feat: add aiera.com.cn 新智元 RSS route#22118
Open
panqingjie00 wants to merge 7 commits into
Open
Conversation
[pull] master from diygod:master
[pull] master from diygod:master
Contributor
Auto Review
|
Contributor
|
Successfully generated as following: http://localhost:1200/aiera/latest - Failed ❌ |
Contributor
|
Successfully generated as following: http://localhost:1200/aiera/latest - Failed ❌ |
Contributor
|
Successfully generated as following: http://localhost:1200/aiera/latest - Success ✔️<?xml version="1.0" encoding="UTF-8"?>
<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
<channel>
<title>新智元 - 最新文章</title>
<link>https://aiera.com.cn</link>
<atom:link href="http://localhost:1200/aiera/latest" rel="self" type="application/rss+xml"></atom:link>
<description>新智元最新 AI 新闻资讯 - Powered by RSSHub</description>
<generator>RSSHub</generator>
<webMaster>contact@rsshub.app (RSSHub)</webMaster>
<language>zh-CN</language>
<lastBuildDate>Wed, 27 May 2026 05:30:09 GMT</lastBuildDate>
<ttl>5</ttl>
<item>
<title>一个月的活一周干完!英伟达世界模型训练速度飙升400%</title>
<description><h3 data-mpa-powered-by="yiban.io" style="outline: 0px;color: rgb(34, 34, 34);letter-spacing: 0.544px;font-family: -apple-system-font, system-ui, &quot;Helvetica Neue&quot;, &quot;PingFang SC&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;background-color: rgb(255, 255, 255);visibility: visible;">
<section data-tools="135编辑器" data-id="88402" style="outline: 0px;line-height: 27.2px;widows: 1;visibility: visible;">
<section data-tools="135编辑器" data-id="88402" style="outline: 0px;letter-spacing: 0.544px;line-height: 27.2px;visibility: visible;">
<section data-style="line-height: 1.8; text-align: justify; font-size: 15px; letter-spacing: 0px; color: rgb(117, 114, 114);white-space: normal;" style="outline: 0px;visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<section style="margin-bottom: 8px;text-align: center;margin-top: 0px;">
<span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_6e284f6c87.jpg" alt="" referrerpolicy="no-referrer"></span><br>
</section>
</section>
</section>
</section>
</section>
</h3>
<h3 style="outline: 0px;color: rgb(34, 34, 34);letter-spacing: 0.544px;font-family: -apple-system-font, system-ui, &quot;Helvetica Neue&quot;, &quot;PingFang SC&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;background-color: rgb(255, 255, 255);visibility: visible;">
<section data-tools="135编辑器" data-id="88402" style="outline: 0px;line-height: 27.2px;widows: 1;visibility: visible;">
<section data-tools="135编辑器" data-id="88402" style="outline: 0px;letter-spacing: 0.544px;line-height: 27.2px;visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<p style="margin-top: -1.2em;margin-right: 8px;margin-left: 8px;outline: 0px;font-size: 17px;letter-spacing: 0.544px;text-align: center;line-height: 1.75em;visibility: visible;"><span style="outline: 0px;letter-spacing: 1px;visibility: visible;"><strong style="outline: 0px;font-family: inherit;font-size: 1em;text-decoration: inherit;visibility: visible;"><span style="outline: 0px;font-size: 18px;color: rgb(255, 255, 255);line-height: 1.4;font-family: inherit;font-weight: inherit;text-decoration: inherit;background-color: rgb(127, 127, 127);visibility: visible;"><span leaf="">&nbsp;&nbsp;</span></span></strong><strong style="outline: 0px;font-size: 1em;font-family: inherit;text-decoration: inherit;visibility: visible;"><span style="outline: 0px;font-size: 18px;color: rgb(255, 255, 255);line-height: 1.4;font-family: inherit;font-weight: inherit;text-decoration: inherit;background-color: rgb(127, 127, 127);visibility: visible;"><span leaf="">新智元报道 &nbsp;</span></span></strong></span></p>
</section>
</section>
</section>
</section>
</h3>
<section powered-by="xiumi.us" style="margin-bottom: 0px;outline: 0px;color: rgb(34, 34, 34);letter-spacing: 0.544px;font-family: -apple-system-font, system-ui, &quot;Helvetica Neue&quot;, &quot;PingFang SC&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;background-color: rgb(255, 255, 255);visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<p style="text-align: center;"><span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_3b253ef414-60.png" alt="" referrerpolicy="no-referrer"></span></p>
<section powered-by="xiumi.us" class="js_darkmode__5" style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);margin: 0px;outline: 0px;background-color: rgb(255, 255, 255);color: rgb(34, 34, 34);letter-spacing: 0.544px;white-space: normal;font-family: -apple-system-font, system-ui, &quot;Helvetica Neue&quot;, &quot;PingFang SC&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;visibility: visible;">
<section style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;visibility: visible;">
<section style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;visibility: visible;">
<h5 class="js_darkmode__6" style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);margin: 10px 8px 0px;padding: 10px;outline: 0px;font-size: 14px;background-color: rgb(248, 248, 248);color: rgb(0, 0, 0);letter-spacing: 0.544px;font-family: Arial, Helvetica, sans-serif;border-radius: 3px;line-height: 1.75em;visibility: visible;word-break: break-all !important;word-spacing: 1px !important;"><span style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;letter-spacing: 1px;font-size: 15px;visibility: visible;"><strong style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;visibility: visible;"><span leaf="" style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;visibility: visible;">【新智元导读】<span textstyle="" style="font-weight: normal;">英伟达世界动作模型 DreamZero 训练一次要烧 8 张 H100 整整 25 天,RLinf 从算子融合到 I/O 全链路系统级重构,把训练吞吐拉高近 4 倍——1 个月的活,1 周就能干完。</span></span></strong></span></h5>
</section>
</section>
</section>
</section>
</section>
</section>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;margin-top: 24px;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">在通往 AGI 的道路上,世界模型(World Model)被视为让 AI 真正理解并预测物理世界的关键拼图。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">英伟达近期重磅发布的世界动作模型(WAM) DreamZero 一经发布就在两项机器人基准测试 RoboArena 、MolmoSpaces 上双双登顶,在具身智能领域获得极大关注。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">与传统VLA等模型不同,WAM将视频这一具备完整时空信息的载体当作自己的核心学习材料,并以一种「先理解世界如何变化,再决定自己如何行动」的模式,使模型天然获得互联网视频所蕴含的海量物理经验。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">它不再需要大量重复演示来学习单一动作,而是能从多样化的数据中学习世界的物理规律,从而在从未见过的环境和任务中依然保持稳定执行能力。</span></span></strong></p>
<section style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.75em;letter-spacing: 0.034em;font-style: normal;font-weight: normal;margin-left: 8px;margin-right: 8px;margin-bottom: 0px;" nodeleaf="">
<img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_7732b8c4c5.png" alt="" referrerpolicy="no-referrer"><br>
</section>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;text-align: center;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;&nbsp;</span></span><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 13px;letter-spacing: 1px;color: rgb(136, 136, 136);">当前最优的VLA模型与DreamZero世界模型在任务成功率、泛化性、跨本体等方面的直观对比</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">上面的表格直观的展示出 DreamZero 模型相比开源最优的 VLA 模型 π0.5,在任务成功率、任务泛化性、后训练对成功率的提升效果、以及跨真机本体的泛化性等方面具有明显的优势,实现了超过 2x 的成功率提升。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">它的范式革新不仅大幅降低了学习成本,也让机器人的形态适配与技能拓展不再受限于大量专属数据,为多机型协同、快速部署与低成本迭代提供了可行路径。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">然而,</span></span><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">以 Diffusion 架构为主体的 WAM 多模态模型,也给算力和显存带来了巨大的挑战。</span></span></strong></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">参考官方开源的 DreamZero 训练代码,采用 8 台 H100 训练 24750 万帧数据,完整训练周期长达 25 天,高昂的训练成本和耗时成为行业复现的主要门槛。</span></span></strong></p>
