https://example.org/ Recent content on 前海分享 Hugo en-us Sun, 19 Jul 2026 14:48:46 +0800 https://example.org/environmentconfig/mamba-ssm%E7%8E%AF%E5%A2%83%E9%85%8D%E7%BD%AE/ Sun, 19 Jul 2026 14:48:46 +0800 https://example.org/environmentconfig/mamba-ssm%E7%8E%AF%E5%A2%83%E9%85%8D%E7%BD%AE/ <h2 id="前言">前言</h2> <p>Mamba-SSM还需要安装Causal_conv</p> <p>只能通过在GitHub上下载whl文件来进行安装。</p> <h2 id="实际配置">实际配置</h2> <p>即便下载了符合cuda和torch版本的mamba-ssm，依旧是失败。</p> <p><strong>唯一成功的方案是如下配置：</strong></p> <div class="code-block"> <button class="code-copy" type="button" hidden aria-label="Copy code to clipboard"> <span class="code-copy-label" aria-hidden="true">Copy</span> </button> <pre tabindex="0"><code>9527MY on Apr 3 i solve this problem.我解决了这个问题。 at cuda 12.4在cuda 12.4 pytorch 2.4 python 3.11 iintalled that我安装了那个 mamba_ssm-2.2.2+cu122torch2.4cxx11abiFALSE-cp311-cp311-linux_x86_64.whl causal_conv1d-1.4.0+cu122torch2.4cxx11abiFALSE-cp311-cp311-linux_x86_64.whl you can find at 你可以在 https://github.com/Dao-AlLab/causal-conv1d/releases/tag/v1.4.0https://github.com/state-spaces/mamba/releases/tag/v2.2.2L1</code></pre> </div> https://example.org/posts/flash-attention2/ Sat, 18 Jul 2026 21:52:15 +0800 https://example.org/posts/flash-attention2/ <h2 id="前言">前言</h2> <p>直接使用 <code>pip install flash-attn</code> 经常会遇到各种各样的报错，不推荐把报错信息交给AI解决，因为它们往往会乱答给出不管用的 fix。我建议直接前往官方仓库下载适配的 wheel 文件并用其进行安装。访问官方仓库的 release 页面：<a href="https://link.zhihu.com/?target=https%3A//github.com/Dao-AILab/flash-attention/releases">https://github.com/Dao-AILab/flash-attention/releases</a> ，可以看到许多批不同版本的 <a href="https://zhida.zhihu.com/search?content_id=268997872&amp;content_type=Article&amp;match_order=2&amp;q=flash-attn&amp;zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3ODQ1NTQ1MjgsInEiOiJmbGFzaC1hdHRuIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MjY4OTk3ODcyLCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.jNTb2B4uAgwt8X24HEO0T1fvfTxV3-rPaNbu4sx5dwg&amp;zhida_source=entity">flash-attn</a> ，再点开 Assets 就能看见该版本具体兼容的各个文件，我们需要选择一个适配的。</p> <p>一个典型的 flash-attn <code>.whl</code> 的文件名如 <code>flash_attn-2.7.3+cu11torch2.6cxx11abiFALSE-cp312-cp312-linux_x86_64.whl</code> 记录了该 <code>.whl</code> 文件对应的安装环境， “flash_attn-<strong>2.7.3</strong>+cu<strong>11</strong>torch<strong>2.6</strong>cxx11abi<strong>FALSE</strong>-cp<strong>312</strong>-cp<strong>312</strong>” 其中 <strong>标黑下划线</strong> 的各部分是我们要着重查明的。</p> <h2 id="确定配置">确定配置</h2> <h3 id="1-python-版本">1. Python 版本</h3> <p>在终端进入环境输入指令：</p> <div class="code-block"> <button class="code-copy" type="button" hidden aria-label="Copy code to clipboard"> <span class="code-copy-label" aria-hidden="true">Copy</span> </button> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;-webkit-text-size-adjust:none;"><code class="language-text" data-lang="text"><span style="display:flex;"><span>python --version</span></span></code></pre></div> </div><p>返回如：</p> <div class="code-block"> <button class="code-copy" type="button" hidden aria-label="Copy code to clipboard"> <span class="code-copy-label" aria-hidden="true">Copy</span> </button> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;-webkit-text-size-adjust:none;"><code class="language-text" data-lang="text"><span style="display:flex;"><span>Python 3.12.9</span></span></code></pre></div> </div><h3 id="2-cuda-版本">2. CUDA 版本</h3> <p>输入指令：</p> <div class="code-block"> <button class="code-copy" type="button" hidden aria-label="Copy code to clipboard"> <span class="code-copy-label" aria-hidden="true">Copy</span> </button> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;-webkit-text-size-adjust:none;"><code class="language-text" data-lang="text"><span style="display:flex;"><span>nvcc -V</span></span></code></pre></div> </div><p>返回如：</p> <div class="code-block"> <button class="code-copy" type="button" hidden aria-label="Copy code to clipboard"> <span class="code-copy-label" aria-hidden="true">Copy</span> </button> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;-webkit-text-size-adjust:none;"><code class="language-text" data-lang="text"><span style="display:flex;"><span>nvcc: NVIDIA (R) Cuda compiler driver </span></span><span style="display:flex;"><span>Copyright (c) 2005-2022 NVIDIA Corporation </span></span><span style="display:flex;"><span>Built on Wed_Sep_21_10:33:58_PDT_2022 </span></span><span style="display:flex;"><span>Cuda compilation tools, release 11.8, V11.8.89 </span></span><span style="display:flex;"><span>Build cuda_11.8.r11.8/compiler.31833905_0</span></span></code></pre></div> </div><p>这里查询的是<strong>系统本地</strong>安装的 <strong>NVIDIA <a href="https://zhida.zhihu.com/search?content_id=268997872&amp;content_type=Article&amp;match_order=1&amp;q=CUDA+Toolkit&amp;zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3ODQ1NTQ1MjgsInEiOiJDVURBIFRvb2xraXQiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyNjg5OTc4NzIsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.QuaBnGH2aw1M07mVeb0-MgfmcyNAr0Og1LI2RTCTaKQ&amp;zhida_source=entity">CUDA Toolkit</a></strong> 版本。可以看出此处 CUDA 版本为 11.8 ，对应 cu<strong>11</strong> ；如果 CUDA 版本为 12.x 则对应 cu<strong>12</strong> 。</p> https://example.org/search/ Mon, 01 Jan 0001 00:00:00 +0000 https://example.org/search/