<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[feixiang_9的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/feixiang_9</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; feixiang_9]]></copyright><item><title><![CDATA[主流NVIDIA算力卡对比]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/170338806</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/170338806</guid><author>feixiang_9</author><pubDate>Mon, 09 Nov 2026 10:04:16 +0800</pubDate><description><![CDATA[数据来源：NVIDIA 官方规格页与数据中心资料（H200 产品页、HGX H100/H200/B200 参考架构、Blackwell Ultra Datasheet、HGX/GB300 规格页）。：表中 Tensor Core 数值；NVIDIA 官方常公布 sparse（稀疏）数值，sparse 通常是 dense 的 2 倍（FP4 一代为 1.5 倍关系，见备注）。]]></description><category></category></item><item><title><![CDATA[RTX 4090 24GB 部署 MiniMax-H3 实战：SGLang + INT8 量化，50 秒生成带音频视频]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/170191545</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/170191545</guid><author>feixiang_9</author><pubDate>Fri, 06 Nov 2026 12:05:04 +0800</pubDate><description><![CDATA[回头看，这次部署真正难的不是跑通，而是"在 24GB 显存 + 79GB 磁盘的夹缝里做减法"磁盘约束逼出了预量化单文件方案（144GB → 48GB）；显存约束逼出了分层卸载（52GB → 18GB）；版本约束则贡献了本文一半的踩坑记录。先算账再动手：把每个组件的体积、显存、内存列出来，方案基本就唯一了；预量化权重 + 覆盖加载是最省事的路子，但一定要确认 embedding 是 BF16；分层卸载只卸载 DiT 和文本编码器，VAE 常驻 GPU，性能差别巨大；从 0 开始加。]]></description><category></category></item><item><title><![CDATA[Jev：不写一个字、只做判断的 TypeSafe System One 模型]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/166361987</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/166361987</guid><author>feixiang_9</author><pubDate>Tue, 22 Sep 2026 14:06:37 +0800</pubDate><description><![CDATA[obsidian。]]></description><category></category></item><item><title><![CDATA[Laya：不做生成、只做决策的开源非自回归模型]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/166360961</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/166360961</guid><author>feixiang_9</author><pubDate>Tue, 22 Sep 2026 13:45:26 +0800</pubDate><description><![CDATA[它由 Convai Innovations 发布，被广泛视为 TypeSafe Jev 的开源平替。obsidian。]]></description><category></category></item><item><title><![CDATA[基于 vLLM + RTX 4090 部署 Qwen3.8-27B-W4A16-AWQ 实战指南]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/166251397</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/166251397</guid><author>feixiang_9</author><pubDate>Mon, 21 Sep 2026 17:31:21 +0800</pubDate><description><![CDATA[在单卡 RTX 4090（24GB）上部署 27B 参数的大模型，听起来像是不可能完成的任务？通过 AWQ 4-bit 量化 + vLLM 推理框架，我们不仅实现了，还获得了相当不错的性能。本文将完整记录在环境下，部署阿里多模态模型的全过程，包括踩过的坑和解决方案。要求说明必须 PyTorch 2.13+cu130vLLM 0.28 硬编码要求必须驱动 ≥ R580cu130 ABI 需要必须 transformers 5.xQwen3.5 架构在 4.x 不识别。]]></description><category></category></item><item><title><![CDATA[基于sglang+4090 部署Qwen3.8-27B-W4A16-AWQ ]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/164370594</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/164370594</guid><author>feixiang_9</author><pubDate>Fri, 04 Sep 2026 15:57:22 +0800</pubDate><description><![CDATA[基于sglang+4090 部署Qwen3.8-27B-W4A16-AWQ]]></description><category></category></item><item><title><![CDATA[GPU细粒度资源虚拟化方案说明]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/164085905</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/164085905</guid><author>feixiang_9</author><pubDate>Thu, 27 Aug 2026 13:58:27 +0800</pubDate><description><![CDATA[对AI业务应用进行容器化改造后，可采用GPU虚拟化技术实现GPU算力、显存的细粒度分配，解决Kubernetes原生仅支持整卡分配GPU的痛点，既支持多个业务应用共享单张GPU资源，也支持单个业务应用聚合多张物理GPU的切片资源，按需灵活分配算力与显存，有效提升GPU集群整体资源利用率。]]></description><category></category></item><item><title><![CDATA[英伟达国内主流数据中心GPU总表]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/164085312</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/164085312</guid><author>feixiang_9</author><pubDate>Wed, 26 Aug 2026 12:42:18 +0800</pubDate><description><![CDATA[不适合现代大模型推理；FP64双精度性能较强，多用于传统HPC科学计算，新项目不建议采购。补充说明：P100属于Pascal架构，]]></description><category></category></item><item><title><![CDATA[MIG / vGPU时间切片 / MPS 三者对比表]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/164085135</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/164085135</guid><author>feixiang_9</author><pubDate>Wed, 26 Aug 2026 12:34:09 +0800</pubDate><description><![CDATA[如果你需要，我可以把前面全部GPU型号连同MIG支持情况合并成一张完整总表。]]></description><category></category></item><item><title><![CDATA[一张图理解kafka的原理]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/164019043</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/164019043</guid><author>feixiang_9</author><pubDate>Mon, 24 Aug 2026 11:41:04 +0800</pubDate><description><![CDATA[]]></description><category></category></item><item><title><![CDATA[【算力平台多维分类表】]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/163731956</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/163731956</guid><author>feixiang_9</author><pubDate>Thu, 13 Aug 2026 18:24:02 +0800</pubDate><description><![CDATA[]]></description><category></category></item><item><title><![CDATA[【k8s应用-vSphere CSI Driver on Kubernete】]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/163701249</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/163701249</guid><author>feixiang_9</author><pubDate>Wed, 12 Aug 2026 16:42:15 +0800</pubDate><description><![CDATA[确认 Kubernetes、vCenter 和 ESXi 环境满足版本要求：存储：VMware Datastore（SSD/HDD 均可）本方案使用 vSphere CSI Driver 将 VMware 块存储接入 Kubernetes，通过 CSI 标准接口提供 PVC 持久化存储。]]></description><category></category></item><item><title><![CDATA[k8s基础-pod创建过程]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/161948800</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/161948800</guid><author>feixiang_9</author><pubDate>Sat, 13 Jun 2026 11:11:34 +0800</pubDate><description><![CDATA[阶段控制平面组件工作平面组件交互方式接收 Pod 定义客户端 → API Server调度 PodScheduler存储挂载kubelet, CSI 插件创建 Sandboxkubelet 调用 CRI/CNI启动容器kubelet 调用 CRI探针检查kubelet服务发现网络规则kube-proxyDNS 解析CoreDNS（控制平面 Pod）]]></description><category></category></item><item><title><![CDATA[Calico网络架构图 & 跨主机通信原理]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/161904359</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/161904359</guid><author>feixiang_9</author><pubDate>Thu, 11 Jun 2026 23:56:44 +0800</pubDate><description><![CDATA[]]></description><category></category></item><item><title><![CDATA[k8s基础原理-pod的生命周期流程]]></title><link>https://blog.youkuaiyun.com/feixiang_9/article/details/161898546</link><guid>https://blog.youkuaiyun.com/feixiang_9/article/details/161898546</guid><author>feixiang_9</author><pubDate>Thu, 11 Jun 2026 16:55:42 +0800</pubDate><description><![CDATA[]]></description><category></category></item></channel></rss>