              <i class="iconfont icon-search"></i>
            </a>
          </li>
          
        
          <li class="nav-item" id="color-toggle-btn">
            <a class="nav-link" target="_self" href="javascript:;" aria-label="Color Toggle">
              <i class="iconfont icon-dark" id="color-toggle-icon"></i>
            </a>
          </li>
        
      </ul>
    </div>
  </div>
</nav>

  
<div id="banner" class="banner" parallax=true
     style="background: url('https://img.limour.top/2023/08/29/64ee08e108638.webp') no-repeat center center; background-size: cover;">
  <div class="full-bg-img">
    <div class="mask flex-center" style="background-color: rgba(0, 0, 0, 0.3)">
      <div class="banner-text text-center fade-in-up">
        <div class="h2">
          
            <span id="subtitle">【探索】从零开始训练 GPT</span>
          
        </div>

        
  <div class="mt-3">
    
      <span class="post-meta mr-2">
        <i class="iconfont icon-author" aria-hidden="true"></i>
        Limour
      </span>
    
    
      <span class="post-meta">
        <i class="iconfont icon-date-fill" aria-hidden="true"></i>
        <time datetime="2024-01-18 22:19" pubdate>
          January 18, 2024 pm
        </time>
      </span>
    
  </div>

  <div class="mt-1">
    
      <span class="post-meta mr-2">
        <i class="iconfont icon-chart"></i>
        
          1.7k words
        
      </span>
    

      <span class="post-meta mr-2">
        <i class="iconfont icon-clock-fill"></i>
        
        
          15 mins
        
      </span>
    

  </div>


      </div>

      
    </div>
  </div>
</div>

</div>

  </header>

  <main>
    
      
<div class="container-fluid nopadding-x">
  <div class="row nomargin-x">
    <div class="side-col d-none d-lg-block col-lg-2">
      

    </div>

    <div class="col-lg-8 nopadding-x-md">
      <div class="container nopadding-x-md" id="board-ctn">
        <div id="board">
          <article class="post-content mx-auto">
            <h1 id="seo-header">【探索】从零开始训练 GPT</h1>
            
