<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[个人小站：www.xiaowangyun.com]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/scc1371815174</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; scc1371815174]]></copyright><item><title><![CDATA[昇思25天学习打卡Day01]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/139945924</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/139945924</guid><author>scc1371815174</author><pubDate>Tue, 25 Jun 2024 08:57:20 +0800</pubDate><description><![CDATA[昇思25天学习打卡]]></description><category></category></item><item><title><![CDATA[昇思25天学习打卡]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/139818924</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/139818924</guid><author>scc1371815174</author><pubDate>Thu, 20 Jun 2024 08:59:40 +0800</pubDate><description><![CDATA[昇思25天学习打卡。]]></description><category></category></item><item><title><![CDATA[MindSpore实践图神经网络之GCN]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/139335634</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/139335634</guid><author>scc1371815174</author><pubDate>Thu, 30 May 2024 21:57:47 +0800</pubDate><description><![CDATA[图卷积网络（GCN）于2016年提出，旨在对图结构数据进行半监督学习。它提出了一种基于卷积神经网络有效变体的可扩展方法，可直接在图上操作。该模型在图边缘的数量上线性缩放，并学习隐藏层表示，这些表示编码了局部图结构和节点特征。GCN（图卷积神经网络) 类似CNN（卷积神经网络），只不过CNN用于二维数据结构，GCN用于图数据结构。GCN实际上跟CNN的作用一样，就是一个特征提取器，只不过它的对象是图数据。GCN精妙地设计了一种从图数据中提取特征的方法。GCN包含两个图卷积层。]]></description><category></category></item><item><title><![CDATA[MindSpore实践图神经网络之图数据]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/139335597</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/139335597</guid><author>scc1371815174</author><pubDate>Thu, 30 May 2024 21:56:55 +0800</pubDate><description><![CDATA[通常一个图（graph)G是由一系列的节点(vertices)V以及边（eage）E，简单的图如下所示。图中包含节点V = {a, b, c, d}，和边E = {(a, b), (b, c), (c, d), (d, b)}，针对图中的连接关系通常需借助数学的方式进行描述，如常用的基于邻接矩阵的方式，用于描述上述图连接关系的矩阵C如下，其中a、 b、c、d对应为第1、2、 3、4个节点。]]></description><category></category></item><item><title><![CDATA[MindSpore实践图神经网络之环境篇]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/139307849</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/139307849</guid><author>scc1371815174</author><pubDate>Wed, 29 May 2024 22:52:38 +0800</pubDate><description><![CDATA[MindSpore在Windows11系统下的环境配置。MindSpore环境配置大概分为三步：（1）安装Python环境，（2）安装MindSpore，（3）验证是否成功如果是GPU环境还需安装CUDA等环境，可以参照官网 https://www.mindspore.cn/install接下来介绍几款开发IDE以及辅助工具，此处省略了上一节Minicanda安装Python环境，没有安装可以参照：【MindSpore易点通】图神经网络实践系列01开篇。]]></description><category></category></item><item><title><![CDATA[MindSpore实现图神经网络之理论篇]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/139307807</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/139307807</guid><author>scc1371815174</author><pubDate>Wed, 29 May 2024 22:50:42 +0800</pubDate><description><![CDATA[但是，现实世界中存在许多复杂系统，这些系统常常表现为复杂网络的形式，如社交网络，交通网络，蛋白质交互网络，由于这些复杂网络的不规则性，传统的神经网络很难直接应用于图数据。消息传递机制指的是要更新某一个点的特征，需要考虑周边和它相连的点的特征，这里就要用到消息函数。然后用另一个网络整合周边节点信息，接着加和到这个节点本身的特征中，这两步更新模式就构成了一个基本的消息传递机制。图神经网络的每一个特征提取层都会对节点的特征和连接的特征进行更新，这种更新模式被叫作消息传递机制。]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day06智能体]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/138292483</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/138292483</guid><author>scc1371815174</author><pubDate>Sun, 28 Apr 2024 21:57:24 +0800</pubDate><description><![CDATA[可以感知环境中的动态条件能采取动作影响环境能运用推理能力。]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day05作业]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/138013186</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/138013186</guid><author>scc1371815174</author><pubDate>Sun, 21 Apr 2024 09:10:22 +0800</pubDate><description><![CDATA[以API Server方式启动 lmdeploy，开启 W4A16量化，调整KV Cache的占用比例为0.4，分别使用命令行客户端与Gradio网页客户端与模型对话。使用W4A16量化，调整KV Cache的占用比例为0.4，使用Python代码集成的方式运行internlm2-chat-1.8b模型。设置KV Cache最大占用比例为0.4，开启W4A16量化，以命令行方式与模型对话。使用 LMDeploy 运行视觉多模态大模型 llava gradio demo。]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day05LMDeploy 高阶]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/138013037</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/138013037</guid><author>scc1371815174</author><pubDate>Sun, 21 Apr 2024 09:02:37 +0800</pubDate><description><![CDATA[llava多模态模型，使用pipeline推理]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day05LMDeploy服务]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/138012796</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/138012796</guid><author>scc1371815174</author><pubDate>Sun, 21 Apr 2024 09:01:03 +0800</pubDate><description><![CDATA[将大模型封装为API接口服务，供客户端访问]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day05LMDeploy实践]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/138012115</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/138012115</guid><author>scc1371815174</author><pubDate>Sat, 20 Apr 2024 21:29:47 +0800</pubDate><description><![CDATA[选择镜像；选择10% A100*1GPU；点击“立即创建”。]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day05 LMDeploy 量化部署]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/138011987</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/138011987</guid><author>scc1371815174</author><pubDate>Sat, 20 Apr 2024 21:23:29 +0800</pubDate><description><![CDATA[书生·浦语大模型实战营Day05 LMDeploy 量化部署

