<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Machine Learning on Jaehun's Blog</title><link>https://jaehun.me/tags/machine-learning/</link><description>Recent content in Machine Learning 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/tags/machine-learning/index.xml" rel="self" type="application/rss+xml"/><item><title>[논문리뷰]: NVIDIA Nemotron 3: Efficient and Open Intelligence</title><link>https://jaehun.me/posts/%EB%85%BC%EB%AC%B8%EB%A6%AC%EB%B7%B0-nvidia-nemotron-3-efficient-and-open-intelligence/</link><pubDate>Tue, 16 Dec 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/%EB%85%BC%EB%AC%B8%EB%A6%AC%EB%B7%B0-nvidia-nemotron-3-efficient-and-open-intelligence/</guid><description>&lt;p&gt;&lt;a&#10; href="https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-White-Paper.pdf"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="nvidia-nemotron-3-하이브리드-mambatransformer-moe로-정확도처리량-프런티어를-당기다"&gt;NVIDIA Nemotron 3: 하이브리드 Mamba–Transformer MoE로 “정확도/처리량” 프런티어를 당기다&lt;a href="#nvidia-nemotron-3-%ed%95%98%ec%9d%b4%eb%b8%8c%eb%a6%ac%eb%93%9c-mambatransformer-moe%eb%a1%9c-%ec%a0%95%ed%99%95%eb%8f%84%ec%b2%98%eb%a6%ac%eb%9f%89-%ed%94%84%eb%9f%b0%ed%8b%b0%ec%96%b4%eb%a5%bc-%eb%8b%b9%ea%b8%b0%eb%8b%a4" 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;&lt;p&gt;&lt;strong&gt;Nemotron 3&lt;/strong&gt; 는 MoE 하이브리드 Mamba–Transformer, &lt;strong&gt;LatentMoE&lt;/strong&gt; , &lt;strong&gt;MTP&lt;/strong&gt; , &lt;strong&gt;NVFP4&lt;/strong&gt; 학습, &lt;strong&gt;멀티-환경 RL&lt;/strong&gt; 을 결합해 “정확도 대비 추론 처리량(accuracy-to-inference-throughput)”을 끌어올리고, &lt;strong&gt;최대 1M tokens 컨텍스트&lt;/strong&gt; 와 &lt;strong&gt;상대 처리량 3.3×&lt;/strong&gt; 를 핵심 메시지로 제시한다 (근거: §Intro/§2.2/§2.3/§2.4/§2.5/§2.6/Fig.2).&lt;/p&gt;</description></item><item><title>[논문리뷰]: Radial Attention: O(n log n) Sparse Attention with Energy Decay for Long Video Generation</title><link>https://jaehun.me/posts/%EB%85%BC%EB%AC%B8%EB%A6%AC%EB%B7%B0-radial-attention-on-log-n-sparse-attention-with-energy-decay-for-long-video-generation/</link><pubDate>Mon, 15 Dec 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/%EB%85%BC%EB%AC%B8%EB%A6%AC%EB%B7%B0-radial-attention-on-log-n-sparse-attention-with-energy-decay-for-long-video-generation/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2506.19852v1"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="radial-attention-on-log-n-sparse-attention-with-energy-decay로-긴-비디오를-싸게-뽑기"&gt;Radial Attention: O(n log n) Sparse Attention with Energy Decay로 “긴 비디오”를 싸게 뽑기&lt;a href="#radial-attention-on-log-n-sparse-attention-with-energy-decay%eb%a1%9c-%ea%b8%b4-%eb%b9%84%eb%94%94%ec%98%a4%eb%a5%bc-%ec%8b%b8%ea%b2%8c-%eb%bd%91%ea%b8%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;&lt;h2 id="한-줄-요약-tldr"&gt;한 줄 요약 (TL;DR)&lt;a href="#%ed%95%9c-%ec%a4%84-%ec%9a%94%ec%95%bd-tldr" 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;&lt;p&gt;사후(softmax 이후) attention 에너지가 거리(시간/공간)와 함께 &lt;strong&gt;지수적으로 감쇠&lt;/strong&gt; 한다는 관찰을 바탕으로, 계산 밀도도 같은 방식으로 감쇠시키는 &lt;strong&gt;정적(static) 마스크&lt;/strong&gt;를 설계해 긴 비디오 생성의 훈련/추론 비용을 크게 줄인다. (근거: §4.1–§4.2, Eq.3–Eq.4)&lt;/p&gt;</description></item><item><title>[논문리뷰] Pretraining Large Language Models with NVFP4</title><link>https://jaehun.me/posts/%EB%85%BC%EB%AC%B8%EB%A6%AC%EB%B7%B0-pretraining-large-language-models-with-nvfp4/</link><pubDate>Thu, 09 Oct 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/%EB%85%BC%EB%AC%B8%EB%A6%AC%EB%B7%B0-pretraining-large-language-models-with-nvfp4/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2509.25149v1"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;h1 id="nvfp4로-4-bit-프리트레이닝을-실전으로-12b를-10t-토큰까지-fp8과-사실상-동급"&gt;NVFP4로 4-bit 프리트레이닝을 실전으로: 12B를 10T 토큰까지, FP8과 사실상 동급&lt;a href="#nvfp4%eb%a1%9c-4-bit-%ed%94%84%eb%a6%ac%ed%8a%b8%eb%a0%88%ec%9d%b4%eb%8b%9d%ec%9d%84-%ec%8b%a4%ec%a0%84%ec%9c%bc%eb%a1%9c-12b%eb%a5%bc-10t-%ed%86%a0%ed%81%b0%ea%b9%8c%ec%a7%80-fp8%ea%b3%bc-%ec%82%ac%ec%8b%a4%ec%83%81-%eb%8f%99%ea%b8%89" 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;/h1&gt;&lt;h2 id="tldr"&gt;TL;DR&lt;a href="#tldr" 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;&lt;p&gt;12B 하이브리드 Mamba-Transformer를 &lt;strong&gt;10T tokens&lt;/strong&gt; 에서 &lt;strong&gt;NVFP4(4-bit)&lt;/strong&gt; 