<p data-pm-slice="0 0 []" style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span data-eleid="3"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">为助力前沿研究更高效地落地,</span></span></span><span data-eleid="4"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">无问芯穹与清华大学等联合推出的大规模强化学习框架 RLinf 已正式上线了对 DreamZero 训练的深度支持。</span></span></span></p>
<p data-pm-slice="0 0 []" style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span data-eleid="4"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">在实现功能适配的基础之上更进一步,依托 RLinf 强大的底层系统优化能力,对 DreamZero 的训练管线进行了深度的重构与加速。</span></span></span></p>
<p data-pm-slice="0 0 []" style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span data-eleid="4"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">相比 DreamZero 官方提供的基线训练脚本,RLinf 成功实现了近 4 倍的训练吞吐加速,且具有更好的收敛效果。</span></span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">RLinf 是如何极致榨干 GPU 的每一滴算力,达成 4 倍训练加速的?接下来将为您一文拆解背后的核心优化思路与逻辑。</span></span></p>
<section style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.75em;letter-spacing: 0.034em;font-style: normal;font-weight: normal;margin-left: 8px;margin-right: 8px;margin-bottom: 0px;" nodeleaf="">
<img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_cd5464740f.png" alt="" referrerpolicy="no-referrer"><br>
</section>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;text-align: left;margin-bottom: 0px;"><span leaf="" style="line-height: 1.75em;color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 13px;letter-spacing: 1px;color: rgb(136, 136, 136);">代码链接:https://github.com/RLinf/RLinf</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;text-align: left;margin-bottom: 0px;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 13px;letter-spacing: 1px;color: rgb(136, 136, 136);">Hugging Face链接:https://huggingface.co/RLinf</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;text-align: left;margin-bottom: 0px;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 13px;letter-spacing: 1px;color: rgb(136, 136, 136);">使用文档链接:https://rlinf.readthedocs.io/zh-cn/latest/rst_source/examples/embodied/sft_dreamzero.html</span></span></p>
<section style="text-align: center;margin-top: 48px;margin-bottom: 0px;">
<span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_df9d89c9b2-206.png" alt="" referrerpolicy="no-referrer"></span><br>
</section>
<section style="text-align: center;margin-bottom: 0px;margin-top: 8px;line-height: 1.75em;">
<span style="color: rgb(0, 0, 0);font-size: 19px;letter-spacing: 1px;"><strong><span leaf="">核心揭秘</span></strong></span><br>
</section>
<section style="text-align: center;margin-bottom: 0px;margin-top: 8px;line-height: 1.75em;">
<span style="color: rgb(0, 0, 0);font-size: 19px;letter-spacing: 1px;"><strong><span leaf="">近 4 倍加速背后的</span></strong></span><br>
</section>
<section style="text-align: center;margin-bottom: 0px;margin-top: 8px;line-height: 1.75em;">
<span style="color: rgb(0, 0, 0);font-size: 19px;letter-spacing: 1px;"><strong><span leaf="">3 大优化维度</span></strong></span><br>
</section>
<h2 style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><br></span></strong></h2>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">为了打破官方脚本的性能瓶颈,RLinf 系统优化团队从计算图、FSDP2并行优化与全局参数调优、数据处理管线进行了深度优化。</span></span></p>
<section style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.75em;letter-spacing: 0.034em;font-style: normal;font-weight: normal;margin-left: 8px;margin-right: 8px;" nodeleaf="">
<img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_50d4030b04.png" alt="" referrerpolicy="no-referrer"><br>
</section>
<section style="margin-bottom: 0px;">
<p style="line-height: 1.75em;margin: 24px 8px 8px;"><strong style="letter-spacing: 0.578px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;font-size: var(--articleFontsize);"><span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_779d8ee2aa-131.png" alt="" referrerpolicy="no-referrer"></span></strong></p>
</section>
<section style="margin-bottom: 0px;line-height: 1.75em;margin-left: 8px;margin-right: 8px;">
<span style="letter-spacing: 1px;"><strong><span leaf="">极致的算子/计算图优化:Torch Compile + CUDA Graph</span></strong></span><br>
</section>
<h3 style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;</span></span><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><br></span></strong></h3>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">Python 层面的算子与调度开销往往是限制 GPU 峰值性能的「隐形杀手」。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">在 RLinf 中,我们深度融合了&nbsp;</span></span><code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">torch.compile</span></span></code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;和 CUDA Graph 技术:</span></span></p>
<ul style="list-style-type: disc;margin-left: 8px;margin-right: 8px;" class="list-paddingleft-1">
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">Torch Compile</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:通过底层编译优化,对算子进行深度融合(Kernel Fusion),包括 WanRMSNorm、adaLN-zero 等 Diffusion 架构中的低效算子。</span></span></p>
</li>
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">CUDA</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">&nbsp;Graph</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:将计算图固化,消除 GPU launch 的 CPU 调度瓶颈,在DreamZero的训练中,CausalWanSelfAttention 部分的kernel launch较为密集,CUDA Graph 可以做到有效优化。</span></span></p>
</li>
</ul>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">通过该项优化技术,DreamZero 5B 和 14B 模型在不改变原有mbs=1(此处 mbs 指 mbs per&nbsp;</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">gpu</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">,下同)的配置下分别获得 50%(从1.8s/step降到1.2s/step)和 34%(从9s/step降到6.7s/step)的训练加速。</span></span></strong></p>
<section style="margin-bottom: 0px;">
<p style="line-height: 1.75em;margin: 0px 8px 8px;"><strong style="letter-spacing: 0.578px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;font-size: var(--articleFontsize);"><span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_779d8ee2aa-131.png" alt="" referrerpolicy="no-referrer"></span></strong></p>
</section>
<section style="margin-bottom: 0px;line-height: 1.75em;margin-left: 8px;margin-right: 8px;">
<span style="letter-spacing: 1px;"><strong><span leaf="">计算与显存的联合优化:解锁全方位性能调优</span></strong></span><br>
</section>
<h3><strong><span leaf="" style="color:rgba(0, 0, 0, 0.9);font-size:17px;font-family:&quot;mp-quote&quot;, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height:1.6;letter-spacing:0.034em;font-style:normal;font-weight:normal;"><br></span></strong></h3>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">支持任意 Microbatch Size、并行方式的参数调优以及 Recompute(激活重计算),是业界训练大模型时必不可少的性能调优手段。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">然而,在 DreamZero 官方的 baseline 中,存在着明显的工程局限,例如默认使用 DeepSpeed 的 zero2 offload 并行方法、image encoder 不拼 batch 逐样本执行等,大大降低了性能的调优空间。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">RLinf 团队从底层夯实了工程底座,彻底修复了这些痛点,交付了一套健壮且高度可配的调优矩阵:</span></span></p>
<ul style="list-style-type: disc;margin-left: 8px;margin-right: 8px;" class="list-paddingleft-1">
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">稳定适配 FSDP2</span><span textstyle="" style="font-size: 15px;">&nbsp;</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:FSDP2 是 PyTorch 官方团队推出的最新 ZeRO 实现,也是 RLinf &nbsp;面向中等规模大模型的默认并行方案。此前,在 DreamZero 官方代码中使用的 DeepSpeed 方案存在一定的局限性:由于 ZeRO3 与 VAE 模块中 causal conv 的上下文维护机制存在兼容性冲突,开发者往往被迫回退至性能较低的 ZeRO2 offload 模式。此外,DeepSpeed 在反向传播阶段的 post backward hook 产生了较高的 CPU 侧开销,制约了整体训练吞吐。通过向 FSDP2 训练后端的迁移,我们彻底解决了上述架构冲突与性能瓶颈。用户现在可以根据显存配置需求,在不同的分片策略间灵活切换,确保训练过程的高效与稳定。</span></span></p>
</li>
