<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
    <channel>
        <title>Linear Attention - 标签 - vllbc02&#39;s blogs</title>
        <link>https://blog.vllbc.top/tags/linear-attention/</link>
        <description>Linear Attention - 标签 - vllbc02&#39;s blogs</description>
        <generator>Hugo -- gohugo.io</generator><language>zh-CN</language><copyright>This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.</copyright><lastBuildDate>Wed, 29 Jul 2026 00:00:00 &#43;0000</lastBuildDate><atom:link href="https://blog.vllbc.top/tags/linear-attention/" rel="self" type="application/rss+xml" /><item>
    <title>从 GPT-2 到 Kimi K3：大模型架构如何演进</title>
    <link>https://blog.vllbc.top/%E4%BB%8Egpt-2%E5%88%B0kimi-k3/</link>
    <pubDate>Wed, 29 Jul 2026 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/%E4%BB%8Egpt-2%E5%88%B0kimi-k3/</guid>
    <description><![CDATA[<blockquote>
<p>[!info] 来源 整理自 ali（<a
href="https://x.com/waterloo_intern"><span class="citation"
data-cites="waterloo_intern">@waterloo_intern</span></a>）于 2026-07-27
发布的 X Article：<a
href="https://x.com/waterloo_intern/status/2081762065392541951">22580:
From GPT2 to Kimi3,
Explained</a>。本文不是逐句翻译，而是对原文论证主线的中文重组。</p>]]></description>
</item>
<item>
    <title>Gated DeltaNet</title>
    <link>https://blog.vllbc.top/gated-deltanet/</link>
    <pubDate>Sat, 06 Jun 2026 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/gated-deltanet/</guid>
    <description><![CDATA[<h1 id="gated-deltanet">Gated DeltaNet</h1>
<blockquote>
<p>来自 Sebastian Raschka 的 <em>The Big LLM Architecture
Comparison</em>。一个<strong>线性 attention 变体</strong>，是 Qwen3-Next
的核心组件。</p>
</blockquote>
<h2 id="一句话">一句话</h2>
<p>Gated DeltaNet 是 <strong>DeltaNet + Mamba 风格 gating</strong>
的组合：</p>
<ul>
<li>DeltaNet：用 q / k / v / α / β 维护一个<strong>fast-weight
memory</strong>，按 delta rule 更新（线性 O(n) 缩放）</li>
<li>Mamba gating：额外用门控机制控制 memory 写入/读取</li>
</ul>
<p>→ 本质是”线性 attention 的现代变种”，常被拿来做 3:1 混合中的”线性那 3
层”。</p>]]></description>
</item>
<item>
    <title>KDA（Kimi Delta Attention）</title>
    <link>https://blog.vllbc.top/kdakimi-delta-attention/</link>
    <pubDate>Sat, 06 Jun 2026 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/kdakimi-delta-attention/</guid>
    <description><![CDATA[<h1 id="kdakimi-delta-attention">KDA（Kimi Delta Attention）</h1>
<blockquote>
<p>来自 Sebastian Raschka 的 <em>The Big LLM Architecture
Comparison</em>。Kimi Linear 对 [[Gated DeltaNet]] 的精细改进。</p>
</blockquote>
<h2 id="一句话">一句话</h2>
<p>KDA = 把 Gated DeltaNet 中的<strong>标量 gate</strong> 改成
<strong>channel-wise gate</strong>（每个 feature 通道独立）：</p>]]></description>
</item>
<item>
    <title>Kimi 系列（K2 / K2 Thinking / Kimi Linear）</title>
    <link>https://blog.vllbc.top/kimi-%E7%B3%BB%E5%88%97/</link>
    <pubDate>Sat, 06 Jun 2026 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/kimi-%E7%B3%BB%E5%88%97/</guid>
    <description><![CDATA[<h1 id="kimi-系列k2-k2-thinking-kimi-linear">Kimi 系列（K2 / K2 Thinking
/ Kimi Linear）</h1>
<blockquote>
<p>原文第 8、14 节。Moonshot AI 的旗舰开源系列。K2 是 1T
参数的”超大杯”，Kimi Linear 是 2025 年 10 月推出的线性注意力版本。</p>]]></description>
</item>
<item>
    <title>Nemotron 3（Nano / Super）</title>
    <link>https://blog.vllbc.top/nemotron-3/</link>
    <pubDate>Sat, 06 Jun 2026 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/nemotron-3/</guid>
    <description><![CDATA[<h1 id="nemotron-3nano-super">Nemotron 3（Nano / Super）</h1>
<blockquote>
<p>原文第 18 节。NVIDIA 在 2025-12 发布的 Nemotron 3 系列，开源权重 +
技术报告 + 数据集 + 训练代码（与 Olmo 3 透明度相当）。</p>
</blockquote>
<h2 id="nemotron-3-nano30b-a3b">18.1 Nemotron 3 Nano（30B-A3B）</h2>
<h3 id="架构总览">架构总览</h3>
<p>Nemotron 3 Nano 是 <strong>Mamba-Transformer 混合 +
MoE</strong>：</p>]]></description>
</item>
</channel>
</rss>