              <p id="updated-time" class="note note-info" style="">
                
                  
                    Last updated on March 19, 2024 pm
                  
                
              </p>
            
            
              <div class="markdown-body">
                
                <p><img src="https://img.limour.top/2024/01/18/65a93c6a8065a.webp" srcset="https://jscdn.limour.top/gh/Limour-dev/Sakurairo_Vision/load_svg/inload.svg" lazyload alt="训练中..."></p>
<h2 id="预期结构">预期结构</h2>
<ul>
<li><a target="_blank" rel="noopener" href="https://github.com/Limour-dev/HelloGPT">相关代码已经放到 Github</a></li>
</ul>
<figure class="highlight txt"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br></pre></td><td class="code"><pre><span class="line">HelloGPT(</span><br><span class="line">  (tok_embeddings): Embedding(32765, 768)</span><br><span class="line">  (rotary_emb): RotaryEmbedding(head_dim=64, max_seq_len=1024)</span><br><span class="line">  (layers): ModuleList(</span><br><span class="line">    (0-11): 12 x Decoder(</span><br><span class="line">      (ln1): RMSNorm(hidden_size=768, eps=1e-06)</span><br><span class="line">      (attn): Attention(</span><br><span class="line">        (q_proj): Linear(in_features=768, out_features=768, bias=False)</span><br><span class="line">        (k_proj): Linear(in_features=768, out_features=768, bias=False)</span><br><span class="line">        (v_proj): Linear(in_features=768, out_features=768, bias=False)</span><br><span class="line">        (o_proj): Linear(in_features=768, out_features=768, bias=False)</span><br><span class="line">      )</span><br><span class="line">      (ln2): RMSNorm(hidden_size=768, eps=1e-06)</span><br><span class="line">      (mlp): MLP(</span><br><span class="line">        (gate_proj): Linear(in_features=768, out_features=1536, bias=False)</span><br><span class="line">        (up_proj): Linear(in_features=768, out_features=1536, bias=False)</span><br><span class="line">        (down_proj): Linear(in_features=1536, out_features=768, bias=False)</span><br><span class="line">      )</span><br><span class="line">    )</span><br><span class="line">  )</span><br><span class="line">  (norm): RMSNorm(hidden_size=768, eps=1e-06)</span><br><span class="line">  (ln2): Linear(in_features=768, out_features=32765, bias=False)</span><br><span class="line">)</span><br></pre></td></tr></table></figure>
<h2 id="配置环境">配置环境</h2>
<ul>
<li><a href="/-ji-lu--an-zhuang-conda-bing-geng-huan-qing-hua-yuan">安装conda</a></li>
</ul>
<figure class="highlight powershell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br></pre></td><td class="code"><pre><span class="line"><span class="built_in">cd</span> E:\GPT</span><br><span class="line">conda install mamba <span class="literal">-c</span> conda<span class="literal">-forge</span></span><br><span class="line">mamba create <span class="literal">-n</span> HelloGPT pytorch pytorch<span class="literal">-cuda</span>=<span class="number">12.1</span> <span class="literal">-c</span> pytorch <span class="literal">-c</span> nvidia <span class="literal">-c</span> conda<span class="literal">-forge</span></span><br><span class="line">conda activate HelloGPT</span><br><span class="line">conda install numpy transformers tiktoken tensorboard sentencepiece<span class="literal">-python</span> jieba emoji <span class="literal">-c</span> conda<span class="literal">-forge</span></span><br><span class="line">pip install opencc<span class="literal">-python-reimplemented</span> <span class="literal">-i</span> https://pypi.tuna.tsinghua.edu.cn/simple</span><br><span class="line">python test_cuda.py</span><br><span class="line">python test_SPDA.py</span><br><span class="line">D:\vscode\Code.exe</span><br></pre></td></tr></table></figure>
<h2 id="准备数据">准备数据</h2>
<ul>
<li>下载 <a target="_blank" rel="noopener" href="https://huggingface.co/collections/Limour/r18-novels-galgame-6598f16894cadc9cdcb3f3ab">h-corpus-2023</a></li>
</ul>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br><span class="line">33</span><br><span class="line">34</span><br><span class="line">35</span><br><span class="line">36</span><br><span class="line">37</span><br><span class="line">38</span><br><span class="line">39</span><br><span class="line">40</span><br><span class="line">41</span><br><span class="line">42</span><br><span class="line">43</span><br><span class="line">44</span><br><span class="line">45</span><br><span class="line">46</span><br><span class="line">47</span><br><span class="line">48</span><br><span class="line">49</span><br><span class="line">50</span><br><span class="line">51</span><br><span class="line">52</span><br><span class="line">53</span><br><span class="line">54</span><br><span class="line">55</span><br><span class="line">56</span><br><span class="line">57</span><br><span class="line">58</span><br><span class="line">59</span><br><span class="line">60</span><br><span class="line">61</span><br><span class="line">62</span><br><span class="line">63</span><br><span class="line">64</span><br><span class="line">65</span><br><span class="line">66</span><br><span class="line">67</span><br><span class="line">68</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">import</span> os</span><br><span class="line"></span><br><span class="line"><span class="keyword">class</span> <span class="title class_">Fileset</span>(<span class="title class_ inherited__">list</span>):</span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">__init__</span>(<span class="params">self, path, ext=<span class="string">&#x27;&#x27;</span>, _read=<span class="literal">None</span></span>):</span><br><span class="line">        <span class="keyword">if</span> <span class="built_in">isinstance</span>(path, <span class="built_in">str</span>):</span><br><span class="line">            self.root = path</span><br><span class="line">            self.extend(f <span class="keyword">for</span> f <span class="keyword">in</span> os.listdir(self.root) <span class="keyword">if</span> f.endswith(ext))</span><br><span class="line">            self._read = _read</span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">__getitem__</span>(<span class="params">self, index</span>):</span><br><span class="line">        <span class="keyword">if</span> <span class="built_in">isinstance</span>(index, <span class="built_in">int</span>):  <span class="comment"># index是索引</span></span><br><span class="line">            <span class="keyword">if</span> self._read:</span><br><span class="line">                <span class="keyword">return</span> self._read(os.path.join(self.root, <span class="built_in">super</span>().__getitem__(index)))</span><br><span class="line">            <span class="keyword">else</span>:</span><br><span class="line">                <span class="keyword">return</span> os.path.join(self.root, <span class="built_in">super</span>().__getitem__(index))</span><br><span class="line">        <span class="keyword">else</span>:  <span class="comment"># index是切片</span></span><br><span class="line">            fileset = Fileset(<span class="literal">None</span>)</span><br><span class="line">            fileset.root = self.root</span><br><span class="line">            fileset._read = self._read</span><br><span class="line">            fileset.extend(<span class="built_in">super</span>().__getitem__(index))</span><br><span class="line">            <span class="keyword">return</span> fileset</span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">getFileName</span>(<span class="params">self, index</span>):</span><br><span class="line">        fname, ext = os.path.splitext(<span class="built_in">super</span>().