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">灵活的 Microbatch 设置</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:在 FSDP2 支持 DreamZero 模型训练的初始版本中,Microbatch Size (mbs)、Recompute(激活重计算)与 FSDP2 的策略组合往往会触发复杂的底层计算图冲突,而且 image encoder 不拼 batch 会吞掉一部分开大 mbs 的加速收益。RLinf 通过工程上的努力,彻底解决了 mbs &gt; 1 时与上述特性共存的不兼容问题,并且使得 image encoder 能够高效地拼 batch 执行。这一改进使训练系统具备了更高的灵活性:用户可以不受约束地配置任意 mbs,从而根据硬件资源的显存水位与计算吞吐需求,进行精细化的参数调优,在显存占用与执行效率之间达成更优的工程平衡。举例来说,对 DreamZero 5B 模型的训练,在不开启 Recompute 的情况下,mbs 开到2,相比于原来的 mbs 只能开到1,单步耗时几乎没有变化,1.2s/step 变到 1.3s/step,吞吐增加 85%。</span></span></p>
</li>
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">Recompute机制与加速算子的深度协同</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:针对 PyTorch 原生框架在复杂并行策略下的兼容性局限,RLinf 通过深度的底层工程优化,实现了 Recompute(激活重计算)与 CUDA Graph、FSDP2 的稳定解耦与协同。这一改进将 Recompute 转化为一个高可靠、可量化的性能调优维度。在显存受限的硬件环境下,系统能够以微小的计算耗时为代价,换取显著的显存空间释放,从而支持更大规模的并行任务,大幅提升整体训练吞吐。在 DreamZero 5B的训练中,在不开启 Recompute 情况下,单卡 mbs 只能开到2,最佳速度约 1.2s/step,即1.7 samples/sec/gpu,有 Recompute 情况下,单卡 mbs 开到 32 可获得 7.2 s/step,即 4.4 samples/sec/gpu,同等算力下吞吐提升 158%。可以看到,开启 Recompute 使 mbs 得以大幅增加,从而大大提升算子效率。</span></span></p>
</li>
</ul>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">通过以上FSDP2、mbs、Recompute 的全局参数调优,在 DreamZero 5B 模型训练上,我们在第一项算子优化的基础上(即 1.2 samples/sec/</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">gpu</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">)将训练性能进一步提升了 266%,达到 4.4 samples/sec/gpu。</span></span></strong></p>
<section style="margin-bottom: 0px;">
<p style="line-height: 1.75em;margin: 24px 8px 8px;"><strong style="letter-spacing: 0.578px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;font-size: var(--articleFontsize);"><span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_779d8ee2aa-131.png" alt="" referrerpolicy="no-referrer"></span></strong></p>
</section>
<section style="margin-bottom: 0px;line-height: 1.75em;margin-left: 8px;margin-right: 8px;">
<span style="letter-spacing: 1px;"><strong><span leaf="">突破 I/O 吞吐瓶颈:高效视频数据处理管线</span></strong></span><br>
</section>
<h3 style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;&nbsp;</span></span><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><br></span></strong></h3>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">随着计算密度(即上述两项优化)的显著提升,数据加载效率逐渐成为制约整体训练吞吐的新瓶颈。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">在 DreamZero 的训练实践中,视频数据的解码与预处理过程极其消耗 CPU 资源。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">传统的方案(如 PyAV)在解码性能上难以支撑高频的吞吐需求;而单纯通过增加&nbsp;</span></span><code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">dataset</span></span></code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;的&nbsp;</span></span><code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">num_workers</span></span></code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;来尝试「通过数量换速度」往往治标不治本——过多的数据读取进程会剧烈抢占 CPU 资源,进而导致训练主线程的内核下发(Kernel Launch)出现延迟,反而拖慢了 GPU 的执行节奏。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">为了在「解码速度」与「系统资源开销」之间寻找最优解,RLinf 团队对主流的视频处理库进行了深度的性能 Benchmark:</span></span></p>
<section style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.75em;letter-spacing: 0.034em;font-style: normal;font-weight: normal;margin-left: 8px;margin-right: 8px;" nodeleaf="">
<img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_0d6979035b.png" alt="" referrerpolicy="no-referrer"><br>
</section>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">虽然 Decord 在纯解码速度上略胜一筹,但&nbsp;</span></span><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">Torchcodec</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;在保持同梯队性能的同时,表现出了更优的 CPU 占用稳定性。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">这使得我们能够预留出足够的计算余量给训练主线程,并支持开启更多的&nbsp;</span></span><code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">num_workers</span></span></code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;来并发处理数据。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">相比原生的 PyAV 方案,单个视频的解码时间缩短了近 400</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">ms</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">。在 DreamZero 多视角(左视角、右视角、腕部视角三个视频)的训练场景下,视频解码时间累计节省了 1.2s。</span></span></strong></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">这一&nbsp;</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">I/O</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">&nbsp;端的性能提升,为后续进一步压榨&nbsp;</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">GPU</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">&nbsp;计算潜力提供了充足的数据「弹药」。</span></span></strong></p>
<section style="text-align: center;margin-top: 48px;margin-bottom: 0px;">
<span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_df9d89c9b2-206.png" alt="" referrerpolicy="no-referrer"></span><br>
</section>
<section style="text-align: center;margin-bottom: 0px;margin-top: 8px;line-height: 1.75em;">
<sp |
Contributor
|
Successfully generated as following: http://localhost:1200/aiera/latest - Success ✔️<?xml version="1.0" encoding="UTF-8"?>
<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
<channel>
<title>新智元 - 最新文章</title>
<link>https://aiera.com.cn</link>
<atom:link href="http://localhost:1200/aiera/latest" rel="self" type="application/rss+xml"></atom:link>
<description>新智元最新 AI 新闻资讯 - Powered by RSSHub</description>
<generator>RSSHub</generator>
<webMaster>contact@rsshub.app (RSSHub)</webMaster>
<language>zh-CN</language>
<lastBuildDate>Wed, 27 May 2026 05:47:57 GMT</lastBuildDate>
<ttl>5</ttl>
<item>
<title>一个月的活一周干完!英伟达世界模型训练速度飙升400%</title>
<description><h3 data-mpa-powered-by="yiban.io" style="outline: 0px;color: rgb(34, 34, 34);letter-spacing: 0.544px;font-family: -apple-system-font, system-ui, &quot;Helvetica Neue&quot;, &quot;PingFang SC&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;background-color: rgb(255, 255, 255);visibility: visible;">
<section data-tools="135编辑器" data-id="88402" style="outline: 0px;line-height: 27.2px;widows: 1;visibility: visible;">
<section data-tools="135编辑器" data-id="88402" style="outline: 0px;letter-spacing: 0.544px;line-height: 27.2px;visibility: visible;">
<section data-style="line-height: 1.8; text-align: justify; font-size: 15px; letter-spacing: 0px; color: rgb(117, 114, 114);white-space: normal;" style="outline: 0px;visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<section style="margin-bottom: 8px;text-align: center;margin-top: 0px;">
<span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_6e284f6c87.jpg" alt="" referrerpolicy="no-referrer"></span><br>
</section>
</section>
</section>
</section>
</section>
</h3>
<h3 style="outline: 0px;color: rgb(34, 34, 34);letter-spacing: 0.544px;font-family: -apple-system-font, system-ui, &quot;Helvetica Neue&quot;, &quot;PingFang SC&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;background-color: rgb(255, 255, 255);visibility: visible;">
<section data-tools="135编辑器" data-id="88402" style="outline: 0px;line-height: 27.2px;widows: 1;visibility: visible;">
<section data-tools="135编辑器" data-id="88402" style="outline: 0px;letter-spacing: 0.544px;line-height: 27.2px;visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<p style="margin-top: -1.2em;margin-right: 8px;margin-left: 8px;outline: 0px;font-size: 17px;letter-spacing: 0.544px;text-align: center;line-height: 1.75em;visibility: visible;"><span style="outline: 0px;letter-spacing: 1px;visibility: visible;"><strong style="outline: 0px;font-family: inherit;font-size: 1em;text-decoration: inherit;visibility: visible;"><span style="outline: 0px;font-size: 18px;color: rgb(255, 255, 255);line-height: 1.4;font-family: inherit;font-weight: inherit;text-decoration: inherit;background-color: rgb(127, 127, 127);visibility: visible;"><span leaf="">&nbsp;&nbsp;</span></span></strong><strong style="outline: 0px;font-size: 1em;font-family: inherit;text-decoration: inherit;visibility: visible;"><span style="outline: 0px;font-size: 18px;color: rgb(255, 255, 255);line-height: 1.4;font-family: inherit;font-weight: inherit;text-decoration: inherit;background-color: rgb(127, 127, 127);visibility: visible;"><span leaf="">新智元报道 &nbsp;</span></span></strong></span></p>
