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


			
							123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746
							

<!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="Llama.cpp 能 CPU &amp;amp; GPU 环境混合推理，这里记录一下在 windows10 平台上运行 Qwen-1.8B 的过程，显卡是 1660Ti 。 准备模型  安装conda Tun模式(管理员权限)  1234567conda create -n llamaConvert python=3.10 git -c conda-forgeconda activate llamaCo">
  
  
  <title>【记录】win10平台6G显存运行Qwen-1.8B - 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/Running-Qwen-on-the-Win10-platform-with-6GB-of-video-memory"/>
<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">【记录】win10平台6G显存运行Qwen-1.8B</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="2024-01-01 11:11" pubdate>
          January 1, 2024 am
        </time>
      </span>
    
  </div>

  <div class="mt-1">
    
      <span class="post-meta mr-2">
        <i class="iconfont icon-chart"></i>
        
          1.2k words
        
      </span>
    

      <span class="post-meta mr-2">
        <i class="iconfont icon-clock-fill"></i>
        
        
          10 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">【记录】win10平台6G显存运行Qwen-1.8B</h1>
            
              <p id="updated-time" class="note note-info" style="">
                
                  
                    Last updated on March 29, 2024 am
                  
                
              </p>
            
            
              <div class="markdown-body">
                
                <p><a href="https://hexo.limour.top/go/#aHR0cHM6Ly9naXRodWIuY29tL2dnZXJnYW5vdi9sbGFtYS5jcHA=" rel="noopener external nofollow noreferrer">Llama.cpp</a> 能 CPU &amp; GPU 环境混合推理，这里记录一下在 windows10 平台上运行 <a href="https://hexo.limour.top/go/#aHR0cHM6Ly9odWdnaW5nZmFjZS5jby9Rd2VuL1F3ZW4tMV84Qg==" rel="noopener external nofollow noreferrer">Qwen-1.8B</a> 的过程，显卡是 1660Ti 。</p>
<h2 id="准备模型">准备模型</h2>
<ul>
<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">Tun模式</a>(管理员权限)</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></pre></td><td class="code"><pre><code class="hljs powershell">conda create <span class="hljs-literal">-n</span> llamaConvert python=<span class="hljs-number">3.10</span> git <span class="hljs-literal">-c</span> conda<span class="hljs-literal">-forge</span><br>conda activate llamaConvert<br><span class="hljs-built_in">cd</span> D:\llama<br>git clone <span class="hljs-literal">--depth</span>=<span class="hljs-number">1</span> https://github.com/ggerganov/llama.cpp.git<br><span class="hljs-built_in">cd</span> llama.cpp<br>python <span class="hljs-literal">-m</span> pip install <span class="hljs-literal">-r</span> requirements.txt<br>pip install tiktoken<br></code></pre></td></tr></table></figure>
<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></pre></td><td class="code"><pre><code class="hljs powershell">python <span class="hljs-literal">-c</span> <span class="hljs-string">&quot;from huggingface_hub import snapshot_download; snapshot_download(repo_id=&#x27;Qwen/Qwen-1_8B-Chat&#x27;, local_dir=r&#x27;D:\qwen&#x27;, ignore_patterns=[&#x27;*.h5&#x27;, &#x27;*.ot&#x27;, &#x27;*.msgpack&#x27;, &#x27;*.safetensors&#x27;])&quot;</span><br><span class="hljs-built_in">cd</span> D:\qwen<br>D:\aria2\aria2c.exe <span class="hljs-literal">--all-proxy</span>=<span class="hljs-string">&#x27;http://127.0.0.1:7890&#x27;</span> <span class="hljs-literal">-o</span> <span