__getitem__(index))</span><br><span class="line">        <span class="keyword">return</span> fname</span><br><span class="line"></span><br><span class="line"></span><br><span class="line"><span class="keyword">from</span> tokenizer <span class="keyword">import</span> tokenizer</span><br><span class="line">token_eos = <span class="number">2</span></span><br><span class="line"></span><br><span class="line"></span><br><span class="line"><span class="keyword">def</span> <span class="title function_">readOne</span>(<span class="params">filePath</span>):</span><br><span class="line">    retn = []</span><br><span class="line">    <span class="keyword">with</span> <span class="built_in">open</span>(file=filePath, encoding=<span class="string">&#x27;utf-8&#x27;</span>) <span class="keyword">as</span> f:</span><br><span class="line">        <span class="keyword">for</span> line <span class="keyword">in</span> f:</span><br><span class="line">            retn += tokenizer.encode(line).ids</span><br><span class="line">    retn.append(token_eos)</span><br><span class="line">    <span class="keyword">return</span> retn</span><br><span class="line"></span><br><span class="line"></span><br><span class="line"><span class="keyword">class</span> <span class="title class_">Hcorpus</span>():</span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">__init__</span>(<span class="params">self, path, ext=<span class="string">&#x27;txt&#x27;</span>, fileset_idx=<span class="number">0</span>, fileset_sub_idx=<span class="number">0</span></span>):</span><br><span class="line">        self.fileset = Fileset(path, ext, readOne)</span><br><span class="line">        self.fileset_idx = fileset_idx</span><br><span class="line">        self.fileset_sub_idx = fileset_sub_idx</span><br><span class="line">        <span class="keyword">if</span> self.fileset_sub_idx &lt; <span class="number">0</span>:  <span class="comment"># 再读上一个太复杂了，直接放弃</span></span><br><span class="line">            self.fileset_sub_idx = <span class="number">0</span></span><br><span class="line">        <span class="keyword">if</span> self.fileset_idx &gt;= <span class="built_in">len</span>(self.fileset):</span><br><span class="line">            self.fileset_idx = <span class="number">0</span></span><br><span class="line">        self.cache = self.fileset[self.fileset_idx]</span><br><span class="line">        self.fileset_idx += <span class="number">1</span></span><br><span class="line">        self.cache_idx = self.fileset_sub_idx</span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">__call__</span>(<span class="params">self, size=<span class="number">512</span></span>):</span><br><span class="line">        <span class="keyword">while</span> <span class="built_in">len</span>(self.cache) &lt; self.cache_idx + size:</span><br><span class="line">            <span class="keyword">if</span> self.fileset_idx &gt;= <span class="built_in">len</span>(self.fileset):</span><br><span class="line">                self.fileset_idx = <span class="number">0</span></span><br><span class="line">            self.fileset_sub_idx = self.cache_idx - <span class="built_in">len</span>(self.cache)</span><br><span class="line">            self.cache = self.cache[self.cache_idx:] + self.fileset[self.fileset_idx]</span><br><span class="line">            self.cache_idx = <span class="number">0</span></span><br><span class="line">            self.fileset_idx += <span class="number">1</span></span><br><span class="line">        retn = self.cache[self.cache_idx:self.cache_idx + size]</span><br><span class="line">        self.cache_idx += size</span><br><span class="line">        self.fileset_sub_idx += size</span><br><span class="line">        <span class="keyword">return</span> retn</span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">__repr__</span>(<span class="params">self</span>):</span><br><span class="line">        <span class="keyword">return</span> <span class="string">f&quot;Hcorpus(r&#x27;<span class="subst">&#123;self.fileset.root&#125;</span>&#x27;, fileset_idx=<span class="subst">&#123;self.fileset_idx-<span class="number">1</span>&#125;</span>, fileset_sub_idx=<span class="subst">&#123;self.fileset_sub_idx&#125;</span>)&quot;</span></span><br></pre></td></tr></table></figure>
<h2 id="训练Tokenizer">训练Tokenizer</h2>
<ul>
<li><a target="_blank" rel="noopener" href="https://huggingface.co/docs/tokenizers/quicktour">tokenizer 包的文档</a></li>
<li>繁体转换成简体：<a target="_blank" rel="noopener" href="https://github.com/Limour-dev/HelloGPT/blob/main/train_tokenizer_pre.py">train_tokenizer_pre.py</a></li>
<li>获取常用 emoji：<a target="_blank" rel="noopener" href="https://github.com/Limour-dev/HelloGPT/blob/main/tmp_emoji.py">tmp_emoji.py</a></li>
<li>分词统计词频：<a target="_blank" rel="noopener" href="https://github.com/Limour-dev/HelloGPT/blob/main/train_tokenizer_jieba.py">tokenizer_jieba.py</a></li>
<li>区分词性并构造 BPE 语料：<a target="_blank" rel="noopener" href="https://github.com/Limour-dev/HelloGPT/blob/main/train_tokenizer_jieba_statistics.py">train_tokenizer_jieba_statistics.py</a></li>
<li>训练 BPE 模型：<a target="_blank" rel="noopener" href="https://github.com/Limour-dev/HelloGPT/blob/main/train_tokenizer.py">train_tokenizer.py</a></li>
<li>最终训练好的 BPE 模型：<a target="_blank" rel="noopener" href="https://github.com/Limour-dev/HelloGPT/blob/main/HelloBPE.tokenizer.json">HelloBPE.tokenizer.json</a></li>
</ul>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">from</span> tokenizers <span class="keyword">import</span> Tokenizer</span><br><span class="line">tokenizer = Tokenizer.from_file(<span class="string">&quot;HelloBPE.tokenizer.json&quot;</span>)</span><br></pre></td></tr></table></figure>
<h2 id="定义模型">定义模型</h2>
<h3 id="定义-Decoder">定义 Decoder</h3>
<h4 id="定义-RMSnorm">定义 RMSnorm</h4>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">class</span> <span class="title class_">RMSNorm</span>(nn.Module):</span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">__init__</span>(<span class="params">self, dim: <span class="built_in">int</span>, eps: <span class="built_in">float</span> = <span class="number">1e-6</span></span>):</span><br><span class="line">        <span class="built_in">super</span>().__init__()</span><br><span class="line">        self.eps = eps</span><br><span class="line">        self.weight = nn.Parameter(torch.ones(dim))</span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">forward</span>(<span class="params">self, x</span>):</span><br><span class="line">        x = x * torch.rsqrt(x.<span class="built_in">pow</span>(<span class="number">2</span>).mean(-<span class="number">1</span>, keepdim=<span class="literal">True</span>) + self.eps)</span><br><span class="line">        <span class="keyword">return</span> x * self.weight</span><br></pre></td></tr></table></figure>