</section>
</section>
</section>
</section>
</h3>
<section powered-by="xiumi.us" style="margin-bottom: 0px;outline: 0px;color: rgb(34, 34, 34);letter-spacing: 0.544px;font-family: -apple-system-font, system-ui, &quot;Helvetica Neue&quot;, &quot;PingFang SC&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;background-color: rgb(255, 255, 255);visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<p style="text-align: center;"><span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_3b253ef414-60.png" alt="" referrerpolicy="no-referrer"></span></p>
<section powered-by="xiumi.us" class="js_darkmode__5" style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);margin: 0px;outline: 0px;background-color: rgb(255, 255, 255);color: rgb(34, 34, 34);letter-spacing: 0.544px;white-space: normal;font-family: -apple-system-font, system-ui, &quot;Helvetica Neue&quot;, &quot;PingFang SC&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;visibility: visible;">
<section style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;visibility: visible;">
<section style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;visibility: visible;">
<h5 class="js_darkmode__6" style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);margin: 10px 8px 0px;padding: 10px;outline: 0px;font-size: 14px;background-color: rgb(248, 248, 248);color: rgb(0, 0, 0);letter-spacing: 0.544px;font-family: Arial, Helvetica, sans-serif;border-radius: 3px;line-height: 1.75em;visibility: visible;word-break: break-all !important;word-spacing: 1px !important;"><span style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;letter-spacing: 1px;font-size: 15px;visibility: visible;"><strong style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;visibility: visible;"><span leaf="" style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;visibility: visible;">【新智元导读】<span textstyle="" style="font-weight: normal;">英伟达世界动作模型 DreamZero 训练一次要烧 8 张 H100 整整 25 天,RLinf 从算子融合到 I/O 全链路系统级重构,把训练吞吐拉高近 4 倍——1 个月的活,1 周就能干完。</span></span></strong></span></h5>
</section>
</section>
</section>
</section>
</section>
</section>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;margin-top: 24px;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">在通往 AGI 的道路上,世界模型(World Model)被视为让 AI 真正理解并预测物理世界的关键拼图。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">英伟达近期重磅发布的世界动作模型(WAM) DreamZero 一经发布就在两项机器人基准测试 RoboArena 、MolmoSpaces 上双双登顶,在具身智能领域获得极大关注。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">与传统VLA等模型不同,WAM将视频这一具备完整时空信息的载体当作自己的核心学习材料,并以一种「先理解世界如何变化,再决定自己如何行动」的模式,使模型天然获得互联网视频所蕴含的海量物理经验。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">它不再需要大量重复演示来学习单一动作,而是能从多样化的数据中学习世界的物理规律,从而在从未见过的环境和任务中依然保持稳定执行能力。</span></span></strong></p>
<section style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.75em;letter-spacing: 0.034em;font-style: normal;font-weight: normal;margin-left: 8px;margin-right: 8px;margin-bottom: 0px;" nodeleaf="">
<img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_7732b8c4c5.png" alt="" referrerpolicy="no-referrer"><br>
</section>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;text-align: center;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;&nbsp;</span></span><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 13px;letter-spacing: 1px;color: rgb(136, 136, 136);">当前最优的VLA模型与DreamZero世界模型在任务成功率、泛化性、跨本体等方面的直观对比</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">上面的表格直观的展示出 DreamZero 模型相比开源最优的 VLA 模型 π0.5,在任务成功率、任务泛化性、后训练对成功率的提升效果、以及跨真机本体的泛化性等方面具有明显的优势,实现了超过 2x 的成功率提升。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">它的范式革新不仅大幅降低了学习成本,也让机器人的形态适配与技能拓展不再受限于大量专属数据,为多机型协同、快速部署与低成本迭代提供了可行路径。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">然而,</span></span><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">以 Diffusion 架构为主体的 WAM 多模态模型,也给算力和显存带来了巨大的挑战。</span></span></strong></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">参考官方开源的 DreamZero 训练代码,采用 8 台 H100 训练 24750 万帧数据,完整训练周期长达 25 天,高昂的训练成本和耗时成为行业复现的主要门槛。</span></span></strong></p>
<p data-pm-slice="0 0 []" style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span data-eleid="3"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">为助力前沿研究更高效地落地,</span></span></span><span data-eleid="4"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">无问芯穹与清华大学等联合推出的大规模强化学习框架 RLinf 已正式上线了对 DreamZero 训练的深度支持。</span></span></span></p>
<p data-pm-slice="0 0 []" style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span data-eleid="4"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">在实现功能适配的基础之上更进一步,依托 RLinf 强大的底层系统优化能力,对 DreamZero 的训练管线进行了深度的重构与加速。</span></span></span></p>
<p data-pm-slice="0 0 []" style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span data-eleid="4"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">相比 DreamZero 官方提供的基线训练脚本,RLinf 成功实现了近 4 倍的训练吞吐加速,且具有更好的收敛效果。</span></span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">RLinf 是如何极致榨干 GPU 的每一滴算力,达成 4 倍训练加速的?接下来将为您一文拆解背后的核心优化思路与逻辑。</span></span></p>
<section style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.75em;letter-spacing: 0.034em;font-style: normal;font-weight: normal;margin-left: 8px;margin-right: 8px;margin-bottom: 0px;" nodeleaf="">
<img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_cd5464740f.png" alt="" referrerpolicy="no-referrer"><br>
</section>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;text-align: left;margin-bottom: 0px;"><span leaf="" style="line-height: 1.75em;color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 13px;letter-spacing: 1px;color: rgb(136, 136, 136);">代码链接:https://github.com/RLinf/RLinf</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;text-align: left;margin-bottom: 0px;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 13px;letter-spacing: 1px;color: rgb(136, 136, 136);">Hugging Face链接:https://huggingface.co/RLinf</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;text-align: left;margin-bottom: 0px;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 13px;letter-spacing: 1px;color: rgb(136, 136, 136);">使用文档链接:https://rlinf.readthedocs.io/zh-cn/latest/rst_source/examples/embodied/sft_dreamzero.html</span></span></p>
<section style="text-align: center;margin-top: 48px;margin-bottom: 0px;">
<span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_df9d89c9b2-206.png" alt="" referrerpolicy="no-referrer"></span><br>
</section>
<section style="text-align: center;margin-bottom: 0px;margin-top: 8px;line-height: 1.75em;">
<span style="color: rgb(0, 0, 0);font-size: 19px;letter-spacing: 1px;"><strong><span leaf="">核心揭秘</span></strong></span><br>
</section>
<section style="text-align: center;margin-bottom: 0px;margin-top: 8px;line-height: 1.75em;">
<span style="color: rgb(0, 0, 0);font-size: 19px;letter-spacing: 1px;"><strong><span leaf="">近 4 倍加速背后的</span></strong></span><br>
</section>
<section style="text-align: center;margin-bottom: 0px;margin-top: 8px;line-height: 1.75em;">
<span style="color: rgb(0, 0, 0);font-size: 19px;letter-spacing: 1px;"><strong><span leaf="">3 大优化维度</span></strong></span><br>
</section>