class="hljs-string">&#x27;model-00001-of-00002.safetensors&#x27;</span> <span class="hljs-string">&quot;https://huggingface.co/Qwen/Qwen-1_8B-Chat/resolve/main/model-00001-of-00002.safetensors?download=true&quot;</span><br>D:\aria2\aria2c.exe <span class="hljs-literal">--all-proxy</span>=<span class="hljs-string">&#x27;http://127.0.0.1:7890&#x27;</span> <span class="hljs-literal">-o</span> <span class="hljs-string">&#x27;model-00002-of-00002.safetensors&#x27;</span> <span class="hljs-string">&quot;https://huggingface.co/Qwen/Qwen-1_8B-Chat/resolve/main/model-00002-of-00002.safetensors?download=true&quot;</span><br></code></pre></td></tr></table></figure>
<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"><span class="hljs-built_in">cd</span> D:\llama\llama.cpp<br>python <span class="hljs-built_in">convert-hf</span><span class="hljs-literal">-to-gguf</span>.py D:\qwen<br><span class="hljs-comment"># Model successfully exported to &#x27;D:\qwen\ggml-model-f16.gguf&#x27;</span><br></code></pre></td></tr></table></figure>
<h2 id="运行模型">运行模型</h2>
<ul>
<li>下载 <a href="https://hexo.limour.top/go/#aHR0cHM6Ly9naXRodWIuY29tL2dnZXJnYW5vdi9sbGFtYS5jcHAvcmVsZWFzZXM=" rel="noopener external nofollow noreferrer">llama-b1732-bin-win-cublas-cu12.2.0-x64.zip</a></li>
<li>提取文件到 <code>D:\llama</code></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></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><span class="hljs-built_in">cd</span> D:\llama ; .\main.exe <span class="hljs-comment">## 检查能否正确运行</span><br><span class="hljs-built_in">cd</span> D:\llama ; .\quantize.exe <span class="hljs-literal">--help</span> <span class="hljs-comment">## 自己决定量化方式</span><br>.\quantize.exe D:\qwen\ggml<span class="hljs-literal">-model-f16</span>.gguf .\qwen<span class="hljs-literal">-1_8-f16</span>.gguf <span class="hljs-built_in">COPY</span><br>.\server.exe <span class="hljs-literal">-m</span> .\qwen<span class="hljs-literal">-1_8-f16</span>.gguf <span class="hljs-literal">-c</span> <span class="hljs-number">4096</span> <span class="hljs-literal">--n-gpu-layers</span> <span class="hljs-number">50</span> <span class="hljs-comment">## 调节 n-gpu-layers 平衡 CPU &amp; GPU</span><br></code></pre></td></tr></table></figure>
<ul>
<li>访问 <code>http://127.0.0.1:8080</code> 选择 <code>Completion</code> 进行测试</li>
</ul>
<h2 id="微调模型">微调模型</h2>
<ul>
<li><a href="https://hexo.limour.top/go/#aHR0cHM6Ly9odWdnaW5nZmFjZS5jby9kYXRhc2V0cy9hNjg2ZDM4MC9oLWNvcnB1cy0yMDIz" rel="noopener external nofollow noreferrer">h-corpus数据集</a></li>
<li><a href="https://hexo.limour.top/go/#aHR0cHM6Ly9naXRodWIuY29tL1F3ZW5MTS9Rd2VuL2Jsb2IvbWFpbi9SRUFETUVfQ04ubWQjJUU1JUJFJUFFJUU4JUIwJTgz" rel="noopener external nofollow noreferrer">官方微调教程</a></li>
</ul>
<h2 id="附加-Yi-6B-Chat">附加 Yi-6B-Chat</h2>
<p><a href="https://hexo.limour.top/go/#aHR0cHM6Ly9odWdnaW5nZmFjZS5jby8wMS1haS9ZaS02Qi1DaGF0" rel="noopener external nofollow noreferrer">Yi-6B</a>是零一万物开源的双语语言模型，经过3T多语种语料库的训练，在语言理解、常识推理、阅读理解等方面有一定潜力。</p>