로 프리트레이닝하면 &lt;strong&gt;안정 구간 손실차 &amp;lt;1%&lt;/strong&gt; , 말기 &lt;strong&gt;~1.5%&lt;/strong&gt; 로 &lt;strong&gt;FP8을 근접 추종&lt;/strong&gt; 하고, 다운스트림 성능도 대부분 동급(수학·다국어 일부 +0.9~+3.7pp)이다. &lt;strong&gt;MXFP4 대비 동일 손실에 토큰 +36%&lt;/strong&gt; (1.36T vs 1.0T)가 필요해 &lt;strong&gt;NVFP4의 토큰 효율 우위&lt;/strong&gt; 가 확인된다. (근거: Fig.2, Tab.2, Fig.6)&lt;/p&gt;</description></item><item><title>[논문리뷰] Inference-Time Hyper-Scaling with KV Cache Compression</title><link>https://jaehun.me/posts/%EB%85%BC%EB%AC%B8%EB%A6%AC%EB%B7%B0-inference-time-hyper-scaling-with-kv-cache-compression/</link><pubDate>Tue, 29 Jul 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/%EB%85%BC%EB%AC%B8%EB%A6%AC%EB%B7%B0-inference-time-hyper-scaling-with-kv-cache-compression/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2506.05345v1"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="dynamic-memory-sparsificationdms-kv-캐시-8-압축으로-llm-하이퍼-스케일링을-현실로"&gt;Dynamic Memory Sparsification(DMS): KV 캐시 8× 압축으로 LLM 하이퍼-스케일링을 현실로&lt;a href="#dynamic-memory-sparsificationdms-kv-%ec%ba%90%ec%8b%9c-8-%ec%95%95%ec%b6%95%ec%9c%bc%eb%a1%9c-llm-%ed%95%98%ec%9d%b4%ed%8d%bc-%ec%8a%a4%ec%bc%80%ec%9d%bc%eb%a7%81%ec%9d%84-%ed%98%84%ec%8b%a4%eb%a1%9c" 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;&lt;h2 id="한-줄-요약-tldr"&gt;한 줄 요약 (TL;DR)&lt;a href="#%ed%95%9c-%ec%a4%84-%ec%9a%94%ec%95%bd-tldr" 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;&lt;p&gt;&lt;code&gt;1 K&lt;/code&gt; 스텝만의 경량 재적합과 &lt;strong&gt;지연 퇴출 전략&lt;/strong&gt;을 결합한 &lt;strong&gt;DMS&lt;/strong&gt;는 KV 캐시를 최대 &lt;strong&gt;8×&lt;/strong&gt; 압축하면서도 Qwen-R1 32B 기준 AIME 24 &lt;strong&gt;+9.1 pt&lt;/strong&gt; 등 성능을 오히려 끌어올렸다. 결과적으로 동일 연산·메모리 예산에서 &lt;strong&gt;더 길고·더 많은&lt;/strong&gt; 토큰을 실시간으로 생성할 수 있다.&lt;/p&gt;</description></item><item><title>Accelerating LLM Inference Throughput via Asynchronous KV Cache Prefetching</title><link>https://jaehun.me/posts/accelerating-llm-inference-throughput-via-asynchronous-kv-cache-prefetching/</link><pubDate>Thu, 19 Jun 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/accelerating-llm-inference-throughput-via-asynchronous-kv-cache-prefetching/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2504.06319v1"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h2 id="llm-추론-가속화를-위한-비동기식-kv-캐시-프리페칭-논문-상세-분석"&gt;LLM 추론 가속화를 위한 비동기식 KV 캐시 프리페칭 논문 상세 분석&lt;a href="#llm-%ec%b6%94%eb%a1%a0-%ea%b0%80%ec%86%8d%ed%99%94%eb%a5%bc-%ec%9c%84%ed%95%9c-%eb%b9%84%eb%8f%99%ea%b8%b0%ec%8b%9d-kv-%ec%ba%90%ec%8b%9c-%ed%94%84%eb%a6%ac%ed%8e%98%ec%b9%ad-%eb%85%bc%eb%ac%b8-%ec%83%81%ec%84%b8-%eb%b6%84%ec%84%9d" 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;&lt;p&gt;제공된 논문 &amp;ldquo;Accelerating LLM Inference Throughput via Asynchronous KV Cache Prefetching&amp;quot;은 대규모 언어 모델(LLM) 추론 시 발생하는 메모리 병목 현상을 해결하기 위한 독창적인 방법을 제시합니다. 이 논문은 GPU의 L2 캐시를 활용한 비동기식 프리페칭(prefetching) 기법을 통해 추론 처리량을 획기적으로 개선하는 데 중점을 둡니다.&lt;/p&gt;</description></item><item><title>Hogwild! Inference: Parallel LLM Generation via Concurrent Attention</title><link>https://jaehun.me/posts/hogwild-inference-parallel-llm-generation-via-concurrent-attention/</link><pubDate>Thu, 19 Jun 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/hogwild-inference-parallel-llm-generation-via-concurrent-attention/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2504.06261v3"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h2 id="hogwild-inference-논문-심층-분석"&gt;Hogwild! Inference 논문 심층 분석&lt;a href="#hogwild-inference-%eb%85%bc%eb%ac%b8-%ec%8b%ac%ec%b8%b5-%eb%b6%84%ec%84%9d" 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;&lt;p&gt;제공해주신 &amp;ldquo;Hogwild! Inference: Parallel LLM Generation via Concurrent Attention&amp;rdquo; 논문을 자세히 분석하여 강점과 독창성, 핵심 알고리즘, 그리고 한계점을 설명해 드리겠습니다.