<h2 style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><br></span></strong></h2>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">为了打破官方脚本的性能瓶颈,RLinf 系统优化团队从计算图、FSDP2并行优化与全局参数调优、数据处理管线进行了深度优化。</span></span></p>
<section style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.75em;letter-spacing: 0.034em;font-style: normal;font-weight: normal;margin-left: 8px;margin-right: 8px;" nodeleaf="">
<img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_50d4030b04.png" alt="" referrerpolicy="no-referrer"><br>
</section>
<section style="margin-bottom: 0px;">
<p style="line-height: 1.75em;margin: 24px 8px 8px;"><strong style="letter-spacing: 0.578px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;font-size: var(--articleFontsize);"><span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_779d8ee2aa-131.png" alt="" referrerpolicy="no-referrer"></span></strong></p>
</section>
<section style="margin-bottom: 0px;line-height: 1.75em;margin-left: 8px;margin-right: 8px;">
<span style="letter-spacing: 1px;"><strong><span leaf="">极致的算子/计算图优化:Torch Compile + CUDA Graph</span></strong></span><br>
</section>
<h3 style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;</span></span><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><br></span></strong></h3>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">Python 层面的算子与调度开销往往是限制 GPU 峰值性能的「隐形杀手」。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">在 RLinf 中,我们深度融合了&nbsp;</span></span><code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">torch.compile</span></span></code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;和 CUDA Graph 技术:</span></span></p>
<ul style="list-style-type: disc;margin-left: 8px;margin-right: 8px;" class="list-paddingleft-1">
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">Torch Compile</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:通过底层编译优化,对算子进行深度融合(Kernel Fusion),包括 WanRMSNorm、adaLN-zero 等 Diffusion 架构中的低效算子。</span></span></p>
</li>
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">CUDA</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">&nbsp;Graph</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:将计算图固化,消除 GPU launch 的 CPU 调度瓶颈,在DreamZero的训练中,CausalWanSelfAttention 部分的kernel launch较为密集,CUDA Graph 可以做到有效优化。</span></span></p>
</li>
</ul>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">通过该项优化技术,DreamZero 5B 和 14B 模型在不改变原有mbs=1(此处 mbs 指 mbs per&nbsp;</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">gpu</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">,下同)的配置下分别获得 50%(从1.8s/step降到1.2s/step)和 34%(从9s/step降到6.7s/step)的训练加速。</span></span></strong></p>
<section style="margin-bottom: 0px;">
<p style="line-height: 1.75em;margin: 0px 8px 8px;"><strong style="letter-spacing: 0.578px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;font-size: var(--articleFontsize);"><span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_779d8ee2aa-131.png" alt="" referrerpolicy="no-referrer"></span></strong></p>
</section>
<section style="margin-bottom: 0px;line-height: 1.75em;margin-left: 8px;margin-right: 8px;">
<span style="letter-spacing: 1px;"><strong><span leaf="">计算与显存的联合优化:解锁全方位性能调优</span></strong></span><br>
</section>
<h3><strong><span leaf="" style="color:rgba(0, 0, 0, 0.9);font-size:17px;font-family:&quot;mp-quote&quot;, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height:1.6;letter-spacing:0.034em;font-style:normal;font-weight:normal;"><br></span></strong></h3>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">支持任意 Microbatch Size、并行方式的参数调优以及 Recompute(激活重计算),是业界训练大模型时必不可少的性能调优手段。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">然而,在 DreamZero 官方的 baseline 中,存在着明显的工程局限,例如默认使用 DeepSpeed 的 zero2 offload 并行方法、image encoder 不拼 batch 逐样本执行等,大大降低了性能的调优空间。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">RLinf 团队从底层夯实了工程底座,彻底修复了这些痛点,交付了一套健壮且高度可配的调优矩阵:</span></span></p>
<ul style="list-style-type: disc;margin-left: 8px;margin-right: 8px;" class="list-paddingleft-1">
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">稳定适配 FSDP2</span><span textstyle="" style="font-size: 15px;">&nbsp;</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:FSDP2 是 PyTorch 官方团队推出的最新 ZeRO 实现,也是 RLinf &nbsp;面向中等规模大模型的默认并行方案。此前,在 DreamZero 官方代码中使用的 DeepSpeed 方案存在一定的局限性:由于 ZeRO3 与 VAE 模块中 causal conv 的上下文维护机制存在兼容性冲突,开发者往往被迫回退至性能较低的 ZeRO2 offload 模式。此外,DeepSpeed 在反向传播阶段的 post backward hook 产生了较高的 CPU 侧开销,制约了整体训练吞吐。通过向 FSDP2 训练后端的迁移,我们彻底解决了上述架构冲突与性能瓶颈。用户现在可以根据显存配置需求,在不同的分片策略间灵活切换,确保训练过程的高效与稳定。</span></span></p>
</li>
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">灵活的 Microbatch 设置</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:在 FSDP2 支持 DreamZero 模型训练的初始版本中,Microbatch Size (mbs)、Recompute(激活重计算)与 FSDP2 的策略组合往往会触发复杂的底层计算图冲突,而且 image encoder 不拼 batch 会吞掉一部分开大 mbs 的加速收益。RLinf 通过工程上的努力,彻底解决了 mbs &gt; 1 时与上述特性共存的不兼容问题,并且使得 image encoder 能够高效地拼 batch 执行。这一改进使训练系统具备了更高的灵活性:用户可以不受约束地配置任意 mbs,从而根据硬件资源的显存水位与计算吞吐需求,进行精细化的参数调优,在显存占用与执行效率之间达成更优的工程平衡。举例来说,对 DreamZero 5B 模型的训练,在不开启 Recompute 的情况下,mbs 开到2,相比于原来的 mbs 只能开到1,单步耗时几乎没有变化,1.2s/step 变到 1.3s/step,吞吐增加 85%。</span></span></p>
</li>
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">Recompute机制与加速算子的深度协同</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:针对 PyTorch 原生框架在复杂并行策略下的兼容性局限,RLinf 通过深度的底层工程优化,实现了 Recompute(激活重计算)与 CUDA Graph、FSDP2 的稳定解耦与协同。这一改进将 Recompute 转化为一个高可靠、可量化的性能调优维度。在显存受限的硬件环境下,系统能够以微小的计算耗时为代价,换取显著的显存空间释放,从而支持更大规模的并行任务,大幅提升整体训练吞吐。在 DreamZero 5B的训练中,在不开启 Recompute 情况下,单卡 mbs 只能开到2,最佳速度约 1.2s/step,即1.7 samples/sec/gpu,有 Recompute 情况下,单卡 mbs 开到 32 可获得 7.2 s/step,即 4.4 samples/sec/gpu,同等算力下吞吐提升 158%。可以看到,开启 Recompute 使 mbs 得以大幅增加,从而大大提升算子效率。</span></span></p>
</li>
</ul>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">通过以上FSDP2、mbs、Recompute 的全局参数调优,在 DreamZero 5B 模型训练上,我们在第一项算子优化的基础上(即 1.2 samples/sec/</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">gpu</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">)将训练性能进一步提升了 266%,达到 4.4 samples/sec/gpu。</span></span></strong></p>
<section style="margin-bottom: 0px;">
<p style="line-height: 1.75em;margin: 24px 8px 8px;"><strong style="letter-spacing: 0.578px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;font-size: var(--articleFontsize);"><span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_779d8ee2aa-131.png" alt="" referrerpolicy="no-referrer"></span></strong></p>
</section>
<section style="margin-bottom: 0px;line-height: 1.75em;margin-left: 8px;margin-right: 8px;">
<span style="letter-spacing: 1px;"><strong><span leaf="">突破 I/O 吞吐瓶颈:高效视频数据处理管线</span></strong></span><br>
</section>
<h3 style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;&nbsp;</span></span><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><br></span></strong></h3>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">随着计算密度(即上述两项优化)的显著提升,数据加载效率逐渐成为制约整体训练吞吐的新瓶颈。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">在 DreamZero 的训练实践中,视频数据的解码与预处理过程极其消耗 CPU 资源。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">传统的方案(如 PyAV)在解码性能上难以支撑高频的吞吐需求;而单纯通过增加&nbsp;</span></span><code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">dataset</span></span></code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;的&nbsp;</span></span><code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">num_workers</span></span></code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;来尝试「通过数量换速度」往往治标不治本——过多的数据读取进程会剧烈抢占 CPU 资源,进而导致训练主线程的内核下发(Kernel Launch)出现延迟,反而拖慢了 GPU 的执行节奏。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">为了在「解码速度」与「系统资源开销」之间寻找最优解,RLinf 团队对主流的视频处理库进行了深度的性能 Benchmark:</span></span></p>
<section style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.75em;letter-spacing: 0.034em;font-style: normal;font-weight: normal;margin-left: 8px;margin-right: 8px;" nodeleaf="">