<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></pre></td><td class="code"><pre><code class="hljs powershell"><span class="hljs-built_in">cd</span> D:\models\<span class="hljs-number">01</span>yi<br>D:\aria2\aria2c.exe <span class="hljs-literal">--all-proxy</span>=<span class="hljs-string">&#x27;http://127.0.0.1:7890&#x27;</span> <span class="hljs-literal">-o</span> <span class="hljs-string">&#x27;model-00001-of-00003.safetensors&#x27;</span> <span class="hljs-string">&quot;https://huggingface.co/01-ai/Yi-6B-Chat/resolve/main/model-00001-of-00003.safetensors?download=true&quot;</span><br>D:\aria2\aria2c.exe <span class="hljs-literal">--all-proxy</span>=<span class="hljs-string">&#x27;http://127.0.0.1:7890&#x27;</span> <span class="hljs-literal">-o</span> <span class="hljs-string">&#x27;model-00002-of-00003.safetensors&#x27;</span> <span class="hljs-string">&quot;https://huggingface.co/01-ai/Yi-6B-Chat/resolve/main/model-00002-of-00003.safetensors?download=true&quot;</span><br>D:\aria2\aria2c.exe <span class="hljs-literal">--all-proxy</span>=<span class="hljs-string">&#x27;http://127.0.0.1:7890&#x27;</span> <span class="hljs-literal">-o</span> <span class="hljs-string">&#x27;model-00003-of-00003.safetensors&#x27;</span>  https://huggingface.co/<span class="hljs-number">01</span><span class="hljs-literal">-ai</span>/Yi<span class="hljs-literal">-6B-Chat</span>/resolve/main/model<span class="hljs-literal">-00003-of-00003</span>.safetensors?download=true<br>conda activate llamaConvert<br>python <span class="hljs-literal">-c</span> <span class="hljs-string">&quot;from huggingface_hub import snapshot_download; snapshot_download(repo_id=&#x27;01-ai/Yi-6B-Chat&#x27;, local_dir=r&#x27;D:\models\01yi&#x27;, ignore_patterns=[&#x27;*.h5&#x27;, &#x27;*.ot&#x27;, &#x27;*.msgpack&#x27;, &#x27;*.safetensors&#x27;])&quot;</span><br></code></pre></td></tr></table></figure>
<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></pre></td><td class="code"><pre><code class="hljs powershell">conda activate llamaConvert<br><span class="hljs-built_in">cd</span> D:\llama\llama.cpp<br>python convert.py D:\models\<span class="hljs-number">01</span>yi<br><span class="hljs-comment"># Wrote D:\models\01yi\ggml-model-f16.gguf</span><br>conda activate llamaCpp<br><span class="hljs-built_in">cd</span> D:\llama ; .\quantize.exe <span class="hljs-literal">--help</span><br>.\quantize.exe D:\models\<span class="hljs-number">01</span>yi\ggml<span class="hljs-literal">-model-f16</span>.gguf .\<span class="hljs-number">01</span>yi<span class="hljs-literal">-6b-Q4_K_M</span>.gguf Q4_K_M<br>.\server.exe <span class="hljs-literal">-m</span> .\<span class="hljs-number">01</span>yi<span class="hljs-literal">-6b-Q4_K_M</span>.gguf <span class="hljs-literal">-c</span> <span class="hljs-number">4096</span> <span class="hljs-literal">--n-gpu-layers</span> <span class="hljs-number">50</span><br></code></pre></td></tr></table></figure>
<h2 id="附加-百川2">附加 百川2</h2>
<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><span class="line">10</span><br><span class="line">11</span><br></pre></td><td class="code"><pre><code class="hljs powershell"><span class="hljs-built_in">cd</span> D:\models\baichuan<br>D:\aria2\aria2c.exe <span class="hljs-literal">--all-proxy</span>=<span class="hljs-string">&#x27;http://127.0.0.1:7890&#x27;</span> <span class="hljs-literal">-o</span> <span class="hljs-string">&#x27;pytorch_model.bin&#x27;</span> <span class="hljs-string">&quot;https://huggingface.co/baichuan-inc/Baichuan2-7B-Chat/resolve/main/pytorch_model.bin?download=true&quot;</span><br>conda activate llamaConvert<br>python <span class="hljs-literal">-c</span> <span class="hljs-string">&quot;from huggingface_hub import snapshot_download; snapshot_download(repo_id=&#x27;baichuan-inc/Baichuan2-7B-Chat&#x27;, local_dir=r&#x27;D:\models\baichuan&#x27;, ignore_patterns=[&#x27;*.h5&#x27;, &#x27;*.bin&#x27;, &#x27;*.ot&#x27;, &#x27;*.msgpack&#x27;, &#x27;*.safetensors&#x27;])&quot;</span><br><span class="hljs-built_in">cd</span> D:\llama\llama.cpp<br>python convert.py D:\models\baichuan<br><span class="hljs-comment"># Wrote D:\models\baichuan\ggml-model-f16.gguf</span><br>conda activate llamaCpp<br><span class="hljs-built_in">cd</span> D:\llama ; .