&lt;/p&gt;</description></item><item><title>MMInference: Accelerating Pre-filling for Long-Context Visual Language Models via Modality-Aware Permutation Sparse Attention</title><link>https://jaehun.me/posts/mminference-accelerating-pre-filling-for-long-context-visual-language-models-via-modality-aware-permutation-sparse-attention/</link><pubDate>Thu, 19 Jun 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/mminference-accelerating-pre-filling-for-long-context-visual-language-models-via-modality-aware-permutation-sparse-attention/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2504.16083v2"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h2 id="mminference-논문-리뷰-vlm의-긴-컨텍스트-추론-순열로-속도의-벽을-넘다"&gt;MMInference 논문 리뷰: VLM의 긴 컨텍스트 추론, &amp;lsquo;순열&amp;rsquo;로 속도의 벽을 넘다&lt;a href="#mminference-%eb%85%bc%eb%ac%b8-%eb%a6%ac%eb%b7%b0-vlm%ec%9d%98-%ea%b8%b4-%ec%bb%a8%ed%85%8d%ec%8a%a4%ed%8a%b8-%ec%b6%94%eb%a1%a0-%ec%88%9c%ec%97%b4%eb%a1%9c-%ec%86%8d%eb%8f%84%ec%9d%98-%eb%b2%bd%ec%9d%84-%eb%84%98%eb%8b%a4" 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;&lt;p&gt;최근 Vision Language Model(VLM)은 이미지와 텍스트를 넘어 긴 비디오까지 이해하는 능력으로 무한한 가능성을 보여주고 있습니다. [cite_start]하지만 수백만 개의 토큰으로 이루어진 긴 비디오를 입력받을 때, 모델이 본격적인 답변 생성을 시작하기 전 입력 전체를 처리하는 &amp;lsquo;Pre-filling&amp;rsquo; 단계에서 엄청난 지연이 발생합니다. [cite: 2] [cite_start]이는 어텐션 메커니즘의 연산량이 입력 길이의 제곱에 비례하여 증가하기 때문인데, 현실적인 서비스 적용에 큰 걸림돌이 되어 왔습니다. [cite: 2, 19]&lt;/p&gt;</description></item><item><title>Slim attention: cut your context memory in half without loss– K-cache is all you need for MHA</title><link>https://jaehun.me/posts/slim-attention-cut-your-context-memory-in-half-without-loss-k-cache-is-all-you-need-for-mha/</link><pubDate>Mon, 16 Jun 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/slim-attention-cut-your-context-memory-in-half-without-loss-k-cache-is-all-you-need-for-mha/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2503.05840v2"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h2 id="slim-attention-논문-심층-분석"&gt;Slim Attention 논문 심층 분석&lt;a href="#slim-attention-%eb%85%bc%eb%ac%b8-%ec%8b%ac%ec%b8%b5-%eb%b6%84%ec%84%9d" 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;&lt;p&gt;제공해주신 &amp;ldquo;Slim attention: cut your context memory in half without loss&amp;rdquo; 논문을 정독하고 요청하신 내용에 따라 상세하게 분석해 드리겠습니다.&lt;/p&gt;</description></item><item><title>TransMLA: Multi-Head Latent Attention Is All You Need</title><link>https://jaehun.me/posts/transmla-multi-head-latent-attention-is-all-you-need/</link><pubDate>Mon, 16 Jun 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/transmla-multi-head-latent-attention-is-all-you-need/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2502.07864v5"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;p&gt;네, TransMLA 논문(arXiv:2502.07864v5)을 심도 있게 분석하여 강점과 독창성, 핵심 알고리즘, 그리고 한계점을 상세히 설명해 드리겠습니다.&lt;/p&gt;</description></item><item><title>Accelerating MoE Model Inference with Expert Sharding</title><link>https://jaehun.me/posts/accelerating-moe-model-inference-with-expert-sharding/</link><pubDate>Thu, 05 Jun 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/accelerating-moe-model-inference-with-expert-sharding/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2503.08467v1"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;p&gt;제공해주신 논문 &amp;ldquo;Accelerating MoE Model Inference with Expert Sharding&amp;rdquo; (MOESHARD)에 대한 자세한 분석은 다음과 같습니다.&lt;/p&gt;</description></item><item><title>An Empirical Study of Qwen3 Quantization</title><link>https://jaehun.me/posts/an-empirical-study-of-qwen3-quantization/</link><pubDate>Mon, 12 May 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/an-empirical-study-of-qwen3-quantization/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2505.02214v1"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h3 id="-결론-요약"&gt;✅ 결론 요약&lt;a href="#-%ea%b2%b0%eb%a1%a0-%ec%9a%94%ec%95%bd" 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;/h3&gt;&lt;p&gt;Qwen3의 정밀한 사전학습 덕분에 고성능을 보이나, 3bit 이하의 ultra-low-bit quantization에 매우 민감하여 기존 모델보다 성능 저하가 크다. 4bit 이상에서는 GPTQ와 AWQ가 최상의 trade-off를 제공하며, BiLLM은 1bit에서도 유의미한 결과를 보인다. 이는 Qwen3이 redundancy가 적은 모델 구조라는 점에서 비롯된 것이며, 향후 LLM quantization 연구가 이를 고려해 더욱 정교한 전략을 개발해야 함을 시사한다.&lt;/p&gt;</description></item><item><title>MELODI: Exploring Memory Compression for Long Contexts</title><link>https://jaehun.me/posts/melodi-exploring-memory-compression-for-long-contexts/</link><pubDate>Mon, 12 May 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/melodi-exploring-memory-compression-for-long-contexts/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2410.03156v1"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h3 id="-결론-요약"&gt;📌 결론 요약&lt;a href="#-%ea%b2%b0%eb%a1%a0-%ec%9a%94%ec%95%bd" 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;/h3&gt;&lt;p&gt;논문 &amp;ldquo;MELODI: Exploring Memory Compression for Long Contexts&amp;quot;는 Transformer의 긴 문맥 처리 문제를 해결하기 위해 &lt;strong&gt;계층적 메모리 압축(hierarchical memory compression)&lt;/strong&gt; 구조를 제안합니다.