<h4 id="定义-RoPE">定义 RoPE</h4>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">class</span> <span class="title class_">RotaryEmbedding</span>(nn.Module):</span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">__init__</span>(<span class="params">self, head_dim: <span class="built_in">int</span>, max_seq_len: <span class="built_in">int</span>, device=device, theta: <span class="built_in">float</span> = <span class="number">10000.0</span></span>):</span><br><span class="line">        <span class="built_in">super</span>().__init__()</span><br><span class="line">        self.head_dim = head_dim</span><br><span class="line">        self.set_max_seq_len(max_seq_len, device, theta)</span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">set_max_seq_len</span>(<span class="params">self, max_seq_len: <span class="built_in">int</span>, device=device, theta: <span class="built_in">float</span> = <span class="number">10000.0</span></span>):</span><br><span class="line">        self.max_seq_len = max_seq_len</span><br><span class="line">        freqs = <span class="number">1.0</span> / (theta ** (torch.arange(<span class="number">0</span>, self.head_dim, <span class="number">2</span>).<span class="built_in">float</span>().to(device) / self.head_dim))</span><br><span class="line">        t = torch.arange(max_seq_len, device=device)  <span class="comment"># type: ignore</span></span><br><span class="line">        freqs = torch.outer(t, freqs).<span class="built_in">float</span>()  <span class="comment"># 外积</span></span><br><span class="line">        self.freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  <span class="comment"># 复数，模 1，角度 freqs</span></span><br><span class="line">        self.freqs_cis.requires_grad = <span class="literal">False</span>  <span class="comment"># filter(lambda p : p.requires_grad, model.parameters())</span></span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">rotary_emb</span>(<span class="params">self, x</span>):</span><br><span class="line">        x_ = torch.view_as_complex(x.<span class="built_in">float</span>().reshape(*x.shape[:-<span class="number">1</span>], -<span class="number">1</span>, <span class="number">2</span>))</span><br><span class="line">        x_out = torch.view_as_real(x_ * self.local_freqs_cis).flatten(<span class="number">3</span>)</span><br><span class="line">        <span class="keyword">return</span> x_out.type_as(x)</span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">forward</span>(<span class="params">self, start_pos: <span class="built_in">int</span>, seqlen: <span class="built_in">int</span></span>):</span><br><span class="line">        self.local_freqs_cis = self.freqs_cis[start_pos: start_pos + seqlen].view(<span class="number">1</span>, seqlen, <span class="number">1</span>, -<span class="number">1</span>)  <span class="comment"># cacheKV 相关，可忽略</span></span><br><span class="line">        self.local_freqs_cis.requires_grad = <span class="literal">False</span></span><br><span class="line">        <span class="keyword">return</span> self.rotary_emb</span><br></pre></td></tr></table></figure>
<h4 id="定义-Attention">定义 Attention</h4>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">class</span> <span class="title class_">Attention</span>(nn.Module):</span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">__init__</span>(<span class="params">self, hidden_size, n_heads, cacheKV, max_batch_size, max_seq_len, device=device</span>):</span><br><span class="line">        <span class="built_in">super</span>().__init__()</span><br><span class="line">        self.n_heads = n_heads</span><br><span class="line">        self.head_dim = hidden_size // n_heads</span><br><span class="line">        self.q_proj = nn.Linear(hidden_size, hidden_size, bias=<span class="literal">False</span>)</span><br><span class="line">        self.k_proj = nn.Linear(hidden_size, hidden_size, bias=<span class="literal">False</span>)</span><br><span class="line">        self.v_proj = nn.Linear(hidden_size, hidden_size, bias=<span class="literal">False</span>)</span><br><span class="line">        self.o_proj = nn.Linear(hidden_size, hidden_size, bias=<span class="literal">False</span>)</span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">forward</span>(<span class="params">self, hidden_states, rotary_emb, start_pos=<span class="number">0</span>, mask=<span class="literal">None</span>, is_causal=<span class="literal">True</span></span>):</span><br><span class="line">        bsz, seqlen, hidden_size = hidden_states.shape</span><br><span class="line"></span><br><span class="line">        q = self.q_proj(hidden_states)</span><br><span class="line">        k = self.k_proj(hidden_states)</span><br><span class="line">        v = self.v_proj(hidden_states)</span><br><span class="line"></span><br><span class="line">        q = q.view(bsz, seqlen, self.n_heads, self.head_dim)</span><br><span class="line">        k = k.view(bsz, seqlen, self.n_heads, self.head_dim)</span><br><span class="line">        v = v.view(bsz, seqlen, self.n_heads, self.head_dim)</span><br><span class="line"></span><br><span class="line">        q = rotary_emb(q)</span><br><span class="line">        k = rotary_emb(k)</span><br><span class="line"></span><br><span class="line">        q = q.transpose(<span class="number">1</span>, <span class="number">2</span>)  <span class="comment"># (bs, n_heads, seqlen, head_dim)</span></span><br><span class="line">        k = k.transpose(<span class="number">1</span>, <span class="number">2</span>)  <span class="comment"># (bs, n_local_heads, cache_len + seqlen, head_dim)</span></span><br><span class="line">        v = v.transpose(<span class="number">1</span>, <span class="number">2</span>)  <span class="comment"># (bs, n_local_heads, cache_len + seqlen, head_dim)</span></span><br><span class="line"></span><br><span class="line">        output = F.scaled_dot_product_attention(q, k, v, attn_mask=mask, is_causal=is_causal)</span><br><span class="line"></span><br><span class="line">        output = output.transpose(<span class="number">1</span>, <span class="number">2</span>).contiguous().view(bsz, seqlen, hidden_size)</span><br><span class="line">        <span class="keyword">return</span> self.o_proj(output)</span><br></pre></td></tr></table></figure>
<h4 id="定义-MLP">定义 MLP</h4>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">class</span> <span class="title class_">MLP</span>(nn.Module):</span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">__init__</span>(<span class="params">self, hidden_size</span>):</span><br><span class="line">        <span class="built_in">super</span>().__init__()</span><br><span class="line">        intermediate_size = <span class="built_in">int</span>(<span class="number">2</span> * hidden_size)</span><br><span class="line">        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=<span class="literal">False</span>)</span><br><span class="line">        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=<span class="literal">False</span>)</span><br><span class="line">        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=<span class="literal">False</span>)</span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">forward</span>(<span class="params">self, x</span>):</span><br><span class="line">        gate = F.silu(self.gate_proj(x))</span><br><span class="line">        intermediate_states = self.up_proj(x)</span><br><span class="line">        <span class="keyword">return</span> self.down_proj(gate * intermediate_states)</span><br></pre></td></tr></table></figure>
<h4 id="组装-Decoder">组装 Decoder</h4>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">class</span> <span class="title class_">Decoder</span>(nn.Module):</span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">__init__</span>(<span class="params">self, hidden_size, n_heads, cacheKV, max_batch_size, max_seq_len</span>):</span><br><span class="line">        <span class="built_in">super</span>().__init__()</span><br><span class="line">        self.ln1 = RMSNorm(hidden_size)</span><br><span class="line">        self.attn = Attention(hidden_size, n_heads, cacheKV, max_batch_size, max_seq_len)</span><br><span class="line">        self.ln2 = RMSNorm(hidden_size)</span><br><span class="line">        self.mlp = MLP(hidden_size)</span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">forward</span>(<span class="params">self, x, rotary_emb, start_pos, mask=<span class="literal">None</span>, is_causal=<span class="literal">True</span></span>):</span><br><span class="line">        x = x + self.attn(self.ln1(x), rotary_emb, start_pos, mask, is_causal)</span><br><span class="line">        <span class="keyword">return</span> x + self.mlp(self.ln2(x))</span><br></pre></td></tr></table></figure>
