https://gpux.ai/blog/sdxl_50_faster

1<!DOCTYPE html><html lang="en"><head><meta charSet="utf-8"/><title>GPUX - Blog</title><meta content="width=device-width,initial-scale=1" name="viewport"/><meta content="
2    GPUX is a distributed supercomputer. 
3  " name="description"/><link rel="stylesheet" href="/build/_assets/index-KQCKENBX.css"/><link rel="stylesheet" href="/build/_assets/navbar-YLEYHBO2.css"/><link rel="stylesheet" href="/build/_assets/footer-SKDWH2HP.css"/><link rel="stylesheet" href="/build/_assets/ai-NZXUWXYO.css"/></head><body><div class="news-banner" style="color:#fde590"><p>🚀 April 20th, 2023 -<a href="https://blog.cloudflare.com/launchpad-fall-22/" target="_blank" rel="noopener noreferrer"><span class="news-banner-link">V2 is launched!</span></a><span class="news-banner-long"> 1s starts from cold! </span> 🎉</p></div><div class="navbar web-align"><div class="navbar-item"><a href="/"><img src="/build/_assets/logo2-E5ZJL7F7.png" alt=""/></a></div><div class="navbar-container"><div class="navbar-item"><div style="padding-left:3px"><a class="Link-tag" href="/blog/sdxl_50_faster"><div class="navbar-btn" style="color:black;background:white"><p>Blog</p></div></a></div><div style="padding-left:3px"><a href="https://discord.com/invite/jjBSjSF" target="_blank" rel="noreferrer" class="Link-tag"><div class="navbar-btn" style="min-height:40px"><div><svg width="100%" viewBox="0 -58.5 256 256" version="1.1" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" preserveAspectRatio="xMidYMid"><g><path d="M216.856339,16.5966031 C200.285002,8.84328665 182.566144,3.2084988 164.041564,0 C161.766523,4.11318106 159.108624,9.64549908 157.276099,14.0464379 C137.583995,11.0849896 118.072967,11.0849896 98.7430163,14.0464379 C96.9108417,9.64549908 94.1925838,4.11318106 91.8971895,0 C73.3526068,3.2084988 55.6133949,8.86399117 39.0420583,16.6376612 C5.61752293,67.146514 -3.4433191,116.400813 1.08711069,164.955721 C23.2560196,181.510915 44.7403634,191.567697 65.8621325,198.148576 C71.0772151,190.971126 75.7283628,183.341335 79.7352139,175.300261 C72.104019,172.400575 64.7949724,168.822202 57.8887866,164.667963 C59.7209612,163.310589 61.5131304,161.891452 63.2445898,160.431257 C105.36741,180.133187 151.134928,180.133187 192.754523,160.431257 C194.506336,161.891452 196.298154,163.310589 198.110326,164.667963 C191.183787,168.842556 183.854737,172.420929 176.223542,175.320965 C180.230393,183.341335 184.861538,190.991831 190.096624,198.16893 C211.238746,191.588051 232.743023,181.531619 254.911949,164.955721 C260.227747,108.668201 245.831087,59.8662432 216.856339,16.5966031 Z M85.4738752,135.09489 C72.8290281,135.09489 62.4592217,123.290155 62.4592217,108.914901 C62.4592217,94.5396472 72.607595,82.7145587 85.4738752,82.7145587 C98.3405064,82.7145587 108.709962,94.5189427 108.488529,108.914901 C108.508531,123.290155 98.3405064,135.09489 85.4738752,135.09489 Z M170.525237,135.09489 C157.88039,135.09489 147.510584,123.290155 147.510584,108.914901 C147.510584,94.5396472 157.658606,82.7145587 170.525237,82.7145587 C183.391518,82.7145587 193.761324,94.5189427 193.539891,108.914901 C193.539891,123.290155 183.391518,135.09489 170.525237,135.09489 Z" fill="#5865F2"></path></g></svg></div></div></a></div></div></div><div class="navbar-mobile-menu"><div class="menu-icon"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" height="27" width="27" class="rsl1IconClass"><path d="M0 0h24v24H0z" fill="none"></path><path d="M3 18h18v-2H3v2zm0-5h18v-2H3v2zm0-7v2h18V6H3z"></path></svg></div></div></div><div class="blog-container web-align"><div class="blog-container-left"><div class="blog-container-wrapper"><h1 class="blog-title">Make StableDiffusionXL 50% faster on RTX 4090</h1><span class="blog-small-text">⚡AI Technology</span><img src="/build/_assets/sdxl_big-KSFONQYT.jpg" alt=""/><hr/><h1>The Problem</h1><p class="blog-desc">Pytorch2 has a lot of optimization improvements but the upstream version when running <pre><code>pip3 install pytorch</code></pre> will pull pytorch2.0.1 with cu117.</p><p class="blog-desc">The problem with this is cu117 does not properly support newer GPUs like the RTX4090 or H100, infact cu117 wont even run on a H100.</p><p class="blog-desc">Popular cloud gpu providers that are used for deep learning often carry the cu117 default, hurting performance.</p><h1>The Fix</h1><p class="blog-desc">To fix this we should be targetting cu118 until its adapted in upstream pip packages.</p><p class="blog-desc">This means upgrading to an image cuda version of atleast 11.8, this can be done in docker by using<pre><code>docker.io/nvidia/cuda:12.2.0-devel-ubuntu22.04</code></pre> as our base image. Or by upgrading our NVIDIA driver + CUDA version.</p><p class="blog-desc">Also we need to pull pytorch with cu118 by doing<pre><code>pip3 install torch==2.0.1+cu118 \<br/>--extra-index-url https://download.pytorch.org/whl/cu118</code></pre></p><p class="blog-desc">We can grab torchvision and torchaudio as well<pre><code>pip3 install torch==2.0.1+cu118 \<br/>torchvision==0.15.2+cu118 \<br/>torchaudio \<br/>--extra-index-url https://download.pytorch.org/whl/cu118</code></pre></p><h1>The Reward</h1><p class="blog-desc">Enjoy an overall 50% speedup on all ADA LOVELACE workloads now!<pre><code>-- cu117 -- <br/>68%|███████▏  | 34/50 [00:04&lt;00:01,  4.88it/s]<br/>70%|███████▍  | 35/50 [00:04&lt;00:01,  4.88it/s]<br/>-- cu118 -- <br/>72%|███████▏  | 36/50 [00:04&lt;00:01,  7.78it/s]<br/>74%|███████▍  | 37/50 [00:04&lt;00:01,  7.78it/s]<br/>76%|███████▌  | 38/50 [00:04&lt;00:01,  7.78it/s]<br/>78%|███████▊  | 39/50 [00:05&lt;00:01,  7.78it/s]<br/>80%|████████  | 40/50 [00:05&lt;00:01,  7.78it/s]<br/></code></pre></p></div></div><div class="blog-container-right">

