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


			
							123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733
							

<!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="角色扮演的体验是否舒适主要受角色卡、大模型和生成时间三个因素的影响。 优秀的角色卡往往附带大量的设定，这会极大的拖慢第一次生成的时间，并且随着对话的进行，上下文长度很容易超过kv_cache的上限，这些很破坏沉浸式的体验。 此外，大模型在进行角色扮演时，除了进行必要的对话生成外，还需要生成旁白增加想象空间。 对博主这些相比填空更喜欢选项的玩家，给出提问建议也是非常必要的：在建议的基础上修改比自己从">
  
  
  <title>【探索】6G显存畅玩无限长度的LLM角色扮演 - 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/c/font_1749284_5i9bdhy70f8.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":"1.9.8","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},"leancloud":{"app_id":null,"app_key":null,"server_url":null,"path":"window.location.pathname","ignore_local":false},"umami":{"src":null,"website_id":null,"domains":null,"start_time":"2024-01-01T00:00:00.000Z","token":null,"api_server":null},"woyaola":null,"cnzz":null},"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/Enjoy-unlimited-length-LLM-role-playing-with-6GB-of-VRAM"/>
<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】6G显存畅玩无限长度的LLM角色扮演</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="2024-02-10 09:02" pubdate>
          February 10, 2024 am
        </time>
      </span>
    
