<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[devascend的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/devascend</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; devascend]]></copyright><item><title><![CDATA[万卡集群训练任务挂了？别慌，这本秘籍帮你解决！]]></title><link>https://blog.youkuaiyun.com/devascend/article/details/156569405</link><guid>https://blog.youkuaiyun.com/devascend/article/details/156569405</guid><author>devascend</author><pubDate>Sun, 04 Jan 2026 15:10:45 +0800</pubDate><description><![CDATA[大规模AI集群运行过程中频繁因为各类硬件、软件故障导致训练任务中断，如何快速发现故障、缩短故障恢复时间MTTR成为提高AI集群可用度的重点方向。本文汇总了昇腾断点续训特性提供的故障检测、故障恢复能力以及面临的问题和挑战。]]></description><category></category></item><item><title><![CDATA[RLHF-PPO 昇腾训推共卡方案案例总结（下）]]></title><link>https://blog.youkuaiyun.com/devascend/article/details/156390267</link><guid>https://blog.youkuaiyun.com/devascend/article/details/156390267</guid><author>devascend</author><pubDate>Mon, 29 Dec 2025 15:17:11 +0800</pubDate><description><![CDATA[RLHF-PPO 昇腾训推共卡方案案例总结（下）]]></description><category></category></item><item><title><![CDATA[RLHF-PPO 昇腾训推共卡方案案例总结（上）]]></title><link>https://blog.youkuaiyun.com/devascend/article/details/156385769</link><guid>https://blog.youkuaiyun.com/devascend/article/details/156385769</guid><author>devascend</author><pubDate>Mon, 29 Dec 2025 11:18:52 +0800</pubDate><description><![CDATA[RLHF-PPO 昇腾训推共卡方案案例总结（上）]]></description><category></category></item><item><title><![CDATA[基于KTransformers在300I DUO单机部署满血版DeepSeek-R1的实践]]></title><link>https://blog.youkuaiyun.com/devascend/article/details/156307416</link><guid>https://blog.youkuaiyun.com/devascend/article/details/156307416</guid><author>devascend</author><pubDate>Fri, 26 Dec 2025 17:59:16 +0800</pubDate><description><![CDATA[基于KTransformers在300I DUO单机部署满血版DeepSeek-R1的实践]]></description><category></category></item><item><title><![CDATA[超节点FSDP2训练MOE大模型：GroupedMatmul NZ使能和性能收益分析]]></title><link>https://blog.youkuaiyun.com/devascend/article/details/156309840</link><guid>https://blog.youkuaiyun.com/devascend/article/details/156309840</guid><author>devascend</author><pubDate>Fri, 26 Dec 2025 17:52:31 +0800</pubDate><description><![CDATA[超节点FSDP2训练MOE大模型：GroupedMatmul NZ使能和性能收益分析]]></description><category></category></item><item><title><![CDATA[大规模训练集群性能问题（下降或抖动）分享及性能问题解决方案]]></title><link>https://blog.youkuaiyun.com/devascend/article/details/156310186</link><guid>https://blog.youkuaiyun.com/devascend/article/details/156310186</guid><author>devascend</author><pubDate>Fri, 26 Dec 2025 17:05:27 +0800</pubDate><description><![CDATA[大规模训练集群性能问题（下降或抖动）分享及性能问题解决方案]]></description><category></category></item><item><title><![CDATA[超节点FSDP2训练MoE大模型昇腾联创设计（下）]]></title><link>https://blog.youkuaiyun.com/devascend/article/details/156201030</link><guid>https://blog.youkuaiyun.com/devascend/article/details/156201030</guid><author>devascend</author><pubDate>Wed, 24 Dec 2025 14:26:48 +0800</pubDate><description><![CDATA[为了进一步挖掘 XTuner V1 训练方案的上限，实验室研究团队与华为昇腾技术团队在 Ascend A3 超节点上进行联合优化，充分利用超节点硬件特性，FSDP2首次在Qwen 235B MoE上实现了相比传统3D并行更高的 MFU（Model FLOPS Utilization，模型浮点运算利用率）。在理论算力落后 NVIDIA H800 近 20% 的情况下，最终实现训练吞吐超过 H800 近 5%，MFU 反超 20% 以上。]]></description><category></category></item><item><title><![CDATA[超节点FSDP2训练MoE大模型昇腾联创设计（上）]]></title><link>https://blog.youkuaiyun.com/devascend/article/details/156196412</link><guid>https://blog.youkuaiyun.com/devascend/article/details/156196412</guid><author>devascend</author><pubDate>Wed, 24 Dec 2025 14:26:23 +0800</pubDate><description><![CDATA[2025年9月8日，上海人工智能实验室开源书生大模型新一代训练引擎XTuner V1。XTuner V1 是伴随上海AI实验室“通专融合”技术路线的持续演进，以及书生大模型研发实践而成长起来的新一代训练引擎。相较于传统的 3D 并行训练引擎，XTuner V1 不仅能应对更加复杂的训练场景，还具备更快的训练速度，尤其在超大规模稀疏混合专家（MoE）模型训练中优势显著。]]></description><category></category></item></channel></rss>