\quantize.exe <span class="hljs-literal">--help</span><br>.\quantize.exe D:\models\baichuan\ggml<span class="hljs-literal">-model-f16</span>.gguf .\baichuan<span class="hljs-literal">-7b-Q3_K_M</span>.gguf Q3_K_M<br>.\server.exe <span class="hljs-literal">-m</span> .\baichuan<span class="hljs-literal">-7b-Q3_K_M</span>.gguf <span class="hljs-literal">-c</span> <span class="hljs-number">2048</span> <span class="hljs-literal">--n-gpu-layers</span> <span class="hljs-number">30</span><br></code></pre></td></tr></table></figure>
<h2 id="附加-tigerbot-13b">附加 tigerbot-13b</h2>
<p><a href="https://hexo.limour.top/go/#aHR0cHM6Ly9odWdnaW5nZmFjZS5jby9UaWdlclJlc2VhcmNoL3RpZ2VyYm90LTEzYi1jaGF0LXY1" rel="noopener external nofollow noreferrer">tigerbot-13b</a> 在 <a href="https://hexo.limour.top/go/#aHR0cHM6Ly9naXRodWIuY29tL2plaW5sZWUxOTkxL2NoaW5lc2UtbGxtLWJlbmNobWFyaw==" rel="noopener external nofollow noreferrer">chinese-llm-benchmark</a> 上排名靠前。</p>
<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><span class="line">10</span><br><span class="line">11</span><br></pre></td><td class="code"><pre><code class="hljs powershell"><span class="hljs-built_in">cd</span> D:\models\tigerbot<br>D:\aria2\aria2c.exe <span class="hljs-literal">--all-proxy</span>=<span class="hljs-string">&#x27;http://127.0.0.1:7890&#x27;</span> <span class="hljs-literal">-o</span> <span class="hljs-string">&#x27;pytorch_model-00001-of-00003.bin&#x27;</span> <span class="hljs-literal">--max-download-limit</span>=<span class="hljs-number">6</span>M <span class="hljs-string">&quot;https://huggingface.co/TigerResearch/tigerbot-13b-chat-v5/resolve/main/pytorch_model-00001-of-00003.bin?download=true&quot;</span><br>D:\aria2\aria2c.exe <span class="hljs-literal">--all-proxy</span>=<span class="hljs-string">&#x27;http://127.0.0.1:7890&#x27;</span> <span class="hljs-literal">-o</span> <span class="hljs-string">&#x27;pytorch_model-00002-of-00003.bin&#x27;</span> <span class="hljs-literal">--max-download-limit</span>=<span class="hljs-number">6</span>M <span class="hljs-string">&quot;https://huggingface.co/TigerResearch/tigerbot-13b-chat-v5/resolve/main/pytorch_model-00002-of-00003.bin?download=true&quot;</span><br>D:\aria2\aria2c.exe <span class="hljs-literal">--all-proxy</span>=<span class="hljs-string">&#x27;http://127.0.0.1:7890&#x27;</span> <span class="hljs-literal">-o</span> <span class="hljs-string">&#x27;pytorch_model-00003-of-00003.bin&#x27;</span> <span class="hljs-literal">--max-download-limit</span>=<span class="hljs-number">6</span>M <span class="hljs-string">&quot;https://huggingface.co/TigerResearch/tigerbot-13b-chat-v5/resolve/main/pytorch_model-00003-of-00003.bin?download=true&quot;</span><br>conda activate llamaConvert<br>python <span class="hljs-literal">-c</span> <span class="hljs-string">&quot;from huggingface_hub import snapshot_download; snapshot_download(repo_id=&#x27;TigerResearch/tigerbot-13b-chat-v5&#x27;, local_dir=r&#x27;D:\models\tigerbot&#x27;, ignore_patterns=[&#x27;*.h5&#x27;, &#x27;*.bin&#x27;, &#x27;*.ot&#x27;, &#x27;*.msgpack&#x27;, &#x27;*.safetensors&#x27;])&quot;</span><br><span class="hljs-built_in">cd</span> D:\llama\llama.cpp<br>python convert.py D:\models\tigerbot <span class="hljs-literal">--padvocab</span><br><span class="hljs-built_in">cd</span> D:\llama ; .