  </div>

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

      <span class="post-meta mr-2">
        <i class="iconfont icon-clock-fill"></i>
        
        
          8 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】6G显存畅玩无限长度的LLM角色扮演</h1>
            
              <p id="updated-time" class="note note-info" style="">
                
                  
                    Last updated on March 19, 2024 pm
                  
                
              </p>
            
            
              <div class="markdown-body">
                
                <p>角色扮演的体验是否舒适主要受角色卡、大模型和生成时间三个因素的影响。</p>
<p>优秀的角色卡往往附带大量的设定，这会极大的拖慢第一次生成的时间，并且随着对话的进行，上下文长度很容易超过kv_cache的上限，这些很破坏沉浸式的体验。</p>
<p>此外，大模型在进行角色扮演时，除了进行必要的对话生成外，还需要生成旁白增加想象空间。</p>
<p>对博主这些相比填空更喜欢选项的玩家，给出提问建议也是非常必要的：在建议的基础上修改比自己从零写一个情景更简单，同时也完整保留了控制剧情走向的权力。</p>
<p>以上这些都让本就稀缺的kv_cache更加雪上加霜。</p>
<p>万幸，StreamingLLM 发现了kv_cache具有良好的平移性，而 llama.cpp 也提供了对kv_cache进行底层操作的api：可以指定范围的 kv_cache_seq_rm 和 kv_cache_seq_shift。基于这两个api，我们将实现对kv_cache的 token 级微操，榨干kv_cache的全部价值。</p>
<p>博主实践表明，在充分利用kv_cache的基础上，哪怕是 huggingface space 免费的2vCPU容器也可以游玩角色扮演，而笔记本端6G显存的1660Ti可以做到畅玩角色扮演。</p>
<h2 id="体验-DEMO">体验 DEMO</h2>
<ul>
<li><a href="https://hexo.limour.top/go/#aHR0cHM6Ly9odWdnaW5nZmFjZS5jby9zcGFjZXMvTGltb3VyL2xsYW1hLXB5dGhvbi1zdHJlYW1pbmdsbG0=" rel="noopener external nofollow noreferrer">Limour/llama-python-streamingllm</a></li>
<li>同一时间仅支持一个人用，用之前点 Reset 按钮恢复初始的 kv_cache</li>
<li>按 Submit 没反应，说明有人在用，等一段时间后再 Reset</li>
<li>最好是 Duplicate 后，设为私密来使用</li>
</ul>
<h2 id="代码仓库">代码仓库</h2>
<ul>
<li><a href="https://hexo.limour.top/go/#aHR0cHM6Ly9naXRodWIuY29tL0xpbW91ci1kZXYvbGxhbWEtcHl0aG9uLXN0cmVhbWluZ2xsbQ==" rel="noopener external nofollow noreferrer">llama-python-streamingllm</a></li>
<li><a href="/-ji-lu--an-zhuang-conda-bing-geng-huan-qing-hua-yuan">安装conda</a></li>
<li><a href="/Use-Tunnel-to-speed-up-the-connection-of-VPS">学术上网</a>(管理员权限)</li>
<li>使用前需要修改 <code>rp_config.json</code> 里的模型路径和参数，指定为你已经下载了的<code>GGUF</code>格式模型的路径</li>
<li>推荐 <a href="https://hexo.limour.top/go/#aHR0cHM6Ly9odWdnaW5nZmFjZS5jby9UaGVCbG9rZS9DYXVzYWxMTS03Qi1HR1VGL2Jsb2IvbWFpbi9jYXVzYWxsbV83Yi5RNV9LX00uZ2d1Zg==" rel="noopener external nofollow noreferrer">causallm_7b.Q5_K_M.gguf</a></li>
<li>或者自己用 <a href="https://hexo.limour.top/go/#aHR0cHM6Ly9odWdnaW5nZmFjZS5jby9kYXRhc2V0cy9MaW1vdXIvYi1jb3JwdXM=" rel="noopener external nofollow noreferrer">Galgame</a> 解包的对话数据集微调一个合适的模型。</li>
</ul>
<h3 id="二选一：GPU版本的环境">二选一：GPU版本的环境</h3>
<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></pre></td><td class="code"><pre><code class="hljs powershell">conda create <span class="hljs-literal">-n</span> llamaCpp libcublas cuda<span class="hljs-literal">-toolkit</span> git <span class="hljs-literal">-c</span> nvidia <span class="hljs-literal">-c</span> conda<span class="hljs-literal">-forge</span><br>conda activate llamaCpp<br>conda install python=<span class="hljs-number">3.10</span> gradio <span class="hljs-literal">-c</span> conda<span class="hljs-literal">-forge</span><br><span class="hljs-comment"># 然后去 release 下载相应的包 https://github.com/Limour-dev/llama-cpp-python-cuBLAS-wheels/releases</span><br>pip install <span class="hljs-literal">--force-reinstall</span> llama_cpp_python<span class="hljs-literal">-0</span>.<span class="hljs-number">2.39</span>+cu122<span class="hljs-literal">-cp310-cp310-win_amd64</span>.whl<br></code></pre></td></tr></table></figure>
<h3 id="二选一：CPU版本的环境">二选一：CPU版本的环境</h3>
<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></pre></td><td class="code"><pre><code class="hljs powershell">conda create <span class="hljs-literal">-n</span> llamaCpp python=<span class="hljs-number">3.10</span> gradio git <span class="hljs-literal">-c</span> conda<span class="hljs-literal">-forge</span><br>conda activate llamaCpp<br>pip install llama<span class="hljs-literal">-cpp-python</span>==<span class="hljs-number">0.2</span>.<span class="hljs-number">39</span><br></code></pre></td></tr></table></figure>
<h3 id="下载并运行">下载并运行</h3>
<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></pre></td><td class="code"><pre><code class="hljs powershell">conda activate llamaCpp<br>git clone <span class="hljs-literal">--depth</span>=<span class="hljs-number">1</span> https://github.com/Limour<span class="hljs-literal">-dev</span>/llama<span class="hljs-literal">-python-streamingllm</span>.git<br><span class="hljs-built_in">cd</span> llama<span class="hljs-literal">-python-streamingllm</span><br>mkdir cache<br>python .\gradio_streamingllm.py<br></code></pre></td></tr></table></figure>
<h2 id="核心内容">核心内容</h2>
<ul>
<li><code>Submit</code> 会将 msg 发送给模型，然后流式生成回答</li>
<li><code>Retry</code> 会重新生成最近一次的 msg 所对应的回答</li>
<li><code>旁白</code> 会流式生成一份旁白到 <code>VO</code> 框</li>
<li><code>建议</code> 会以 usr 的身份流式生成一份 msg 供修改</li>