<h3 id="组装模型">组装模型</h3>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">class</span> <span class="title class_">HelloGPT</span>(nn.Module):</span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">__init__</span>(<span class="params">self, vocab_size=<span class="number">32765</span>, hidden_size=<span class="number">768</span>, n_heads=<span class="number">12</span>, max_seq_len=<span class="number">1024</span>, n_layers=<span class="number">12</span>, cacheKV=<span class="literal">False</span>, max_batch_size=<span class="number">1</span></span>):</span><br><span class="line">        <span class="built_in">super</span>().__init__()</span><br><span class="line">        <span class="comment"># hidden_size &gt; 8.33 * ln(vocab_size)</span></span><br><span class="line">        self.tok_embeddings = nn.Embedding(vocab_size, hidden_size)</span><br><span class="line">        self.rotary_emb = RotaryEmbedding(hidden_size // n_heads, max_seq_len * <span class="number">2</span>)</span><br><span class="line">        self.rotary_emb.requires_grad = <span class="literal">False</span></span><br><span class="line">        self.layers = nn.ModuleList()</span><br><span class="line">        <span class="keyword">for</span> layer_id <span class="keyword">in</span> <span class="built_in">range</span>(n_layers):</span><br><span class="line">            self.layers.append(Decoder(hidden_size, n_heads, cacheKV, max_batch_size, max_seq_len))</span><br><span class="line">        self.norm = RMSNorm(hidden_size)</span><br><span class="line">        self.ln2 = nn.Linear(hidden_size, vocab_size, bias=<span class="literal">False</span>)</span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">forward</span>(<span class="params">self, input_ids: torch.Tensor, start_pos=<span class="number">0</span>, no_mask=<span class="literal">True</span></span>):</span><br><span class="line">        _bsz, seqlen = input_ids.shape</span><br><span class="line">        h = self.tok_embeddings(input_ids)</span><br><span class="line"></span><br><span class="line">        <span class="comment"># 预计算，减少每一层的重复计算</span></span><br><span class="line">        rotary_emb = self.rotary_emb(start_pos, seqlen)</span><br><span class="line">        <span class="keyword">for</span> layer <span class="keyword">in</span> self.layers:</span><br><span class="line">            h = layer(h, rotary_emb, start_pos)</span><br><span class="line"></span><br><span class="line">        h = self.norm(h)</span><br><span class="line">        h = self.ln2(h)</span><br><span class="line">        <span class="keyword">return</span> h.<span class="built_in">float</span>()</span><br></pre></td></tr></table></figure>
<h2 id="训练模型">训练模型</h2>
<h3 id="数据载入">数据载入</h3>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br></pre></td><td class="code"><pre><span class="line">data = Hcorpus(<span class="string">r&#x27;D:\datasets\h-corpus&#x27;</span>)</span><br><span class="line"><span class="keyword">def</span> <span class="title function_">get_batch</span>(<span class="params">size=<span class="number">512</span>, bsz=<span class="number">8</span></span>):</span><br><span class="line">    x = []</span><br><span class="line">    y = []</span><br><span class="line">    <span class="keyword">for</span> i <span class="keyword">in</span> <span class="built_in">range</span>(bsz):</span><br><span class="line">        tmp = data(size+<span class="number">1</span>)</span><br><span class="line">        x.append(tmp[:size])</span><br><span class="line">        y.append(tmp[<span class="number">1</span>:])</span><br><span class="line">    <span class="keyword">return</span> torch.tensor(x).to(device), torch.tensor(y).to(device)</span><br></pre></td></tr></table></figure>
<h3 id="模型载入">模型载入</h3>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br></pre></td><td class="code"><pre><span class="line">model = HelloGPT(n_layers=<span class="number">8</span>, max_seq_len=<span class="number">768</span>)</span><br><span class="line">model.to(device)</span><br></pre></td></tr></table></figure>
<h3 id="训练模型-2">训练模型</h3>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br></pre></td><td class="code"><pre><span class="line"><span class="comment">## 初始化训练器</span></span><br><span class="line">criterion = nn.CrossEntropyLoss()  <span class="comment"># 交叉熵损失函数</span></span><br><span class="line">optimizer = torch.optim.Adam(train_parameters, lr=<span class="number">6e-4</span>)  <span class="comment"># Adam 优化器</span></span><br><span class="line">scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=<span class="number">5</span>, T_mult=<span class="number">2</span>)  <span class="comment"># 余弦退火学习率</span></span><br><span class="line">torch.manual_seed(<span class="number">1337</span>)  <span class="comment"># 魔术随机种子</span></span><br><span class="line"></span><br><span class="line">total_loss = <span class="number">0</span></span><br><span class="line">print_iter = <span class="number">20</span></span><br><span class="line"><span class="keyword">for</span> epoch <span class="keyword">in</span> <span class="built_in">range</span>(<span class="number">1</span>, <span class="number">100001</span>):</span><br><span class="line">    optimizer.zero_grad(set_to_none=<span class="literal">True</span>)  <span class="comment"># 清空梯度，节省显存</span></span><br><span class="line">    x, y = get_batch(size=<span class="number">384</span>, bsz=<span class="number">4</span>)  <span class="comment"># x 是训练语料 y 是 x 移动了一位，当做预测目标</span></span><br><span class="line">    y_ = model(x)  <span class="comment"># 通过 x 预测的 y</span></span><br><span class="line">    loss = criterion(y_.view(-<span class="number">1</span>, <span class="number">32765</span>), y.view(-<span class="number">1</span>))  <span class="comment"># 计算损失</span></span><br><span class="line">    loss.backward()  <span class="comment"># 反向传播梯度</span></span><br><span class="line">    torch.nn.utils.clip_grad_norm_(train_parameters, <span class="number">0.5</span>)  <span class="comment"># 梯度裁剪，减轻过拟合</span></span><br><span class="line">    optimizer.step()  <span class="comment"># 通过梯度优化训练参数</span></span><br><span class="line">    scheduler.step()  <span class="comment"># 计算下一步的学习率</span></span><br><span class="line">    total_loss += loss  <span class="comment"># 累计损失</span></span><br><span class="line"></span><br><span class="line">    <span class="keyword">if</span> epoch % print_iter == <span class="number">0</span>:</span><br><span class="line">        <span class="built_in">print</span>(data)</span><br><span class="line">        <span class="built_in">print</span>(<span class="string">f&#x27;epoch: <span class="subst">&#123;epoch&#125;</span>  lr: <span class="subst">&#123;scheduler.get_last_lr()[<span class="number">0</span>]:<span class="number">.4</span>e&#125;</span> loss: <span class="subst">&#123;total_loss / print_iter:<span class="number">.4</span>e&#125;</span>&#x27;</span>)</span><br><span class="line">        total_loss = <span class="number">0</span></span><br></pre></td></tr></table></figure>