3<div class="blog-container_card"><img src="https://assets-netstorage.groww.in/web-assets/billion_groww_desktop/prod/build/client/images/loggedOutStateImage.f28263ce.svg" alt=""/><p class="container-card-heading">Building an AI App?</p><p class="container-card-para">Contact us for information to get started</p><button><a style="color:inherit;text-decoration:inherit">JOIN DISCORD</a></button></div></div></div><div class="footer "><div class="web-align"><div class="footer-top"><div class="footer-list"><div class="footer-list-left"><img src="/build/_assets/logo1-3QPGHFMJ.png" alt="" class="footer-logo"/><p class="address">GPUX Inc.</p><p class="address">18 King Street East • Suite 1400</p><p class="address">Toronto • Canada </p><p class="underline">Contact Us</p><div class="social"><a href="https://www.linkedin.com/company/gpux-ai" target="_blank"><img src="https://assets-netstorage.groww.in/web-assets/billion_groww_desktop/prod/build/client/images/linkedin_icon_groww.b15f8240.svg" alt=""/></a><a href="https://discord.gg/jjBSjSF" target="_blank"><img src="/build/_assets/discord-KT36GJJG.png" alt="" class="discord_icon"/></a><a href="https://github.com/gpuedge" target="_blank"><img src="/build/_assets/github_footer-ARMK4DGC.png" alt="" class="discord_icon"/></a></div></div><div class="footer-list-right"><div class="footer-col-3"><div class="col-head">BLOG</div><a class="Link-tag" href="/blog/ai"><p class="container-related-work">AI</p></a><a class="Link-tag" href="/blog/render"><p class="container-related-work">Blender</p></a><a class="Link-tag" href="/blog/transcode"><p class="container-related-work">Transcode</p></a><a class="Link-tag" href="/blog/live"><p class="container-related-work">Live Streaming</p></a><a class="Link-tag" href="/blog/homomorphic_encryption"><p class="container-related-work">Homomorphic Encryption</p></a></div><div class="footer-col-3"><div class="col-head">GPUX</div><a class="Link-tag" href="/blog/sdxl_50_faster"><p class="container-related-work">Blog</p></a></div><div class="footer-col-3"><div class="col-head">LINKS</div><p><a href="https://app.gpux.ai/" target="_blank" rel="noreferrer" style="text-decoration:inherit;color:inherit">App (V2)</a></p></div><div class="footer-col-3"><div class="col-head">LEGAL</div><a class="Link-tag" href="/terms-of-service"><p class="container-related-work">Terms of Service</p></a><a class="Link-tag" href="/privacy"><p class="container-related-work">Privacy Policy</p></a></div></div></div></div><div class="footer-line"></div><div class="footer-bottom"><div class="footer-company">ⓒ  2021-2022 GPUx. All rights reserved, Built with<!-- --> <span class="heart">♥</span></div><div class="footer-store"><div class="button-footer-app"><a href="https://github.com/gpuedge/farm/releases" target="_blank"><img src="/build/_assets/windows_logo-F426UIXQ.png" alt=""/><div><p>EARN ON</p><h1>WINDOWS 10</h1></div></a></div><div class="button-footer-app"><a href="https://github.com/gpuedge/farm/releases" target="_blank"><img src="/build/_assets/linux-KKYEA7LV.png" alt=""/><div><p>EARN ON</p><h1>LINUX OS</h1></div></a></div><div class="button-footer-app"><a href="https://github.com/gpuedge" target="_blank"><img src="/build/_assets/github-JL2PIHTB.png" alt=""/><div><p>GET IT ON</p><h1>GITHUB</h1></div></a></div></div></div></div></div>