<li>上面四个功能的基础就是下面的基于 StreamingLLM 原理的 venv 开头的函数</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><span class="line">69</span><br><span class="line">70</span><br><span class="line">71</span><br><span class="line">72</span><br><span class="line">73</span><br><span class="line">74</span><br><span class="line">75</span><br><span class="line">76</span><br><span class="line">77</span><br><span class="line">78</span><br><span class="line">79</span><br><span class="line">80</span><br><span class="line">81</span><br><span class="line">82</span><br><span class="line">83</span><br><span class="line">84</span><br><span class="line">85</span><br></pre></td><td class="code"><pre><code class="hljs python"><span class="hljs-keyword">class</span> <span class="hljs-title class_">StreamingLLM</span>(<span class="hljs-title class_ inherited__">Llama</span>):<br>    <span class="hljs-keyword">pass</span><br>    <span class="hljs-keyword">def</span> <span class="hljs-title function_">kv_cache_seq_trim</span>(<span class="hljs-params">self</span>):<br>        self._ctx.kv_cache_seq_rm(-<span class="hljs-number">1</span>, self.n_tokens, -<span class="hljs-number">1</span>)<br><br>    <span class="hljs-keyword">def</span> <span class="hljs-title function_">kv_cache_seq_ltrim</span>(<span class="hljs-params">self, n_keep, n_discard=<span class="hljs-number">256</span>, n_past=-<span class="hljs-number">1</span></span>):<br>        <span class="hljs-keyword">if</span> n_past &lt; <span class="hljs-number">0</span>:<br>            n_past = self.n_tokens<br>        self._ctx.kv_cache_seq_rm(-<span class="hljs-number">1</span>, n_keep, n_keep + n_discard)<br>        self._ctx.kv_cache_seq_shift(<span class="hljs-number">0</span>, n_keep + n_discard, n_past, -n_discard)<br>        self.input_ids[n_keep:n_past - n_discard] = self.input_ids[n_keep + n_discard:n_past]<br>        self.n_tokens = n_past - n_discard<br><br>    <span class="hljs-keyword">def</span> <span class="hljs-title function_">_venv_init</span>(<span class="hljs-params">self</span>):<br>        self.venv = [<span class="hljs-number">0</span>]<br>        self.venv_idx_map = []<br><br>    <span class="hljs-keyword">def</span> <span class="hljs-title function_">venv_create</span>(<span class="hljs-params">self, name: <span class="hljs-built_in">str</span></span>):<br>        self.venv.append(<span class="hljs-number">0</span>)<br>        self.venv_idx_map.append(name)<br>        <span class="hljs-keyword">return</span> name<br><br>    <span class="hljs-keyword">def</span> <span class="hljs-title function_">venv_disband</span>(<span class="hljs-params">self, name_set</span>):<br>        <span class="hljs-keyword">if</span> <span class="hljs-built_in">len</span>(self.venv) &lt;= <span class="hljs-number">1</span>:<br>            <span class="hljs-keyword">return</span> <span class="hljs-literal">False</span><br>        name_set = &#123;x <span class="hljs-keyword">for</span> x <span class="hljs-keyword">in</span> name_set <span class="hljs-keyword">if</span> x <span class="hljs-keyword">in</span> self.venv_idx_map&#125;<br>        <span class="hljs-keyword">if</span> <span class="hljs-keyword">not</span> name_set:<br>            <span class="hljs-keyword">return</span> <span class="hljs-literal">False</span><br>        <span class="hljs-keyword">while</span> self.venv_idx_map:<br>            <span class="hljs-keyword">if</span> self.venv_idx_map[<span class="hljs-number">0</span>] <span class="hljs-keyword">in</span> name_set:<br>                self.venv_idx_map.pop(<span class="hljs-number">0</span>)  <span class="hljs-comment"># 删除</span><br>                tmp = self.venv.pop(<span class="hljs-number">1</span>)  <span class="hljs-comment"># 对应的 venv 移入上一层</span><br>                self.venv[<span class="hljs-number">0</span>] += tmp<br>            <span class="hljs-keyword">else</span>:<br>                <span class="hljs-keyword">break</span><br>        <span class="hljs-keyword">return</span> <span class="hljs-literal">True</span><br><br>    <span class="hljs-keyword">def</span> <span class="hljs-title function_">venv_revision</span>(<span class="hljs-params">self, name: <span class="hljs-built_in">str</span></span>):<br>        <span class="hljs-keyword">if</span> <span class="hljs-built_in">len</span>(self.venv) &lt;= <span class="hljs-number">1</span>:<br>            <span class="hljs-keyword">return</span> <span class="hljs-literal">False</span><br>        <span class="hljs-keyword">if</span> name <span class="hljs-keyword">not</span> <span class="hljs-keyword">in</span> self.venv_idx_map:<br>            <span class="hljs-keyword">return</span> <span class="hljs-literal">False</span><br>        _s = <span class="hljs-number">0</span><br>        <span class="hljs-keyword">while</span> self.venv_idx_map:<br>            <span class="hljs-keyword">if</span> self.venv_idx_map[-<span class="hljs-number">1</span>] == name:<br>                <span