<img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_0d6979035b.png" alt="" referrerpolicy="no-referrer"><br>
</section>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">虽然 Decord 在纯解码速度上略胜一筹,但&nbsp;</span></span><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">Torchcodec</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;在保持同梯队性能的同时,表现出了更优的 CPU 占用稳定性。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">这使得我们能够预留出足够的计算余量给训练主线程,并支持开启更多的&nbsp;</span></span><code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">num_workers</span></span></code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;来并发处理数据。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">相比原生的 PyAV 方案,单个视频的解码时间缩短了近 400</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">ms</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">。在 DreamZero 多视角(左视角、右视角、腕部视角三个视频)的训练场景下,视频解码时间累计节省了 1.2s。</span></span></strong></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">这一&nbsp;</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">I/O</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">&nbsp;端的性能提升,为后续进一步压榨&nbsp;</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">GPU</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">&nbsp;计算潜力提供了充足的数据「弹药」。</span></span></strong></p>
<section style="text-align: center;margin-top: 48px;margin-bottom: 0px;">
<span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_df9d89c9b2-206.png" alt="" referrerpolicy="no-referrer"></span><br>
</section>
<section style="text-align: center;margin-bottom: 0px;margin-top: 8px;line-height: 1.75em;">
<sp |
Author
|
@DIYgod please review, think you. |
Contributor
|
Successfully generated as following: http://localhost:1200/aiera/latest - Success ✔️<?xml version="1.0" encoding="UTF-8"?>
<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
<channel>
<title>新智元 - 最新文章</title>
<link>https://aiera.com.cn</link>
<atom:link href="http://localhost:1200/aiera/latest" rel="self" type="application/rss+xml"></atom:link>
<description>新智元最新 AI 新闻资讯 - Powered by RSSHub</description>
<generator>RSSHub</generator>
<webMaster>contact@rsshub.app (RSSHub)</webMaster>
<language>zh-CN</language>
<image>
<url>https://aiera.com.cn/wp-content/uploads/2025/01/%E6%96%B0%E6%99%BA%E5%85%83%E7%BD%91%E7%AB%99logo-150x150.png</url>
<title>新智元 - 最新文章</title>
<link>https://aiera.com.cn</link>
</image>
<lastBuildDate>Wed, 27 May 2026 07:00:59 GMT</lastBuildDate>
<ttl>5</ttl>
<item>
<title>一个月的活一周干完!英伟达世界模型训练速度飙升400%</title>
<description><h3 data-mpa-powered-by="yiban.io" style="outline: 0px;color: rgb(34, 34, 34);letter-spacing: 0.544px;font-family: -apple-system-font, system-ui, &quot;Helvetica Neue&quot;, &quot;PingFang SC&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;background-color: rgb(255, 255, 255);visibility: visible;">
<section data-tools="135编辑器" data-id="88402" style="outline: 0px;line-height: 27.2px;widows: 1;visibility: visible;">
<section data-tools="135编辑器" data-id="88402" style="outline: 0px;letter-spacing: 0.544px;line-height: 27.2px;visibility: visible;">
<section data-style="line-height: 1.8; text-align: justify; font-size: 15px; letter-spacing: 0px; color: rgb(117, 114, 114);white-space: normal;" style="outline: 0px;visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<section style="margin-bottom: 8px;text-align: center;margin-top: 0px;">
<span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_6e284f6c87.jpg" alt="" referrerpolicy="no-referrer"></span><br>
</section>
</section>
</section>
</section>
</section>
</h3>
<h3 style="outline: 0px;color: rgb(34, 34, 34);letter-spacing: 0.544px;font-family: -apple-system-font, system-ui, &quot;Helvetica Neue&quot;, &quot;PingFang SC&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;background-color: rgb(255, 255, 255);visibility: visible;">
<section data-tools="135编辑器" data-id="88402" style="outline: 0px;line-height: 27.2px;widows: 1;visibility: visible;">
<section data-tools="135编辑器" data-id="88402" style="outline: 0px;letter-spacing: 0.544px;line-height: 27.2px;visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<p style="margin-top: -1.2em;margin-right: 8px;margin-left: 8px;outline: 0px;font-size: 17px;letter-spacing: 0.544px;text-align: center;line-height: 1.75em;visibility: visible;"><span style="outline: 0px;letter-spacing: 1px;visibility: visible;"><strong style="outline: 0px;font-family: inherit;font-size: 1em;text-decoration: inherit;visibility: visible;"><span style="outline: 0px;font-size: 18px;color: rgb(255, 255, 255);line-height: 1.4;font-family: inherit;font-weight: inherit;text-decoration: inherit;background-color: rgb(127, 127, 127);visibility: visible;"><span leaf="">&nbsp;&nbsp;</span></span></strong><strong style="outline: 0px;font-size: 1em;font-family: inherit;text-decoration: inherit;visibility: visible;"><span style="outline: 0px;font-size: 18px;color: rgb(255, 255, 255);line-height: 1.4;font-family: inherit;font-weight: inherit;text-decoration: inherit;background-color: rgb(127, 127, 127);visibility: visible;"><span leaf="">新智元报道 &nbsp;</span></span></strong></span></p>
</section>
</section>
</section>
</section>
</h3>
<section powered-by="xiumi.us" style="margin-bottom: 0px;outline: 0px;color: rgb(34, 34, 34);letter-spacing: 0.544px;font-family: -apple-system-font, system-ui, &quot;Helvetica Neue&quot;, &quot;PingFang SC&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;background-color: rgb(255, 255, 255);visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<section style="outline: 0px;visibility: visible;">
<p style="text-align: center;"><span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_3b253ef414-60.png" alt="" referrerpolicy="no-referrer"></span></p>
<section powered-by="xiumi.us" class="js_darkmode__5" style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);margin: 0px;outline: 0px;background-color: rgb(255, 255, 255);color: rgb(34, 34, 34);letter-spacing: 0.544px;white-space: normal;font-family: -apple-system-font, system-ui, &quot;Helvetica Neue&quot;, &quot;PingFang SC&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;visibility: visible;">
<section style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;visibility: visible;">
<section style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;visibility: visible;">
<h5 class="js_darkmode__6" style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);margin: 10px 8px 0px;padding: 10px;outline: 0px;font-size: 14px;background-color: rgb(248, 248, 248);color: rgb(0, 0, 0);letter-spacing: 0.544px;font-family: Arial, Helvetica, sans-serif;border-radius: 3px;line-height: 1.75em;visibility: visible;word-break: break-all !important;word-spacing: 1px !important;"><span style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;letter-spacing: 1px;font-size: 15px;visibility: visible;"><strong style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;visibility: visible;"><span leaf="" style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);outline: 0px;visibility: visible;">【新智元导读】<span textstyle="" style="font-weight: normal;">英伟达世界动作模型 DreamZero 训练一次要烧 8 张 H100 整整 25 天,RLinf 从算子融合到 I/O 全链路系统级重构,把训练吞吐拉高近 4 倍——1 个月的活,1 周就能干完。</span></span></strong></span></h5>
</section>
</section>
</section>
</section>
</section>
