地下空间战略 · FP8量化实战

一、 权威必看

EN: At the 13th International Underground Space Development Conference (IFUS 2026) held on July 4, Academician Qian Qihu highlighted that China’s urban development is shifting from large-scale expansion to quality improvement. He stated that underground space has become a critical carrier for enhancing urban resilience and serving national strategies. The conference, co-hosted by Tongji University and Shenzhen University, focused on the integration of artificial intelligence across the entire chain of underground space planning, design, construction, and operation. 中： 7月4日，在由同济大学、深圳大学等共同主办的第十三届国际地下空间开发大会（IFUS 2026）上，中国工程院院士钱七虎指出，我国城市发展已正式从大规模增量扩张转向全量提质增效的新阶段。他强调，地下空间的战略价值日益凸显，成为提升城市韧性、拓展发展空间和服务国家战略的重要载体。本届大会以“共智共生 融合发展”为主题，近百位国内外权威院士及行业领军企业代表齐聚一堂，共同探索现代化城市地下空间产业发展的新范式，特别是人工智能与地下空间全链条的深度融合。

二、 深度与多元

EN: Academician Qian Qihu specifically referenced the “15th Five-Year Plan” outline released in March, which explicitly calls for the comprehensive implementation of the “AI+” action. Recent urban renewal plans further require underground pipe network inspections, comprehensive utility corridor construction, and the acceleration of urban infrastructure lifeline safety projects. These tasks rely heavily on the deep cross-fusion of digitalization, automation, and intelligent technologies to reshape national security dimensions and ecological foundations. 中： 钱七虎特别提到，今年3月发布的“十五五”规划纲要明确提出全面实施“人工智能+”行动，近期城市更新“十五五”规划进一步要求实施地下管网体检、推进综合管廊建设以及加快城市基础设施生命线安全工程。构建CIM（城市信息模型）组织底座已成为关键任务，这些任务的落地都离不开数字化、自动化与智能技术的深度交叉融合。从战略纵深到城市命脉再到生态基石，地下空间的深度开发正以前所未有的方式重塑国家安全的维度与韧性，这不仅是工程技术的升级，更是国家治理能力的现代化体现。

三、 科技与财经

EN: In the developer community, a detailed technical analysis on Juejin demonstrates how FP8 quantization can halve the VRAM usage of Llama 3 70B while tripling throughput without hardware changes. By utilizing online dynamic quantization and offline static quantization combined with AutoFP8 and llm-compressor, developers can fully utilize Tensor Cores. Additionally, KV Cache FP8 allows for 128K context windows without Out-Of-Memory errors, offering a practical solution for deploying large language models on limited hardware. 中： 在开发者社区，一篇来自掘金的技术文章深入解析了FP8量化在vLLM与SGLang部署DeepSeek及Llama 3 70B时的实战效果。通过不换GPU且不降低精度的方式，FP8量化成功使显存占用减半，吞吐量提升3倍。文章指出，在线动态量化仅需一条命令即可验证，而离线静态量化结合AutoFP8与llm-compressor工具则能跑满Tensor Core性能。此外，KV Cache FP8技术的应用使得128K长上下文不再出现OOM（内存溢出）错误，为在有限硬件资源下高效部署大模型提供了极具价值的技术路径。

四、 国际视野

EN: On Weibo, the trending topic “Wang Leehom’s deep wound” has sparked widespread discussion among netizens. This topic reflects current public attention and diverse viewpoints on social media regarding celebrity personal lives. While it serves as a focal point for online discourse, it remains a subject of internet speculation rather than confirmed official news, highlighting the rapid spread of cultural and entertainment news in the digital age. 中： 在微博实时热搜榜上，“王力宏好深的伤口”话题引发了广大网友的关注和讨论。该话题反映了当前网络舆论对娱乐明星私人生活的关注焦点，网民们在此背景下交换着各自的观点与情绪。作为社交媒体上的热点，它更多体现的是公众对文化人物命运的情感投射与讨论，而非具有实质政策或技术影响的国际地缘事件。此类话题的传播速度极快，但也提醒我们在面对网络流传观点时需保持客观中立，区分娱乐八卦与严肃新闻的界限。

五、 青年与生活

EN: V2EX users have issued a warning regarding a connection pool bug in AutoBangumi versions v3.2.2–v3.2.6, which can cause abnormal disk writes when used with Clash/Mihomo Party proxies. The issue causes proxy programs to continuously flush warning logs, leading to high disk I/O. This bug was fixed in version v3.2.8 released on July 2. Users are advised to check their system’s disk write volume and SSD SMART health if they ran these versions during that period. 中： V2EX社区用户发布了一则关于开源项目AutoBangumi的紧急技术提示，指出v3.2.2至v3.2.6版本在Clash或Mihomo Party代理环境下存在连接池Bug。该缺陷会导致代理程序持续刷新Warning日志，从而产生极高的磁盘写入量，可能影响SSD寿命。开发者已于7月2日发布的v3.2.8版本中修复了此问题。建议曾在此期间运行过相关版本的用户尽快检查系统磁盘写入量和SSD的SMART健康状态，以评估潜在硬件损耗风险。

【21ZHAO 综合判断】

EN:

• For Developers: Prioritize hardware efficiency by adopting FP8 quantization techniques for LLM deployment to reduce costs and improve inference speed, as demonstrated by the recent technical breakthroughs in the community.
• For Urban Planners & Tech Investors: Monitor the “AI+” integration in underground infrastructure projects, as this represents a significant policy-driven market opportunity for smart city solutions and CIM platforms. 中：
• 给开发者的建议： 在部署大语言模型时，应优先考虑采用FP8量化技术以优化显存占用并提升推理吞吐量，这是当前降低算力成本、提高硬件利用率的有效技术手段。
• 给城市规划者与投资者的建议： 密切关注“人工智能+”在地下基础设施领域的政策导向，CIM底座与智能管廊建设将是未来几年智慧城市领域的重要市场机遇，具有长期的战略投资价值。

参考来源

• [权威要闻]：梁朝伟被问如何用眼睛演戏，表示自己花很多时间准备，抛开奖项光环，你认为他在华语影坛的演技属于什么级别？ - 原文链接
• [深度解读]：钱七虎院士：地下空间成为提升城市韧性、拓展发展空间和服务国家战略的重要载体 - 原文链接
• [科技财经]：FP8量化实战：vLLM与SGLang部署DeepSeek显存减半、吞吐翻倍——Agent推理引擎篇（二） - 原文链接
• [国际视野]：王力宏好深的伤口—微博热搜焦点话题解析 - 原文链接
• [青年声音]：警告： AutoBangumi v3.2.2–v3.2.6 连接池 Bug 在 Clash/Mihomo Party 代理环境下可能导致异常磁盘写入 - 原文链接