class="hljs-keyword">break</span><br>            self.venv_idx_map.pop()  <span class="hljs-comment"># 删除</span><br>            _s += self.venv.pop()<br>        <span class="hljs-keyword">if</span> _s:<br>            self.n_tokens -= <span class="hljs-built_in">min</span>(_s, self.n_tokens)<br>            self.kv_cache_seq_trim()<br>        <span class="hljs-keyword">return</span> <span class="hljs-literal">True</span><br><br>    <span class="hljs-keyword">def</span> <span class="hljs-title function_">venv_remove</span>(<span class="hljs-params">self, name: <span class="hljs-built_in">str</span></span>):<br>        <span class="hljs-keyword">if</span> <span class="hljs-built_in">len</span>(self.venv) &lt;= <span class="hljs-number">1</span>:<br>            <span class="hljs-keyword">return</span> <span class="hljs-literal">False</span><br>        <span class="hljs-keyword">if</span> name <span class="hljs-keyword">not</span> <span class="hljs-keyword">in</span> self.venv_idx_map:<br>            <span class="hljs-keyword">return</span> <span class="hljs-literal">False</span><br>        venv_idx = self.venv_idx_map.index(name) + <span class="hljs-number">1</span><br>        <span class="hljs-keyword">while</span> self.venv_idx_map:<br>            self.venv_idx_map.pop(venv_idx - <span class="hljs-number">1</span>)  <span class="hljs-comment"># 删除</span><br>            <span class="hljs-keyword">if</span> venv_idx == <span class="hljs-built_in">len</span>(self.venv) - <span class="hljs-number">1</span>:<br>                <span class="hljs-comment"># 最后一层</span><br>                self.n_tokens -= <span class="hljs-built_in">min</span>(self.venv.pop(), self.n_tokens)<br>                self.kv_cache_seq_trim()<br>                <span class="hljs-keyword">break</span><br>            <span class="hljs-keyword">else</span>:<br>                <span class="hljs-comment"># 非最后一层</span><br>                n_keep = self.n_tokens - <span class="hljs-built_in">sum</span>(self.venv[i] <span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(venv_idx, <span class="hljs-built_in">len</span>(self.venv)))<br>                n_discard = self.venv.pop(venv_idx)<br>                self.kv_cache_seq_ltrim(n_keep, n_discard)<br>                <span class="hljs-keyword">try</span>:<br>                    venv_idx = self.venv_idx_map.index(name, venv_idx - <span class="hljs-number">1</span>) + <span class="hljs-number">1</span><br>                <span class="hljs-keyword">except</span> ValueError:  <span class="hljs-comment"># 没有了</span><br>                    <span class="hljs-keyword">break</span><br>        <span class="hljs-keyword">return</span> <span class="hljs-literal">True</span><br><br>    <span class="hljs-keyword">def</span> <span class="hljs-title function_">eval_t</span>(<span class="hljs-params">self, tokens, n_keep=<span class="hljs-number">4</span>, n_discard=<span class="hljs-number">256</span>, im_start=<span class="hljs-literal">None</span></span>):<br>        <span class="hljs-keyword">if</span> self._n_ctx &lt; self.n_tokens + <span class="hljs-built_in">len</span>(tokens):<br>            tmp_n_discard = <span class="hljs-built_in">max</span>(n_discard, self.n_tokens + <span class="hljs-built_in">len</span>(tokens) - self._n_ctx)<br>            self.kv_cache_seq_ltrim(n_keep, tmp_n_discard)<br>        <span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-number">0</span>, <span class="hljs-built_in">len</span>(tokens), self.n_batch):<br>            <span class="hljs-keyword">pass</span><br>            self.n_tokens += n_tokens<br>            self.venv[-<span class="hljs-number">1</span>] += n_tokens<br></code></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】6G显存畅玩无限长度的LLM角色扮演</div>
      <div>https://hexo.limour.top/Enjoy-unlimited-length-LLM-role-playing-with-6GB-of-VRAM</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>February 10, 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-cc-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-cc-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-cc-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="/Sample-size-calculation-for-survival-analysis-in-clinical-research" title="【探索】暴力计算临床研究的样本量">
                        <i class="iconfont icon-arrowleft"></i>
                        <span class="hidden-mobile">【探索】暴力计算临床研究的样本量</span>
                        <span class="visible-mobile">Previous</span>
                      </a>
                    
                  </article>
                  <article class="post-next 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">
                        <span class="hidden-mobile">【探索】将BlueLM-7B-Chat转换为标准的GGUF模型</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.css')
      Fluid.utils.createScript('https://cdn.staticfile.org/waline/2.15.5/waline.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>