\quantize.exe <span class="hljs-literal">--help</span><br>.\quantize.exe D:\models\tigerbot\ggml<span class="hljs-literal">-model-f16</span>.gguf D:\models\tigerbot<span class="hljs-literal">-13B-Chat-Q4_K_M</span>.gguf Q4_K_M<br>.\server.exe <span class="hljs-literal">-m</span> D:\models\tigerbot<span class="hljs-literal">-13B-Chat-Q4_K_M</span>.gguf <span class="hljs-literal">-c</span> <span class="hljs-number">4096</span><br></code></pre></td></tr></table></figure>
<div class="note note-info">
            <p>感觉 6G 显存下，比较好用的是 Yi-6B-Chat-Q4_K_M<br>tigerbot-13b 在 R5 5600H 上推理速度 4.6 tokens/s，CPU 使用率 60%，频率 3.5GHz，应该是内存带宽瓶颈</p>
          </div>
<h2 id="附加-在-Colab-上量化">附加 在 Colab 上量化</h2>
<ul>
<li><a href="https://hexo.limour.top/go/#aHR0cHM6Ly9jb2xhYi5yZXNlYXJjaC5nb29nbGUuY29tL2RyaXZlLzFKVDNYRmpEN0NUUkI5N3B1M1FwZUd1eldBMXlZRUFNNz91c3A9c2hhcmluZw==" rel="noopener external nofollow noreferrer">llm2gguf.ipynb</a></li>
<li><a href="https://hexo.limour.top/go/#aHR0cHM6Ly9odWdnaW5nZmFjZS5jby9MaW1vdXIvQ2F1c2FsTE0tMTRCLUdHVUY=" rel="noopener external nofollow noreferrer">量化后的结果</a></li>
</ul>
<h3 id="安装-llama-cpp">安装 llama.cpp</h3>
<figure class="highlight ipython"><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 ipython">!git clone --depth=<span class="hljs-number">1</span> https://github.com/ggerganov/llama.cpp.git<br>%cd /content/llama.cpp<br>!LLAMA_CUDA=<span class="hljs-number">1</span> make -j<br></code></pre></td></tr></table></figure>
<h3 id="计算-imatrix">计算 imatrix</h3>
<figure class="highlight ipython"><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></pre></td><td class="code"><pre><code class="hljs ipython">%cd /content<br>!wget -O transient.txt.gz https://huggingface.co/datasets/Limour/b-corpus/resolve/main/<span class="hljs-number">00</span>-preview/<span class="hljs-number">00</span>-transient.txt.gz?download=true<br>!gunzip transient.txt.gz<br>!mkdir -p /content/CausalLM-14B-GGUF<br>!wget -O /content/CausalLM-14B-GGUF/causallm_14b.Q8_0.gguf  https://huggingface.co/TheBloke/CausalLM-14B-GGUF/resolve/main/causallm_14b.Q8_0.gguf?download=true<br>!/content/llama.cpp/imatrix -m /content/CausalLM-14B-GGUF/causallm_14b.Q8_0.gguf -f /content/transient.txt -ngl <span class="hljs-number">36</span><br></code></pre></td></tr></table></figure>
<h3 id="登录拥抱脸">登录拥抱脸</h3>
<figure class="highlight ipython"><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></pre></td><td class="code"><pre><code class="hljs ipython"><span class="hljs-keyword">from</span> google.colab <span class="hljs-keyword">import</span> userdata<br><span class="hljs-keyword">from</span> huggingface_hub <span class="hljs-keyword">import</span> login<br><span class="hljs-comment"># login(token=os.environ.get(&quot;HF_TOKEN&quot;), write_permission=True)</span><br>login(token=userdata.get(<span class="hljs-string">&#x27;HF_TOKEN&#x27;</span>), write_permission=<span class="hljs-literal">True</span>)<br><span class="hljs-comment"># from huggingface_hub import notebook_login</span><br><span class="hljs-comment"># notebook_login()</span><br></code></pre></td></tr></table></figure>