&#10;핵심은 **다층 반복 압축 기반의 단기 메모리(SM)**와 **단일층 추가 압축 기반의 장기 메모리(LM)**를 조합한 **“샌드위치 구조”**를 사용하여 긴 문서를 짧은 윈도우(예: 512 tokens)로 효율적으로 처리하는 것입니다.&#10;메모리 사용량은 기존 Dense Memory 방식인 Memorizing Transformer보다 &lt;strong&gt;최대 8배 절감&lt;/strong&gt;하면서도 성능(PPL 기준)은 오히려 &lt;strong&gt;향상&lt;/strong&gt;됩니다.&lt;/p&gt;</description></item><item><title>Comet: Fine-grained Computation-communication Overlapping for Mixture-of-Experts</title><link>https://jaehun.me/posts/comet-fine-grained-computation-communication-overlapping-for-mixture-of-experts/</link><pubDate>Mon, 14 Apr 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/comet-fine-grained-computation-communication-overlapping-for-mixture-of-experts/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2502.19811v3"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;p&gt;논문 **&amp;ldquo;Comet: Fine-grained Computation-communication Overlapping for Mixture-of-Experts&amp;rdquo;**는 MoE 모델의 핵심 병목인 GPU 간 통신 지연을 &lt;strong&gt;fine-grained 수준에서 컴퓨팅과 통신을 정교하게 겹치도록 설계&lt;/strong&gt;함으로써 실행 성능을 크게 개선한 ByteDance의 시스템 최적화 논문입니다.&lt;/p&gt;</description></item><item><title>MegaScale-Infer: Serving Mixture-of-Experts at Scale with Disaggregated Expert Parallelism</title><link>https://jaehun.me/posts/megascale-infer-serving-mixture-of-experts-at-scale-with-disaggregated-expert-parallelism/</link><pubDate>Mon, 14 Apr 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/megascale-infer-serving-mixture-of-experts-at-scale-with-disaggregated-expert-parallelism/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2504.02263v2"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h2 id="-결론-요약-핵심-기여-및-성능"&gt;📌 결론 요약 (핵심 기여 및 성능)&lt;a href="#-%ea%b2%b0%eb%a1%a0-%ec%9a%94%ec%95%bd-%ed%95%b5%ec%8b%ac-%ea%b8%b0%ec%97%ac-%eb%b0%8f-%ec%84%b1%eb%8a%a5" 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;&lt;p&gt;&lt;strong&gt;MegaScale-Infer&lt;/strong&gt;는 대규모 Mixture-of-Experts (MoE) 모델 서빙을 위한 효율적 시스템으로, &lt;strong&gt;Attention과 FFN 모듈을 분리(disaggregate)&lt;/strong&gt; 하여 GPU 활용률을 극대화하고 &lt;strong&gt;최대 1.9×의 GPU throughput 개선&lt;/strong&gt; 및 &lt;strong&gt;1.86× 비용 대비 성능 향상&lt;/strong&gt;을 달성합니다.&lt;/p&gt;</description></item><item><title>SwitchHead: Accelerating Transformers with Mixture-of-Experts Attention</title><link>https://jaehun.me/posts/switchhead-accelerating-transformers-with-mixture-of-experts-attention/</link><pubDate>Mon, 14 Apr 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/switchhead-accelerating-transformers-with-mixture-of-experts-attention/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2312.07987v3"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;p&gt;논문 「SwitchHead: Accelerating Transformers with Mixture-of-Experts Attention」을 매우 자세하게 읽고 분석한 내용을 바탕으로, 논문의 강점과 독창적인 지점, 핵심 알고리즘의 상세한 설명과 함께 예시 입력을 이용한 동작 과정을 소개하고, 마지막으로 한계점을 명확하게 설명하겠습니다.&lt;/p&gt;</description></item><item><title>Duplex: A Device for Large Language Models with Mixture of Experts, Grouped Query Attention, and Continuous Batching</title><link>https://jaehun.me/posts/duplex-a-device-for-large-language-models-with-mixture-of-experts-grouped-query-attention-and-continuous-batching/</link><pubDate>Sun, 13 Apr 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/duplex-a-device-for-large-language-models-with-mixture-of-experts-grouped-query-attention-and-continuous-batching/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2409.01141v1"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h3 id="-결론-요약"&gt;📌 결론 요약&lt;a href="#-%ea%b2%b0%eb%a1%a0-%ec%9a%94%ec%95%bd" 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;/h3&gt;&lt;p&gt;논문 *&amp;ldquo;Duplex: A Device for Large Language Models with Mixture of Experts, Grouped Query Attention, and Continuous Batching&amp;rdquo;*은 저연산량(Op/B)이 지배적인 MoE 및 GQA 기반 LLM 추론을 위한 하드웨어 아키텍처 &lt;strong&gt;Duplex&lt;/strong&gt;를 제안하며, GPU 단독 대비 &lt;strong&gt;최대 2.67×의 추론 속도&lt;/strong&gt;와 &lt;strong&gt;42.03%의 에너지 절감&lt;/strong&gt; 효과를 보여줍니다. 