<h3 id="保存读取">保存读取</h3>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">with</span> <span class="built_in">open</span>(<span class="string">&#x27;tmp_training.pkl&#x27;</span>, <span class="string">&#x27;rb&#x27;</span>) <span class="keyword">as</span> file:</span><br><span class="line">    epoch = pickle.load(file)  <span class="comment"># 读取 epoch 位置</span></span><br><span class="line">    tmp_fileset_idx = pickle.load(file)  <span class="comment"># 读取 data 位置</span></span><br><span class="line">    tmp_fileset_sub_idx = pickle.load(file)</span><br><span class="line"><span class="comment"># 恢复数据位置</span></span><br><span class="line">data = Hcorpus(<span class="string">r&#x27;D:\datasets\h-corpus&#x27;</span>, fileset_idx=tmp_fileset_idx-<span class="number">1</span>, fileset_sub_idx=tmp_fileset_sub_idx)</span><br><span class="line">model = torch.load(<span class="string">f&#x27;tmp_model_<span class="subst">&#123;epoch&#125;</span>.pth&#x27;</span>)  <span class="comment"># 恢复模型</span></span><br><span class="line"><span class="built_in">print</span>(<span class="string">f&#x27;start from epoch: <span class="subst">&#123;epoch&#125;</span>   data: <span class="subst">&#123;data&#125;</span>&#x27;</span>)</span><br><span class="line"></span><br><span class="line">save_iter = <span class="number">5000</span></span><br><span class="line"><span class="keyword">for</span> epoch <span class="keyword">in</span> <span class="built_in">range</span>(<span class="number">1</span>, <span class="number">100001</span>):</span><br><span class="line">    <span class="keyword">pass</span></span><br><span class="line">    <span class="keyword">if</span> epoch % save_iter == <span class="number">0</span>:</span><br><span class="line">        optimizer.zero_grad(set_to_none=<span class="literal">True</span>)  <span class="comment"># 清空梯度，节省显存</span></span><br><span class="line">        <span class="keyword">with</span> <span class="built_in">open</span>(<span class="string">&#x27;tmp_training.pkl&#x27;</span>, <span class="string">&#x27;wb&#x27;</span>) <span class="keyword">as</span> file:</span><br><span class="line">            pickle.dump(epoch, file)  <span class="comment"># 保存 epoch 位置</span></span><br><span class="line">            pickle.dump(data.fileset_idx, file)  <span class="comment"># 保存 data 位置</span></span><br><span class="line">            pickle.dump(data.fileset_sub_idx, file)</span><br><span class="line">        torch.save(model, <span class="string">f&#x27;tmp_model_<span class="subst">&#123;epoch&#125;</span>.pth&#x27;</span>)  <span class="comment"># 保存模型</span></span><br><span class="line">        <span class="built_in">print</span>(<span class="string">f&#x27;save to tmp_model_<span class="subst">&#123;epoch&#125;</span>.pth&#x27;</span>)</span><br></pre></td></tr></table></figure>
<h3 id="可视化">可视化</h3>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br></pre></td><td class="code"><pre><span class="line">writer = SummaryWriter(<span class="string">&#x27;logs&#x27;</span>)  <span class="comment"># tensorboard --logdir logs</span></span><br><span class="line"><span class="keyword">for</span> epoch <span class="keyword">in</span> <span class="built_in">range</span>(<span class="number">1</span>, <span class="number">100001</span>):</span><br><span class="line">    <span class="keyword">pass</span></span><br><span class="line">    writer.add_scalar(<span class="string">&#x27;lr&#x27;</span>, scheduler.get_last_lr()[<span class="number">0</span>], epoch)</span><br><span class="line">    writer.add_scalar(<span class="string">&#x27;loss&#x27;</span>, loss, epoch)</span><br><span class="line">    <span class="keyword">if</span> epoch % print_iter == <span class="number">0</span>:</span><br><span class="line">        <span class="keyword">pass</span></span><br><span class="line">        writer.add_scalar(<span class="string">&#x27;total_loss&#x27;</span>, total_loss / print_iter, epoch)</span><br><span class="line">writer.close()</span><br></pre></td></tr></table></figure>
<h2 id="附加-streaming-llm">附加 streaming_llm</h2>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br><span class="line">33</span><br><span class="line">34</span><br><span class="line">35</span><br><span class="line">36</span><br><span class="line">37</span><br><span class="line">38</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">class</span> <span class="title class_">RotaryEmbedding</span>(nn.Module):</span><br><span class="line">    <span class="keyword">pass</span></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">inverse_rotary_emb</span>(<span class="params">self, x</span>):</span><br><span class="line">        x_ = torch.view_as_complex(x.<span class="built_in">float</span>().reshape(*x.shape[:-<span class="number">1</span>], -<span class="number">1</span>, <span class="number">2</span>))</span><br><span class="line">        x_out = torch.view_as_real(x_ * self.local_freqs_cis_inverse).flatten(<span class="number">3</span>)</span><br><span class="line">        <span class="keyword">return</span> x_out.type_as(x)</span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">inverse_forward</span>(<span class="params">self, start_pos: <span class="built_in">int</span>, seqlen: <span class="built_in">int</span></span>):</span><br><span class="line">        self.local_freqs_cis_inverse = self.freqs_cis[start_pos: start_pos + seqlen].view(<span class="number">1</span>, seqlen, <span class="number">1</span>, -<span class="number">1</span>)  <span class="comment"># cacheKV 相关，可忽略</span></span><br><span class="line">        self.local_freqs_cis_inverse = self.local_freqs_cis_inverse.conj()  <span class="comment"># 乘上共轭就旋转回去了</span></span><br><span class="line">        self.local_freqs_cis.requires_grad = <span class="literal">False</span></span><br><span class="line">        <span class="keyword">return</span> self.inverse_rotary_emb</span><br><span class="line"></span><br><span class="line"><span class="keyword">class</span> <span class="title class_">Attention</span>(nn.Module):</span><br><span class="line">    <span class="keyword">pass</span></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">forward</span>(<span class="params">self, hidden_states, rotary_emb, start_pos=<span class="number">0</span>, mask=<span class="literal">None</span>, is_causal=<span class="literal">True</span></span>):</span><br><span class="line">        <span class="keyword">pass</span></span><br><span class="line">        <span class="keyword">if</span> self.cacheKV:  <span class="comment"># cacheKV 相关，可忽略</span></span><br><span class="line">            self.cache_k[:bsz, start_pos: start_pos + seqlen] = k</span><br><span class="line">            self.cache_v[:bsz, start_pos: start_pos + seqlen] = v</span><br><span class="line">            k = self.cache_k[:bsz, : start_pos + seqlen]</span><br><span class="line">            v = self.cache_v[:bsz, : start_pos + seqlen]</span><br><span class="line"></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">streaming_llm</span>(<span class="params">self, start_pos, seqlen, to_pos, inverse_rotary_emb, rotary_emb, bsz</span>):</span><br><span class="line">        k = self.cache_k[:bsz, start_pos: start_pos + seqlen]</span><br><span class="line">        v = self.cache_v[:bsz, start_pos: start_pos + seqlen]</span><br><span class="line">        k = inverse_rotary_emb(k)</span><br><span class="line">        k = rotary_emb(k)</span><br><span class="line">        self.cache_k[:bsz, to_pos: to_pos + seqlen] = k</span><br><span class="line">        self.cache_v[:bsz, to_pos: to_pos + seqlen] = v</span><br><span class="line"></span><br><span class="line"><span class="keyword">class</span> <span class="title class_">HelloGPT</span>(nn.Module):</span><br><span class="line">    <span class="keyword">pass</span></span><br><span class="line">    <span class="keyword">def</span> <span class="title function_">streaming_llm</span>(<span class="params">self, start_pos, seqlen, to_pos, max_batch_size=<span class="number">1</span></span>):</span><br><span class="line">        rotary_emb = self.rotary_emb(to_pos, seqlen)</span><br><span class="line">        inverse_rotary_emb = self.rotary_emb.inverse_forward(start_pos, seqlen)</span><br><span class="line">        <span class="keyword">for</span> layer <span class="keyword">in</span> self.layers:</span><br><span class="line">            layer.attn.streaming_llm(start_pos, seqlen, to_pos, inverse_rotary_emb, rotary_emb, max_batch_size)</span><br></pre></td></tr></table></figure>
                