<h3 id="跳过-转换模型">(跳过) 转换模型</h3>
<figure class="highlight ipython"><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></pre></td><td class="code"><pre><code class="hljs ipython">%cd llama.cpp<br>!python -m pip install -r requirements.txt<br>!pip install tiktoken<br><span class="hljs-keyword">from</span> huggingface_hub <span class="hljs-keyword">import</span> snapshot_download<br>!mkdir -p ~/CausalLM<br>snapshot_download(repo_id=<span class="hljs-string">&#x27;CausalLM/7B&#x27;</span>, local_dir=<span class="hljs-string">r&#x27;/content/CausalLM&#x27;</span>, ignore_patterns=[<span class="hljs-string">&#x27;*.h5&#x27;</span>, <span class="hljs-string">&#x27;*.ot&#x27;</span>, <span class="hljs-string">&#x27;*.msgpack&#x27;</span>, <span class="hljs-string">&#x27;*.safetensors&#x27;</span>])<br>!python convert.py --vocab-<span class="hljs-built_in">type</span> bpe --pad-vocab --outtype f16 /content/CausalLM <br></code></pre></td></tr></table></figure>
<h3 id="量化模型">量化模型</h3>
<figure class="highlight ipython"><table><tr><td class="gutter"><pre><span class="line">1</span><br></pre></td><td class="code"><pre><code class="hljs ipython">!/content/llama.cpp/quantize --allow-requantize --imatrix /content/imatrix.dat /content/CausalLM-14B-GGUF/causallm_14b.Q8_0.gguf /content/CausalLM-14B-GGUF/causallm_14b.IQ3_XS.gguf IQ3_XS<br></code></pre></td></tr></table></figure>
<h3 id="上传模型">上传模型</h3>
<figure class="highlight ipython"><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></pre></td><td class="code"><pre><code class="hljs ipython"><span class="hljs-keyword">from</span> huggingface_hub <span class="hljs-keyword">import</span> HfApi<br>api = HfApi()<br>api.upload_file(<br>    path_or_fileobj=<span class="hljs-string">&quot;/content/CausalLM-14B-GGUF/causallm_14b.IQ3_XS.gguf&quot;</span>,<br>    path_in_repo=<span class="hljs-string">&quot;causallm_14b.IQ3_XS.gguf&quot;</span>,<br>    repo_id=<span class="hljs-string">&quot;Limour/CausalLM-14B-GGUF&quot;</span><br>)<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/llama/" class="print-no-link">#llama</a>
      
    </div>
  
</div>


  <div class="license-box my-3">
    <div class="license-title">
      <div>【记录】win10平台6G显存运行Qwen-1.8B</div>
      <div>https://hexo.limour.top/Running-Qwen-on-the-Win10-platform-with-6GB-of-video-memory</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 1, 2024</div>
        </div>
      
      
        <div class="license-meta-item license-meta-date">
          <div>Updated on</div>
          <div>March 29, 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="/Azure-AI-prevents-reverse-wool-shearing" title="【避坑】Azure AI 避免反向薅羊毛">
                        <i class="iconfont icon-arrowleft"></i>
                        <span class="hidden-mobile">【避坑】Azure AI 避免反向薅羊毛</span>
                        <span class="visible-mobile">Previous</span>
                      </a>
                    
                  </article>
                  <article class="post-next col-6">
                    
                    
                      <a href="/Lightweight-personal-navigation-page-Flare" title="【记录】轻量个人导航页面 Flare">
                        <span class="hidden-mobile">【记录】轻量个人导航页面 Flare</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>