핵심은 **xPU (GPU 수준 고성능 연산기)**와 **Logic-PIM (로직 다이에 탑재된 저 Op/B 특화 연산기)**를 &lt;strong&gt;동시에 활용&lt;/strong&gt;하여 MoE와 Attention Layer를 공동 처리(co-processing)하는 방식입니다.&lt;/p&gt;</description></item><item><title>Mirage: A Multi-Level Superoptimizer for Tensor Programs</title><link>https://jaehun.me/posts/mirage-a-multi-level-superoptimizer-for-tensor-programs/</link><pubDate>Sun, 13 Apr 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/mirage-a-multi-level-superoptimizer-for-tensor-programs/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2405.05751v2"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;p&gt;논문 『Mirage: A Multi-Level Superoptimizer for Tensor Programs』를 자세히 분석한 내용을 정리하여 설명합니다.&lt;/p&gt;</description></item><item><title>MoEUT: Mixture-of-Experts Universal Transformers</title><link>https://jaehun.me/posts/moeut-mixture-of-experts-universal-transformers/</link><pubDate>Sun, 13 Apr 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/moeut-mixture-of-experts-universal-transformers/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2405.16039v2"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h3 id="-결론-요약-핵심-강점-독창성-핵심-메커니즘"&gt;✅ 결론 요약 (핵심 강점, 독창성, 핵심 메커니즘)&lt;a href="#-%ea%b2%b0%eb%a1%a0-%ec%9a%94%ec%95%bd-%ed%95%b5%ec%8b%ac-%ea%b0%95%ec%a0%90-%eb%8f%85%ec%b0%bd%ec%84%b1-%ed%95%b5%ec%8b%ac-%eb%a9%94%ec%bb%a4%eb%8b%88%ec%a6%98" 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;/h3&gt;&lt;p&gt;&lt;strong&gt;MoEUT&lt;/strong&gt;는 Universal Transformer(UT)의 구조적 강점(레이어 파라미터 공유 기반의 recurrence)을 유지하면서도 기존 한계였던 &lt;strong&gt;parameter-compute ratio 문제&lt;/strong&gt;를 해결한 구조다. 핵심은 다음 세 가지다:&lt;/p&gt;</description></item><item><title>FLEX ATTENTION: A PROGRAMMING MODEL FOR GENERATING OPTIMIZED ATTENTION KERNELS</title><link>https://jaehun.me/posts/flex-attention-a-programming-model-for-generating-optimized-attention-kernels/</link><pubDate>Mon, 07 Apr 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/flex-attention-a-programming-model-for-generating-optimized-attention-kernels/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2412.05496v1"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h3 id="-논문-요약-및-분석-flexattention-a-programming-model-for-generating-optimized-attention-kernels"&gt;📌 &lt;strong&gt;논문 요약 및 분석 (FlexAttention: A Programming Model for Generating Optimized Attention Kernels)&lt;/strong&gt;&lt;a href="#-%eb%85%bc%eb%ac%b8-%ec%9a%94%ec%95%bd-%eb%b0%8f-%eb%b6%84%ec%84%9d-flexattention-a-programming-model-for-generating-optimized-attention-kernels" 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;/h3&gt;&lt;hr&gt;&#10;&lt;h2 id="-논문의-강점-및-독창적인-지점"&gt;✅ &lt;strong&gt;논문의 강점 및 독창적인 지점&lt;/strong&gt;&lt;a href="#-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90-%eb%b0%8f-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90" 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;&lt;ol&gt;&#10;&lt;li&gt;&#10;&lt;p&gt;&lt;strong&gt;유연성 (Flexibility)&lt;/strong&gt;&lt;/p&gt;</description></item><item><title>LeanAttention: Hardware-Aware Scalable Attention Mechanism for the Decode-Phase of Transformers</title><link>https://jaehun.me/posts/leanattention-hardware-aware-scalable-attention-mechanism-for-the-decode-phase-of-transformers/</link><pubDate>Mon, 07 Apr 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/leanattention-hardware-aware-scalable-attention-mechanism-for-the-decode-phase-of-transformers/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2405.10480v2"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;p&gt;논문 『LeanAttention: Hardware-Aware Scalable Attention Mechanism for the Decode-Phase of Transformers』에 대한 상세한 분석을 다음과 같이 제시합니다.&lt;/p&gt;</description></item><item><title>SparseTransX: Efficient Training of Translation-Based Knowledge Graph Embeddings Using Sparse Matrix Operations</title><link>https://jaehun.me/posts/sparsetransx-efficient-training-of-translation-based-knowledge-graph-embeddings-using-sparse-matrix-operations/</link><pubDate>Wed, 02 Apr 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/sparsetransx-efficient-training-of-translation-based-knowledge-graph-embeddings-using-sparse-matrix-operations/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2502.16949"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;p&gt;논문 『SparseTransX: Efficient Training of Translation-Based Knowledge Graph Embeddings Using Sparse Matrix Operations』의 강점, 독창적인 지점, 핵심 알고리즘의 예시와 전체적인 과정, 그리고 한계점을 정리하여 전달드립니다.