3<script>((n) => {
4    if (!window.history.state || !window.history.state.key) {
5      let r = Math.random().toString(32).slice(2);
6      window.history.replaceState({ key: r }, "");
7    }
8    try {
9      let t = JSON.parse(sessionStorage.getItem(n) || "{}")[window.history.state.key];
10      typeof t == "number" && window.scrollTo(0, t);
11    } catch (r) {
12      console.error(r), sessionStorage.removeItem(n);
13    }
14  })("positions")</script>

14<link rel="modulepreload" href="/build/_shared/chunk-PQU25NBF.js"/><link rel="modulepreload" href="/build/root-YJWG4EBQ.js"/><link rel="modulepreload" href="/build/_shared/chunk-QGHTGR7E.js"/><link rel="modulepreload" href="/build/routes/blog-GABXAIE2.js"/><link rel="modulepreload" href="/build/_shared/chunk-27PRSQ6H.js"/><link rel="modulepreload" href="/build/_shared/chunk-IPNYQTEA.js"/><link rel="modulepreload" href="/build/_shared/chunk-KVZIXZ7Y.js"/><link rel="modulepreload" href="/build/routes/blog/sdxl_50_faster-EGM46KAW.js"/>

14<script>window.__remixContext = {'actionData':undefined,'appState':{'trackBoundaries':true,'trackCatchBoundaries':true,'catchBoundaryRouteId':null,'renderBoundaryRouteId':null,'loaderBoundaryRouteId':null,'error':undefined,'catch':undefined},'matches':[{'params':{},'pathname':'/','route':{'id':'root','parentId':undefined,'path':'','index':undefined,'caseSensitive':undefined,'module':'/build/root-YJWG4EBQ.js','imports':undefined,'hasAction':false,'hasLoader':false,'hasCatchBoundary':false,'hasErrorBoundary':false}},{'params':{},'pathname':'/blog','route':{'id':'routes/blog','parentId':'root','path':'blog','index':undefined,'caseSensitive':undefined,'module':'/build/routes/blog-GABXAIE2.js','imports':['/build/_shared/chunk-QGHTGR7E.js'],'hasAction':false,'hasLoader':false,'hasCatchBoundary':false,'hasErrorBoundary':false}},{'params':{},'pathname':'/blog/sdxl_50_faster','route':{'id':'routes/blog/sdxl_50_faster','parentId':'routes/blog','path':'sdxl_50_faster','index':undefined,'caseSensitive':undefined,'module':'/build/routes/blog/sdxl_50_faster-EGM46KAW.js','imports':['/build/_shared/chunk-27PRSQ6H.js','/build/_shared/chunk-IPNYQTEA.js','/build/_shared/chunk-KVZIXZ7Y.js'],'hasAction':false,'hasLoader':false,'hasCatchBoundary':false,'hasErrorBoundary':true}}],'routeData':{}};</script>

14<script src="/build/manifest-2894D64A.js"></script>

14<script type="module">import * as route0 from "/build/root-YJWG4EBQ.js";
15import * as route1 from "/build/routes/blog-GABXAIE2.js";
16import * as route2 from "/build/routes/blog/sdxl_50_faster-EGM46KAW.js";
17window.__remixRouteModules = {"root":route0,"routes/blog":route1,"routes/blog/sdxl_50_faster":route2};</script>

17<script src="/build/entry.client-B3CKKYML.js" type="module"></script>

17<script type="module" src="https://static.cloudflareinsights.com/beacon.min.js/v31edd6df95cf4e85bb4c19e7a9bdbcba1788362987495" integrity="sha512-iIg7k2xntmwu6/uSb5tpc/hySgZc4eoL31yB29W6tJFo2akwjPWcEqnCEdJvGexCL0KEQwVYv5BlowfhVz26hg==" data-cf-beacon='{"version":"2024.11.0","token":"d6d4d912be6a4bb18bccc58e8609ad41","spa":2}' crossorigin="anonymous"></script>

17
18</body></html>