</section>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;margin-top: 24px;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">在通往 AGI 的道路上,世界模型(World Model)被视为让 AI 真正理解并预测物理世界的关键拼图。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">英伟达近期重磅发布的世界动作模型(WAM) DreamZero 一经发布就在两项机器人基准测试 RoboArena 、MolmoSpaces 上双双登顶,在具身智能领域获得极大关注。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">与传统VLA等模型不同,WAM将视频这一具备完整时空信息的载体当作自己的核心学习材料,并以一种「先理解世界如何变化,再决定自己如何行动」的模式,使模型天然获得互联网视频所蕴含的海量物理经验。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">它不再需要大量重复演示来学习单一动作,而是能从多样化的数据中学习世界的物理规律,从而在从未见过的环境和任务中依然保持稳定执行能力。</span></span></strong></p>
<section style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.75em;letter-spacing: 0.034em;font-style: normal;font-weight: normal;margin-left: 8px;margin-right: 8px;margin-bottom: 0px;" nodeleaf="">
<img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_7732b8c4c5.png" alt="" referrerpolicy="no-referrer"><br>
</section>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;text-align: center;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;&nbsp;</span></span><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 13px;letter-spacing: 1px;color: rgb(136, 136, 136);">当前最优的VLA模型与DreamZero世界模型在任务成功率、泛化性、跨本体等方面的直观对比</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">上面的表格直观的展示出 DreamZero 模型相比开源最优的 VLA 模型 π0.5,在任务成功率、任务泛化性、后训练对成功率的提升效果、以及跨真机本体的泛化性等方面具有明显的优势,实现了超过 2x 的成功率提升。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">它的范式革新不仅大幅降低了学习成本,也让机器人的形态适配与技能拓展不再受限于大量专属数据,为多机型协同、快速部署与低成本迭代提供了可行路径。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">然而,</span></span><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">以 Diffusion 架构为主体的 WAM 多模态模型,也给算力和显存带来了巨大的挑战。</span></span></strong></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">参考官方开源的 DreamZero 训练代码,采用 8 台 H100 训练 24750 万帧数据,完整训练周期长达 25 天,高昂的训练成本和耗时成为行业复现的主要门槛。</span></span></strong></p>
<p data-pm-slice="0 0 []" style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span data-eleid="3"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">为助力前沿研究更高效地落地,</span></span></span><span data-eleid="4"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">无问芯穹与清华大学等联合推出的大规模强化学习框架 RLinf 已正式上线了对 DreamZero 训练的深度支持。</span></span></span></p>
<p data-pm-slice="0 0 []" style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span data-eleid="4"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">在实现功能适配的基础之上更进一步,依托 RLinf 强大的底层系统优化能力,对 DreamZero 的训练管线进行了深度的重构与加速。</span></span></span></p>
<p data-pm-slice="0 0 []" style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span data-eleid="4"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">相比 DreamZero 官方提供的基线训练脚本,RLinf 成功实现了近 4 倍的训练吞吐加速,且具有更好的收敛效果。</span></span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">RLinf 是如何极致榨干 GPU 的每一滴算力,达成 4 倍训练加速的?接下来将为您一文拆解背后的核心优化思路与逻辑。</span></span></p>
<section style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.75em;letter-spacing: 0.034em;font-style: normal;font-weight: normal;margin-left: 8px;margin-right: 8px;margin-bottom: 0px;" nodeleaf="">
<img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_cd5464740f.png" alt="" referrerpolicy="no-referrer"><br>
</section>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;text-align: left;margin-bottom: 0px;"><span leaf="" style="line-height: 1.75em;color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 13px;letter-spacing: 1px;color: rgb(136, 136, 136);">代码链接:https://github.com/RLinf/RLinf</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;text-align: left;margin-bottom: 0px;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 13px;letter-spacing: 1px;color: rgb(136, 136, 136);">Hugging Face链接:https://huggingface.co/RLinf</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;text-align: left;margin-bottom: 0px;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 13px;letter-spacing: 1px;color: rgb(136, 136, 136);">使用文档链接:https://rlinf.readthedocs.io/zh-cn/latest/rst_source/examples/embodied/sft_dreamzero.html</span></span></p>
<section style="text-align: center;margin-top: 48px;margin-bottom: 0px;">
<span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_df9d89c9b2-206.png" alt="" referrerpolicy="no-referrer"></span><br>
</section>
<section style="text-align: center;margin-bottom: 0px;margin-top: 8px;line-height: 1.75em;">
<span style="color: rgb(0, 0, 0);font-size: 19px;letter-spacing: 1px;"><strong><span leaf="">核心揭秘</span></strong></span><br>
</section>
<section style="text-align: center;margin-bottom: 0px;margin-top: 8px;line-height: 1.75em;">
<span style="color: rgb(0, 0, 0);font-size: 19px;letter-spacing: 1px;"><strong><span leaf="">近 4 倍加速背后的</span></strong></span><br>
</section>
<section style="text-align: center;margin-bottom: 0px;margin-top: 8px;line-height: 1.75em;">
<span style="color: rgb(0, 0, 0);font-size: 19px;letter-spacing: 1px;"><strong><span leaf="">3 大优化维度</span></strong></span><br>
</section>
<h2 style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><br></span></strong></h2>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">为了打破官方脚本的性能瓶颈,RLinf 系统优化团队从计算图、FSDP2并行优化与全局参数调优、数据处理管线进行了深度优化。</span></span></p>
<section style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.75em;letter-spacing: 0.034em;font-style: normal;font-weight: normal;margin-left: 8px;margin-right: 8px;" nodeleaf="">
<img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_50d4030b04.png" alt="" referrerpolicy="no-referrer"><br>
</section>
<section style="margin-bottom: 0px;">
<p style="line-height: 1.75em;margin: 24px 8px 8px;"><strong style="letter-spacing: 0.578px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;font-size: var(--articleFontsize);"><span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_779d8ee2aa-131.png" alt="" referrerpolicy="no-referrer"></span></strong></p>
</section>
<section style="margin-bottom: 0px;line-height: 1.75em;margin-left: 8px;margin-right: 8px;">
<span style="letter-spacing: 1px;"><strong><span leaf="">极致的算子/计算图优化:Torch Compile + CUDA Graph</span></strong></span><br>
</section>
<h3 style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;</span></span><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><br></span></strong></h3>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">Python 层面的算子与调度开销往往是限制 GPU 峰值性能的「隐形杀手」。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">在 RLinf 中,我们深度融合了&nbsp;</span></span><code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">torch.compile</span></span></code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;和 CUDA Graph 技术:</span></span></p>
<ul style="list-style-type: disc;margin-left: 8px;margin-right: 8px;" class="list-paddingleft-1">
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">Torch Compile</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:通过底层编译优化,对算子进行深度融合(Kernel Fusion),包括 WanRMSNorm、adaLN-zero 等 Diffusion 架构中的低效算子。</span></span></p>
</li>
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">CUDA</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">&nbsp;Graph</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:将计算图固化,消除 GPU launch 的 CPU 调度瓶颈,在DreamZero的训练中,CausalWanSelfAttention 部分的kernel launch较为密集,CUDA Graph 可以做到有效优化。</span></span></p>
</li>
</ul>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">通过该项优化技术,DreamZero 5B 和 14B 模型在不改变原有mbs=1(此处 mbs 指 mbs per&nbsp;</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">gpu</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">,下同)的配置下分别获得 50%(从1.8s/step降到1.2s/step)和 34%(从9s/step降到6.7s/step)的训练加速。</span></span></strong></p>
<section style="margin-bottom: 0px;">
<p style="line-height: 1.75em;margin: 0px 8px 8px;"><strong style="letter-spacing: 0.578px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;font-size: var(--articleFontsize);"><span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_779d8ee2aa-131.png" alt="" referrerpolicy="no-referrer"></span></strong></p>