&lt;/p&gt;</description></item><item><title>NEO: Saving GPU Memory Crisis with CPU Offloading for Online LLM Inference</title><link>https://jaehun.me/posts/neo-saving-gpu-memory-crisis-with-cpu-offloading-for-online-llm-inference/</link><pubDate>Mon, 31 Mar 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/neo-saving-gpu-memory-crisis-with-cpu-offloading-for-online-llm-inference/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2411.01142"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h3 id="-논문-제목-neo-saving-gpu-memory-crisis-with-cpu-offloading-for-online-llm-inference"&gt;📌 &lt;strong&gt;논문 제목:&lt;/strong&gt; NEO: Saving GPU Memory Crisis with CPU Offloading for Online LLM Inference&lt;a href="#-%eb%85%bc%eb%ac%b8-%ec%a0%9c%eb%aa%a9-neo-saving-gpu-memory-crisis-with-cpu-offloading-for-online-llm-inference" 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;/h3&gt;&lt;h3 id="-저자-xuanlin-jiang-yang-zhou-shiyi-cao-ion-stoica-minlan-yu"&gt;📌 &lt;strong&gt;저자:&lt;/strong&gt; Xuanlin Jiang, Yang Zhou, Shiyi Cao, Ion Stoica, Minlan Yu&lt;a href="#-%ec%a0%80%ec%9e%90-xuanlin-jiang-yang-zhou-shiyi-cao-ion-stoica-minlan-yu" 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;/h3&gt;&lt;hr&gt;&#10;&lt;h2 id="1-결론-요약-강점--독창적인-지점"&gt;&lt;strong&gt;1. 결론 요약 (강점 &amp;amp; 독창적인 지점)&lt;/strong&gt;&lt;a href="#1-%ea%b2%b0%eb%a1%a0-%ec%9a%94%ec%95%bd-%ea%b0%95%ec%a0%90--%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90" 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;&lt;p&gt;NEO는 GPU 메모리의 제한으로 인해 발생하는 LLM 추론의 병목을 해결하기 위해 &lt;strong&gt;비대칭 GPU-CPU 파이프라이닝과 부하 인식 스케줄링을 적용한 새로운 시스템&lt;/strong&gt;입니다. 주요 강점은 다음과 같습니다.&lt;/p&gt;</description></item><item><title>PipeFill: Using GPUs During Bubbles in Pipeline-parallel LLM Training</title><link>https://jaehun.me/posts/pipefill-using-gpus-during-bubbles-in-pipeline-parallel-llm-training/</link><pubDate>Tue, 25 Mar 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/pipefill-using-gpus-during-bubbles-in-pipeline-parallel-llm-training/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2410.07192"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;p&gt;논문『PipeFill: Using GPUs During Bubbles in Pipeline-parallel LLM Training』의 핵심 내용을 상세히 분석하여, 논문의 강점, 독창적인 지점, 핵심 알고리즘의 전체적인 과정 및 한계점을 요약하였습니다.&lt;/p&gt;</description></item><item><title>SELF-DATA DISTILLATION FOR RECOVERING QUALITY IN PRUNED LARGE LANGUAGE MODELS</title><link>https://jaehun.me/posts/self-data-distillation-for-recovering-quality-in-pruned-large-language-models/</link><pubDate>Tue, 25 Mar 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/self-data-distillation-for-recovering-quality-in-pruned-large-language-models/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2410.09982"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;p&gt;논문의 핵심을 정리하여 결론부터 간략히 제시한 후, 구체적인 수치를 통해 강점 및 독창적인 지점을 설명하고, 논문에서 제안한 핵심 알고리즘을 예시와 함께 설명하며, 논문의 한계점을 논의하겠습니다.&lt;/p&gt;</description></item><item><title>On Distributed Larger-Than-Memory Subset Selection With Pairwise Submodular Functions</title><link>https://jaehun.me/posts/on-distributed-larger-than-memory-subset-selection-with-pairwise-submodular-functions/</link><pubDate>Mon, 24 Mar 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/on-distributed-larger-than-memory-subset-selection-with-pairwise-submodular-functions/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2402.16442"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h3 id="-결론-요약"&gt;✅ 결론 요약&lt;a href="#-%ea%b2%b0%eb%a1%a0-%ec%9a%94%ec%95%bd" 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;/h3&gt;&lt;p&gt;이 논문은 &amp;ldquo;메모리 용량을 초과하는 대규모 데이터셋에서 대표적인 subset을 효율적으로 선택&amp;quot;하는 문제를 다룬다. 