              </div>
            
            <hr/>
            <div>
              <div class="post-metas my-3">
  
  
    <div class="post-meta">
      <i class="iconfont icon-tags"></i>
      
        <a href="/tags/%E6%8E%A2%E7%B4%A2/" class="print-no-link">#探索</a>
      
        <a href="/tags/llama/" class="print-no-link">#llama</a>
      
    </div>
  
</div>


  <div class="license-box my-3">
    <div class="license-title">
      <div>【探索】从零开始训练 GPT</div>
      <div>https://hexo.limour.top/training-gpt-from-scratch</div>
    </div>
    <div class="license-meta">
      
        <div class="license-meta-item">
          <div>Author</div>
          <div>Limour</div>
        </div>
      
      
        <div class="license-meta-item license-meta-date">
          <div>Posted on</div>
          <div>January 18, 2024</div>
        </div>
      
      
        <div class="license-meta-item license-meta-date">
          <div>Updated on</div>
          <div>March 19, 2024</div>
        </div>
      
      
        <div class="license-meta-item">
          <div>Licensed under</div>
          <div>
            
              
                <a class="print-no-link" target="_blank" href="https://creativecommons.org/licenses/by-nc-sa/4.0/">
                  <span class="hint--top hint--rounded" aria-label="BY - Attribution">
                    <i class="iconfont icon-by"></i>
                  </span>
                </a>
              
                <a class="print-no-link" target="_blank" href="https://creativecommons.org/licenses/by-nc-sa/4.0/">
                  <span class="hint--top hint--rounded" aria-label="NC - Non-commercial">
                    <i class="iconfont icon-nc"></i>
                  </span>
                </a>
              
                <a class="print-no-link" target="_blank" href="https://creativecommons.org/licenses/by-nc-sa/4.0/">
                  <span class="hint--top hint--rounded" aria-label="SA - Share-alike">
                    <i class="iconfont icon-sa"></i>
                  </span>
                </a>
              
            
          </div>
        </div>
      
    </div>
    <div class="license-icon iconfont"></div>
  </div>


                <div class="post-prevnext my-3">
                  <article class="post-prev col-6">
                    
                    
                      <a href="/Convert-BlueLM-7B-Chat-to-the-standard-GGUF-model" title="【探索】将BlueLM-7B-Chat转换为标准的GGUF模型">
                        <i class="iconfont icon-arrowleft"></i>
                        <span class="hidden-mobile">【探索】将BlueLM-7B-Chat转换为标准的GGUF模型</span>
                        <span class="visible-mobile">Previous</span>
                      </a>
                    
                  </article>
                  <article class="post-next col-6">
                    
                    
                      <a href="/Azure-AI-prevents-reverse-wool-shearing" title="【避坑】Azure AI 避免反向薅羊毛">
                        <span class="hidden-mobile">【避坑】Azure AI 避免反向薅羊毛</span>
                        <span class="visible-mobile">Next</span>
                        <i class="iconfont icon-arrowright"></i>
                      </a>
                    
                  </article>
                </div>
              
            </div>

            
    <article id="comments" lazyload>
      
  <div id="waline"></div>
  <script type="text/javascript">
    Fluid.utils.loadComments('#waline', function() {
      Fluid.utils.createCssLink('https://cdn.staticfile.org/waline/2.15.5/waline.min.css')
      Fluid.utils.createScript('https://cdn.staticfile.org/waline/2.15.5/waline.min.js', function() {
        var options = Object.assign(
          {"serverURL":"https://comments.limour.top","path":"window.location.pathname","meta":["nick","mail","link"],"requiredMeta":["nick"],"lang":"zh-CN","emoji":["https://jscdn.limour.top/gh/walinejs/emojis/weibo"],"dark":"html[data-user-color-scheme=\"dark\"]","wordLimit":0,"pageSize":10},
          {
            el: '#waline',
            path: window.location.pathname
          }
        )
        Waline.init(options);
        Fluid.utils.waitElementVisible('#waline .vcontent', () => {
          var imgSelector = '#waline .vcontent img:not(.vemoji)';
          Fluid.plugins.imageCaption(imgSelector);
          Fluid.plugins.fancyBox(imgSelector);
        })
      });
    });
  </script>
  <noscript>Please enable JavaScript to view the comments</noscript>


    </article>
  

          </article>
        </div>
      </div>
    </div>

    <div class="side-col d-none d-lg-block col-lg-2">
      
  <aside class="sidebar" style="margin-left: -1rem">
    <div id="toc">
  <p class="toc-header">
    <i class="iconfont icon-list"></i>
    <span>Table of Contents</span>
  </p>
  <div class="toc-body" id="toc-body"></div>
</div>


  </aside>


    </div>
  </div>
</div>


      <a id="scroll-top-button" aria-label="TOP" href="#" role="button">
        <i class="iconfont icon-arrowup" aria-hidden="true"></i>
      </a>
    

      <div class="modal fade" id="modalSearch" tabindex="-1" role="dialog" aria-labelledby="ModalLabel"
     aria-hidden="true">
  <div class="modal-dialog modal-dialog-scrollable modal-lg" role="document">
    <div class="modal-content">
      <div class="modal-header text-center">
        <h4 class="modal-title w-100 font-weight-bold">Search</h4>
        <button type="button" id="local-search-close" class="close" data-dismiss="modal" aria-label="Close">
          <span aria-hidden="true">&times;</span>
        </button>
      </div>
      <div class="modal-body mx-3">
        <div class="md-form mb-5">
          <input type="text" id="local-search-input" class="form-control validate">
          <label data-error="x" data-success="v" for="local-search-input">Keyword</label>
        </div>
        <div class="list-group" id="local-search-result"></div>
      </div>
    </div>
  </div>
</div>