</section>
<section style="margin-bottom: 0px;line-height: 1.75em;margin-left: 8px;margin-right: 8px;">
<span style="letter-spacing: 1px;"><strong><span leaf="">计算与显存的联合优化:解锁全方位性能调优</span></strong></span><br>
</section>
<h3><strong><span leaf="" style="color:rgba(0, 0, 0, 0.9);font-size:17px;font-family:&quot;mp-quote&quot;, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height:1.6;letter-spacing:0.034em;font-style:normal;font-weight:normal;"><br></span></strong></h3>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">支持任意 Microbatch Size、并行方式的参数调优以及 Recompute(激活重计算),是业界训练大模型时必不可少的性能调优手段。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">然而,在 DreamZero 官方的 baseline 中,存在着明显的工程局限,例如默认使用 DeepSpeed 的 zero2 offload 并行方法、image encoder 不拼 batch 逐样本执行等,大大降低了性能的调优空间。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">RLinf 团队从底层夯实了工程底座,彻底修复了这些痛点,交付了一套健壮且高度可配的调优矩阵:</span></span></p>
<ul style="list-style-type: disc;margin-left: 8px;margin-right: 8px;" class="list-paddingleft-1">
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">稳定适配 FSDP2</span><span textstyle="" style="font-size: 15px;">&nbsp;</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:FSDP2 是 PyTorch 官方团队推出的最新 ZeRO 实现,也是 RLinf &nbsp;面向中等规模大模型的默认并行方案。此前,在 DreamZero 官方代码中使用的 DeepSpeed 方案存在一定的局限性:由于 ZeRO3 与 VAE 模块中 causal conv 的上下文维护机制存在兼容性冲突,开发者往往被迫回退至性能较低的 ZeRO2 offload 模式。此外,DeepSpeed 在反向传播阶段的 post backward hook 产生了较高的 CPU 侧开销,制约了整体训练吞吐。通过向 FSDP2 训练后端的迁移,我们彻底解决了上述架构冲突与性能瓶颈。用户现在可以根据显存配置需求,在不同的分片策略间灵活切换,确保训练过程的高效与稳定。</span></span></p>
</li>
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">灵活的 Microbatch 设置</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:在 FSDP2 支持 DreamZero 模型训练的初始版本中,Microbatch Size (mbs)、Recompute(激活重计算)与 FSDP2 的策略组合往往会触发复杂的底层计算图冲突,而且 image encoder 不拼 batch 会吞掉一部分开大 mbs 的加速收益。RLinf 通过工程上的努力,彻底解决了 mbs &gt; 1 时与上述特性共存的不兼容问题,并且使得 image encoder 能够高效地拼 batch 执行。这一改进使训练系统具备了更高的灵活性:用户可以不受约束地配置任意 mbs,从而根据硬件资源的显存水位与计算吞吐需求,进行精细化的参数调优,在显存占用与执行效率之间达成更优的工程平衡。举例来说,对 DreamZero 5B 模型的训练,在不开启 Recompute 的情况下,mbs 开到2,相比于原来的 mbs 只能开到1,单步耗时几乎没有变化,1.2s/step 变到 1.3s/step,吞吐增加 85%。</span></span></p>
</li>
<li style="font-size: 15px;">
<p style="margin-bottom: 24px;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;font-weight: bold;">Recompute机制与加速算子的深度协同</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.6;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;">:针对 PyTorch 原生框架在复杂并行策略下的兼容性局限,RLinf 通过深度的底层工程优化,实现了 Recompute(激活重计算)与 CUDA Graph、FSDP2 的稳定解耦与协同。这一改进将 Recompute 转化为一个高可靠、可量化的性能调优维度。在显存受限的硬件环境下,系统能够以微小的计算耗时为代价,换取显著的显存空间释放,从而支持更大规模的并行任务,大幅提升整体训练吞吐。在 DreamZero 5B的训练中,在不开启 Recompute 情况下,单卡 mbs 只能开到2,最佳速度约 1.2s/step,即1.7 samples/sec/gpu,有 Recompute 情况下,单卡 mbs 开到 32 可获得 7.2 s/step,即 4.4 samples/sec/gpu,同等算力下吞吐提升 158%。可以看到,开启 Recompute 使 mbs 得以大幅增加,从而大大提升算子效率。</span></span></p>
</li>
</ul>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">通过以上FSDP2、mbs、Recompute 的全局参数调优,在 DreamZero 5B 模型训练上,我们在第一项算子优化的基础上(即 1.2 samples/sec/</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">gpu</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">)将训练性能进一步提升了 266%,达到 4.4 samples/sec/gpu。</span></span></strong></p>
<section style="margin-bottom: 0px;">
<p style="line-height: 1.75em;margin: 24px 8px 8px;"><strong style="letter-spacing: 0.578px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;font-size: var(--articleFontsize);"><span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_779d8ee2aa-131.png" alt="" referrerpolicy="no-referrer"></span></strong></p>
</section>
<section style="margin-bottom: 0px;line-height: 1.75em;margin-left: 8px;margin-right: 8px;">
<span style="letter-spacing: 1px;"><strong><span leaf="">突破 I/O 吞吐瓶颈:高效视频数据处理管线</span></strong></span><br>
</section>
<h3 style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;&nbsp;</span></span><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><br></span></strong></h3>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">随着计算密度(即上述两项优化)的显著提升,数据加载效率逐渐成为制约整体训练吞吐的新瓶颈。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">在 DreamZero 的训练实践中,视频数据的解码与预处理过程极其消耗 CPU 资源。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">传统的方案(如 PyAV)在解码性能上难以支撑高频的吞吐需求;而单纯通过增加&nbsp;</span></span><code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">dataset</span></span></code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;的&nbsp;</span></span><code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">num_workers</span></span></code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;来尝试「通过数量换速度」往往治标不治本——过多的数据读取进程会剧烈抢占 CPU 资源,进而导致训练主线程的内核下发(Kernel Launch)出现延迟,反而拖慢了 GPU 的执行节奏。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">为了在「解码速度」与「系统资源开销」之间寻找最优解,RLinf 团队对主流的视频处理库进行了深度的性能 Benchmark:</span></span></p>
<section style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;line-height: 1.75em;letter-spacing: 0.034em;font-style: normal;font-weight: normal;margin-left: 8px;margin-right: 8px;" nodeleaf="">
<img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_0d6979035b.png" alt="" referrerpolicy="no-referrer"><br>
</section>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">虽然 Decord 在纯解码速度上略胜一筹,但&nbsp;</span></span><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">Torchcodec</span></span></strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;在保持同梯队性能的同时,表现出了更优的 CPU 占用稳定性。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">这使得我们能够预留出足够的计算余量给训练主线程,并支持开启更多的&nbsp;</span></span><code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">num_workers</span></span></code><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;">&nbsp;来并发处理数据。</span></span></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">相比原生的 PyAV 方案,单个视频的解码时间缩短了近 400</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">ms</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">。在 DreamZero 多视角(左视角、右视角、腕部视角三个视频)的训练场景下,视频解码时间累计节省了 1.2s。</span></span></strong></p>
<p style="margin-left: 8px;margin-right: 8px;line-height: 1.75em;"><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">这一&nbsp;</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">I/O</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">&nbsp;端的性能提升,为后续进一步压榨&nbsp;</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">GPU</span></span></strong><strong><span leaf="" style="color: rgba(0, 0, 0, 0.9);font-size: 17px;font-family: mp-quote, &quot;PingFang SC&quot;, system-ui, -apple-system, BlinkMacSystemFont, &quot;Helvetica Neue&quot;, &quot;Hiragino Sans GB&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, Arial, sans-serif;letter-spacing: 0.034em;font-style: normal;font-weight: normal;"><span textstyle="" style="font-size: 15px;letter-spacing: 1px;color: rgb(255, 104, 39);font-weight: bold;">&nbsp;计算潜力提供了充足的数据「弹药」。</span></span></strong></p>
<section style="text-align: center;margin-top: 48px;margin-bottom: 0px;">
<span leaf=""><img decoding="async" src="https://aiera.com.cn/wp-content/uploads/2026/05/aiera_img_df9d89c9b2-206.png" alt="" |
TonyRL
reviewed
May 28, 2026
| }; | ||
|
|
||
| async function handler() { | ||
| const response = await cache.tryGet('aiera:latest', async () => { |
Collaborator
There was a problem hiding this comment.
Don't cache the response as it will refresh the cache duration on each request. This means the entries won't be updated once it's cached.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Involved Issue / 该 PR 相关 Issue
Close #
Example for the Proposed Route(s) / 路由地址示例
New RSS Route Checklist / 新 RSS 路由检查表
PuppeteerNote / 说明
新智元是 AI 新闻资讯平台,使用 WordPress REST API 获取最新文章,包含全文内容和图片。