LMDeploy 量化部署 LLM-VLM

模型部署

部署：将开发完毕的软件投入使用的过程
服务器：CPU、GPU


大模型部署面临的挑战

计算量巨大、参数规模大



内存开销巨大：K V Cache



访存瓶颈、动态请求


模型部署方法
模型剪枝（Pruning）

非结构化剪枝
结构化剪枝


知识蒸馏（Knoledge Distillation，KD）

教师-学生


量化（Quantization）

量化感知训练（Q]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营合集]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/137891180</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/137891180</guid><author>scc1371815174</author><pubDate>Thu, 18 Apr 2024 22:49:54 +0800</pubDate><description><![CDATA[书生·浦语大模型实战营]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day04作业]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/137891857</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/137891857</guid><author>scc1371815174</author><pubDate>Thu, 18 Apr 2024 22:48:16 +0800</pubDate><description><![CDATA[记录复现过程并截图https://openxlab.org.cn/models/detail/sunxiaobei/personal_assistant。]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day04OpenXLab 部署]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/137891263</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/137891263</guid><author>scc1371815174</author><pubDate>Wed, 17 Apr 2024 22:08:00 +0800</pubDate><description><![CDATA[书生·浦语大模型实战营Day04OpenXLab 部署]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day04个人助手微调]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/137844775</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/137844775</guid><author>scc1371815174</author><pubDate>Wed, 17 Apr 2024 21:54:41 +0800</pubDate><description><![CDATA[用 `QLoRA` 的方式来微调一个自己的小助手]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day04微调]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/137844460</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/137844460</guid><author>scc1371815174</author><pubDate>Tue, 16 Apr 2024 23:20:58 +0800</pubDate><description><![CDATA[书生·浦语大模型实战营Day04微调]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day04XTuner多模态]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/137844335</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/137844335</guid><author>scc1371815174</author><pubDate>Tue, 16 Apr 2024 23:12:44 +0800</pubDate><description><![CDATA[书生·浦语大模型实战营Day04XTuner多模态]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day03作业]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/137798300</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/137798300</guid><author>scc1371815174</author><pubDate>Mon, 15 Apr 2024 22:33:07 +0800</pubDate><description><![CDATA[书生·浦语大模型实战营Day03茴香豆作业]]></description><category></category></item><item><title><![CDATA[书生·浦语大模型实战营Day03茴香豆]]></title><link>https://blog.youkuaiyun.com/scc1371815174/article/details/137797954</link><guid>https://blog.youkuaiyun.com/scc1371815174/article/details/137797954</guid><author>scc1371815174</author><pubDate>Mon, 15 Apr 2024 22:28:06 +0800</pubDate><description><![CDATA[书生·浦语大模型实战营Day03茴香豆：搭建你的 RAG 智能助理]]></description><category></category></item></channel></rss>