기존 방법들은 중앙 서버가 전체 subset을 메모리에 올릴 수 있어야 한다는 제약이 있었지만, 이 논문은 &lt;strong&gt;중앙 서버 없이도 분산 환경에서 고품질 subset을 선택&lt;/strong&gt;할 수 있는 새로운 알고리즘 2가지를 제안한다:&lt;/p&gt;</description></item><item><title>SampleAttention: Near-Lossless Acceleration of Long Context LLM Inference with Adaptive Structured Sparse Attention</title><link>https://jaehun.me/posts/sampleattention-near-lossless-acceleration-of-long-context-llm-inference-with-adaptive-structured-sparse-attention/</link><pubDate>Mon, 24 Mar 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/sampleattention-near-lossless-acceleration-of-long-context-llm-inference-with-adaptive-structured-sparse-attention/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2406.15486"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h3 id="-결론-요약"&gt;✅ 결론 요약&lt;a href="#-%ea%b2%b0%eb%a1%a0-%ec%9a%94%ec%95%bd" 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;/h3&gt;&lt;p&gt;&lt;strong&gt;SampleAttention은 기존 LLM의 attention을 거의 정확도 손실 없이 대체하면서, 최대 2.42배 TTFT(Time-to-First-Token) 지연을 줄이는 구조화된 adaptive sparse attention 기법이다.&lt;/strong&gt;&lt;br&gt;&#10;핵심은 두 가지 sparse 패턴인 &lt;code&gt;local window&lt;/code&gt;와 &lt;code&gt;column stripe&lt;/code&gt;를 활용하여 각 attention head에 대해 동적으로 희소 attention mask를 구성하고, FlashAttention 대비 더 높은 하드웨어 효율성과 가속 성능을 달성한다.&lt;/p&gt;</description></item><item><title>TRAINING ULTRA LONG CONTEXT LANGUAGE MODEL WITH FULLY PIPELINED DISTRIBUTED TRANSFORMER</title><link>https://jaehun.me/posts/training-ultra-long-context-language-model-with-fully-pipelined-distributed-transformer/</link><pubDate>Mon, 24 Mar 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/training-ultra-long-context-language-model-with-fully-pipelined-distributed-transformer/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2408.16978"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h3 id="-결론-요약"&gt;✅ 결론 요약&lt;a href="#-%ea%b2%b0%eb%a1%a0-%ec%9a%94%ec%95%bd" 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;/h3&gt;&lt;p&gt;이 논문은 &lt;strong&gt;초장문(long-context)&lt;/strong&gt; LLM을 &lt;strong&gt;저렴한 하드웨어(예: 4 GPU)&lt;/strong&gt; 상에서 효율적으로 훈련할 수 있게 하는 &lt;strong&gt;FPDT (Fully Pipelined Distributed Transformer)&lt;/strong&gt; 구조를 제안함.&lt;br&gt;&#10;기존 대비 &lt;strong&gt;최대 16배 더 긴 시퀀스&lt;/strong&gt;(예: 2M tokens)를 &lt;strong&gt;55% 이상의 MFU(Model FLOPs Utilization)&lt;/strong&gt; 효율로 훈련 가능함.&lt;/p&gt;</description></item><item><title>QServe: W4A8KV4 Quantization and System Co-design for Efficient LLM Serving</title><link>https://jaehun.me/posts/qserve-w4a8kv4-quantization-and-system-co-design-for-efficient-llm-serving/</link><pubDate>Tue, 18 Mar 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/qserve-w4a8kv4-quantization-and-system-co-design-for-efficient-llm-serving/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2405.04532"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h3 id="논문의-강점과-독창적인-지점"&gt;&lt;strong&gt;논문의 강점과 독창적인 지점&lt;/strong&gt;&lt;a href="#%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90" 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;/h3&gt;&lt;h4 id="1-강점"&gt;&lt;strong&gt;1. 강점&lt;/strong&gt;&lt;a href="#1-%ea%b0%95%ec%a0%90" 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;/h4&gt;&lt;ul&gt;&#10;&lt;li&gt;&lt;strong&gt;저비트 정량화(Quantization)의 실용적 개선:&lt;/strong&gt; 기존 INT4 정량화 기법들이 클라우드 기반 LLM 서빙에서 성능 개선을 보이지 못하는 문제를 해결.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;QoQ(W4A8KV4) 알고리즘 제안:&lt;/strong&gt;&#10;&lt;ul&gt;&#10;&lt;li&gt;**4비트 가중치(W4), 8비트 활성화(A8), 4비트 KV 캐시(KV4)**를 적용하여 정량화에 따른 정확도 손실을 최소화.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;진행형 그룹 정량화(Progressive Group Quantization):&lt;/strong&gt; 8비트 중간 표현을 활용하여 INT8 텐서 코어에서 연산을 수행, 기존 INT4 방식보다 높은 성능 제공.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;SmoothAttention 기법:&lt;/strong&gt; 4비트 KV 정량화에 따른 정확도 저하를 완화하는 메커니즘.&lt;/li&gt;&#10;&lt;/ul&gt;&#10;&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;QServe 시스템과 알고리즘 공동 설계(System-Algorithm Co-design):&lt;/strong&gt;&#10;&lt;ul&gt;&#10;&lt;li&gt;&lt;strong&gt;GPU 서빙 성능 극대화:&lt;/strong&gt; CUDA 코어에서 수행되는 비효율적인 연산을 줄이고, 텐서 코어 활용도를 극대화.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;레지스터 수준 병렬성(Register-Level Parallelism) 활용:&lt;/strong&gt; INT4→INT8 변환 시 감산 후 곱셈(Subtraction after Multiplication) 방식을 적용하여 연산량 감소.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;연산 중심 가중치 재배열(Compute-aware Weight Reordering):&lt;/strong&gt; CUDA 코어에서의 포인터 연산량을 줄여 L1 캐시 활용 최적화.&lt;/li&gt;&#10;&lt;/ul&gt;&#10;&lt;/li&gt;&#10;&lt;/ul&gt;&#10;&lt;h4 id="2-독창적인-지점"&gt;&lt;strong&gt;2. 