    
  </main>

  <footer>
    <div class="footer-inner">
  
    <div class="footer-content">
       <a target="_blank" rel="nofollow noopener" href="http://www.beian.gov.cn/portal/registerSystemInfo?recordcode=43130202000203"><img src="https://img.limour.top/2023/08/27/64eadeb81d6a0.webp" srcset="https://jscdn.limour.top/gh/Limour-dev/Sakurairo_Vision/load_svg/inload.svg" lazyload>湘公网安备43130202000203号  </a> <a target="_blank" rel="nofollow noopener" href="https://beian.miit.gov.cn/">湘ICP备20008299号 </a> <a target="_blank" rel="nofollow noopener" href="https://icp.gov.moe/?keyword=20210128">萌ICP备20210128号</a> <br> <a href="https://www.foreverblog.cn/" target="_blank"> <img src="https://img.foreverblog.cn/logo_en_default.png" srcset="https://jscdn.limour.top/gh/Limour-dev/Sakurairo_Vision/load_svg/inload.svg" lazyload alt="" style="width:auto;height:24px"> </a> <br> <a href="https://hexo.io" target="_blank" rel="nofollow noopener"><span>Hexo</span></a> <i class="iconfont icon-love"></i> <a href="https://github.com/fluid-dev/hexo-theme-fluid" target="_blank" rel="nofollow noopener"><span>Fluid</span></a> <i class="iconfont icon-love"></i> <a href="https://github.com/limour-blog/limour-blog.github.io" target="_blank" rel="nofollow noopener"><span>SRC</span></a> <i class="iconfont icon-love"></i> <a href="https://web.archive.org/web/20231130095837/https://effectiveacceleration.tech/" target="_blank" rel="nofollow noopener"><span>e/Acc</span></a> 
    </div>
  
  
</div>

  </footer>

  <!-- Scripts -->
  
  <script  src="https://jscdn.limour.top/npm/nprogress@0.2.0/nprogress.min.js" ></script>
  <link  rel="stylesheet" href="https://jscdn.limour.top/npm/nprogress@0.2.0/nprogress.min.css" />

  <script>
    NProgress.configure({"showSpinner":false,"trickleSpeed":100})
    NProgress.start()
    window.addEventListener('load', function() {
      NProgress.done();
    })
  </script>


<script  src="https://jscdn.limour.top/npm/jquery@3.6.4/dist/jquery.min.js" ></script>
<script  src="https://jscdn.limour.top/npm/bootstrap@4.6.1/dist/js/bootstrap.min.js" ></script>
<script  src="/js/events.js" ></script>
<script  src="/js/plugins.js" ></script>


    <script  src="/js/img-lazyload.js" ></script>
  

<script>
  Fluid.utils.createScript('https://jscdn.limour.top/npm/tocbot@4.20.1/dist/tocbot.min.js', function() {
    var toc = jQuery('#toc');
    if (toc.length === 0 || !window.tocbot) { return; }
    var boardCtn = jQuery('#board-ctn');
    var boardTop = boardCtn.offset().top;

    window.tocbot.init(Object.assign({
      tocSelector     : '#toc-body',
      contentSelector : '.markdown-body',
      linkClass       : 'tocbot-link',
      activeLinkClass : 'tocbot-active-link',
      listClass       : 'tocbot-list',
      isCollapsedClass: 'tocbot-is-collapsed',
      collapsibleClass: 'tocbot-is-collapsible',
      scrollSmooth    : true,
      includeTitleTags: true,
      headingsOffset  : -boardTop,
    }, CONFIG.toc));
    if (toc.find('.toc-list-item').length > 0) {
      toc.css('visibility', 'visible');
    }

    Fluid.events.registerRefreshCallback(function() {
      if ('tocbot' in window) {
        tocbot.refresh();
        var toc = jQuery('#toc');
        if (toc.length === 0 || !tocbot) {
          return;
        }
        if (toc.find('.toc-list-item').length > 0) {
          toc.css('visibility', 'visible');
        }
      }
    });
  });
</script>


  <script src=https://lib.baomitu.com/clipboard.js/2.0.11/clipboard.min.js></script>

  <script>Fluid.plugins.codeWidget();</script>


<script>
  Fluid.utils.createScript('https://jscdn.limour.top/npm/anchor-js@4.3.1/anchor.min.js', function() {
    window.anchors.options = {
      placement: CONFIG.anchorjs.placement,
      visible  : CONFIG.anchorjs.visible
    };
    if (CONFIG.anchorjs.icon) {
      window.anchors.options.icon = CONFIG.anchorjs.icon;
    }
    var el = (CONFIG.anchorjs.element || 'h1,h2,h3,h4,h5,h6').split(',');
    var res = [];
    for (var item of el) {
      res.push('.markdown-body > ' + item.trim());
    }
    if (CONFIG.anchorjs.placement === 'left') {
      window.anchors.options.class = 'anchorjs-link-left';
    }
    window.anchors.add(res.join(', '));

    Fluid.events.registerRefreshCallback(function() {
      if ('anchors' in window) {
        anchors.removeAll();
        var el = (CONFIG.anchorjs.element || 'h1,h2,h3,h4,h5,h6').split(',');
        var res = [];
        for (var item of el) {
          res.push('.markdown-body > ' + item.trim());
        }
        if (CONFIG.anchorjs.placement === 'left') {
          anchors.options.class = 'anchorjs-link-left';
        }
        anchors.add(res.join(', '));
      }
    });
  });
</script>


  <script>Fluid.plugins.imageCaption();</script>

  <script  src="/js/local-search.js" ></script>


<!-- 主题的启动项，将它保持在最底部 -->
<!-- the boot of the theme, keep it at the bottom -->
<script  src="/js/boot.js" ></script>


  <noscript>
    <div class="noscript-warning">Blog works best with JavaScript enabled</div>
  </noscript>
<!-- hexo injector body_end start -->
  <script defer src="/theme-inject/timeliness.js"></script>
<!-- hexo injector body_end end --></body>
</html>