<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>With-Qwen3-8-Flash on Jaehun's Blog</title><link>https://jaehun.me/categories/with-qwen3-8-flash/</link><description>Recent content in With-Qwen3-8-Flash on Jaehun's Blog</description><generator>Hugo</generator><language>ko-kr</language><lastBuildDate>Tue, 08 Sep 2026 13:42:29 +0000</lastBuildDate><atom:link href="https://jaehun.me/categories/with-qwen3-8-flash/index.xml" rel="self" type="application/rss+xml"/><item><title>Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning</title><link>https://jaehun.me/posts/random-attention-rethinking-kv-cache-eviction-for-efficient-reasoning/</link><pubDate>Sun, 06 Sep 2026 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/random-attention-rethinking-kv-cache-eviction-for-efficient-reasoning/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2609.03430v1"target="_blank"&#10; class="inline-flex items-center gap-1"&#10; &gt;논문 링크&lt;svg class="h-3 w-3 flex-shrink-0" id="external-link" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"&gt;&lt;path fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M15 3h6v6m-11 5L21 3m-3 10v6a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2V8a2 2 0 0 1 2-2h6"/&gt;&lt;/svg&gt;&#10; &lt;/a&gt;&lt;/p&gt;&#10;&lt;h2 id="점수는-낭비였다-kv-캐시-eviction에서-무작위-추출이-최강자와-동률이-되는-역설--random-attention-심층-리뷰"&gt;점수는 낭비였다: KV 캐시 eviction에서 &amp;lsquo;무작위 추출&amp;rsquo;이 최강자와 동률이 되는 역설 — Random Attention 심층 리뷰&lt;a href="#%ec%a0%90%ec%88%98%eb%8a%94-%eb%82%ad%eb%b9%84%ec%98%80%eb%8b%a4-kv-%ec%ba%90%ec%8b%9c-eviction%ec%97%90%ec%84%9c-%eb%ac%b4%ec%9e%91%ec%9c%84-%ec%b6%94%ec%b6%9c%ec%9d%b4-%ec%b5%9c%ea%b0%95%ec%9e%90%ec%99%80-%eb%8f%99%eb%a5%a0%ec%9d%b4-%eb%90%98%eb%8a%94-%ec%97%ad%ec%84%a4--random-attention-%ec%8b%ac%ec%b8%b5-%eb%a6%ac%eb%b7%b0" class="heading-anchor" aria-label="이 섹션에 대한 링크"&gt;&lt;svg class="h-4 w-4" aria-hidden="true" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"&gt;&lt;g fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2"&gt;&lt;path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/&gt;&lt;path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/&gt;&lt;/g&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h2&gt;&#10; &lt;blockquote&#10; class="border-primary bg-muted/30 hover:bg-muted/40 my-6 rounded-r-lg border-l-4 py-4 pr-4 pl-6 italic transition-all duration-200 ease-out hover:translate-x-1"&#10; &gt;&#10; &lt;div class="prose prose-sm max-w-none"&gt;&#10; &lt;p&gt;&lt;strong&gt;TL;DR&lt;/strong&gt; — 추론 특화 KV 캐시 eviction 연구는 전부 &amp;ldquo;어떤 토큰이 나중에 중요한가&amp;quot;를 &lt;strong&gt;점수화&lt;/strong&gt;해서 top-K 를 고르는 문제였다. Random Attention 은 프롬프트만 +∞로 강제 보존하고 나머지는 &lt;strong&gt;점수 없이 헤드별로 균등 무작위&lt;/strong&gt; eviction 하는데, 4개 모델·6개 추론 태스크에서 최강 베이스라인(TriAttention)과 동급 정확도(60개 비교 셀 중 31개에서 통계적 우위, 1개에서만 유의 열세)를 보이면서 vLLM 서빙에서는 스코어링 패스가 없어 &lt;strong&gt;32–43% 높은 처리량&lt;/strong&gt;을 낸다 (근거: §1, Fig. 1, Tab. 1, Tab. 4).&lt;/p&gt;</description></item></channel></rss>