독창적인 지점&lt;/strong&gt;&lt;a href="#2-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90" 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;/h4&gt;&lt;table&gt;&#10;&#9;&lt;thead&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;th&gt;기존 기법&lt;/th&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;th&gt;QoQ (논문 기법)&lt;/th&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&lt;/thead&gt;&#10;&#9;&lt;tbody&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;W4A4의 낮은 정확도 문제&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;W4A8로 INT8 텐서 코어 활용 가능&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;W4A16의 높은 메모리 사용량&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;KV4 도입으로 메모리 효율 개선&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;INT4 GEMM에서 발생하는 CUDA Core 연산 병목&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Register-Level Parallelism으로 해결&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;기존 KV 캐시 정량화의 정확도 저하&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;SmoothAttention으로 키(Key) 값 정규화&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&lt;/tbody&gt;&#10;&lt;/table&gt;&#10;&lt;hr&gt;&#10;&lt;h3 id="핵심-알고리즘-과정-설명"&gt;&lt;strong&gt;핵심 알고리즘 과정 설명&lt;/strong&gt;&lt;a href="#%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98-%ea%b3%bc%ec%a0%95-%ec%84%a4%eb%aa%85" 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;/h3&gt;&lt;h4 id="1-qoq-정량화-알고리즘"&gt;&lt;strong&gt;1. QoQ 정량화 알고리즘&lt;/strong&gt;&lt;a href="#1-qoq-%ec%a0%95%eb%9f%89%ed%99%94-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98" 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;/h4&gt;&lt;p&gt;QoQ는 두 단계의 정량화로 이루어짐:&lt;/p&gt;</description></item><item><title>Venn: Resource Management Across Federated Learning Jobs</title><link>https://jaehun.me/posts/venn-resource-management-across-federated-learning-jobs/</link><pubDate>Tue, 18 Mar 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/venn-resource-management-across-federated-learning-jobs/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2312.08298"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;h2 id="논문의-강점과-독창적인-지점"&gt;논문의 강점과 독창적인 지점&lt;a href="#%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90" 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;&lt;p&gt;이 논문은 &lt;strong&gt;연합 학습(Federated Learning, FL)의 리소스 관리 문제&lt;/strong&gt;를 다루며, 특히 다수의 FL 작업이 동일한 디바이스 풀에서 실행될 때 발생하는 &lt;strong&gt;자원 경쟁(Resource Contention)&lt;/strong&gt; 을 해결하는 &lt;strong&gt;Venn&lt;/strong&gt;이라는 새로운 리소스 관리 시스템을 제안합니다.&lt;/p&gt;</description></item><item><title>AI Metropolis: Scaling Large Language Model-based Multi-Agent Simulation with Out-of-order Execution</title><link>https://jaehun.me/posts/ai-metropolis-scaling-large-language-model-based-multi-agent-simulation-with-out-of-order-execution/</link><pubDate>Mon, 17 Mar 2025 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/ai-metropolis-scaling-large-language-model-based-multi-agent-simulation-with-out-of-order-execution/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2411.03519"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;h1 id="q--이-논문을-아주-자세하게-읽고-논문의-강점과-독창적인-지점을-설명해주고-핵심-알고리즘을-예시-입력을-들어서-전체적인-과정을-설명해줘-추가적으로-논문의-한계점에-대해서도-알려줘"&gt;Q : 이 논문을 아주 자세하게 읽고 논문의 강점과 독창적인 지점을 설명해주고 핵심 알고리즘을 예시 입력을 들어서 전체적인 과정을 설명해줘 추가적으로 논문의 한계점에 대해서도 알려줘&lt;a href="#q--%ec%9d%b4-%eb%85%bc%eb%ac%b8%ec%9d%84-%ec%95%84%ec%a3%bc-%ec%9e%90%ec%84%b8%ed%95%98%ea%b2%8c-%ec%9d%bd%ea%b3%a0-%eb%85%bc%eb%ac%b8%ec%9d%98-%ea%b0%95%ec%a0%90%ea%b3%bc-%eb%8f%85%ec%b0%bd%ec%a0%81%ec%9d%b8-%ec%a7%80%ec%a0%90%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a3%bc%ea%b3%a0-%ed%95%b5%ec%8b%ac-%ec%95%8c%ea%b3%a0%eb%a6%ac%ec%a6%98%ec%9d%84-%ec%98%88%ec%8b%9c-%ec%9e%85%eb%a0%a5%ec%9d%84-%eb%93%a4%ec%96%b4%ec%84%9c-%ec%a0%84%ec%b2%b4%ec%a0%81%ec%9d%b8-%ea%b3%bc%ec%a0%95%ec%9d%84-%ec%84%a4%eb%aa%85%ed%95%b4%ec%a4%98-%ec%b6%94%ea%b0%80%ec%a0%81%ec%9c%bc%eb%a1%9c-%eb%85%bc%eb%ac%b8%ec%9d%98-%ed%95%9c%ea%b3%84%ec%a0%90%ec%97%90-%eb%8c%80%ed%95%b4%ec%84%9c%eb%8f%84-%ec%95%8c%eb%a0%a4%ec%a4%98" 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;/h1&gt;&lt;p&gt;논문『AI Metropolis: Scaling Large Language Model-based Multi-Agent Simulation with Out-of-order Execution』의 주요 강점과 독창적인 지점, 핵심 알고리즘 및 한계점을 압축하여 설명하면 다음과 같습니다.&lt;/p&gt;</description></item></channel></rss>