<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[sslfk的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/sslfk</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; sslfk]]></copyright><item><title><![CDATA[Python｜flash_attn 安装方法]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/154834637</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/154834637</guid><author>sslfk</author><pubDate>Fri, 14 Nov 2025 13:46:46 +0800</pubDate><description><![CDATA[对于flash_attn对应的cxx11abiFALSE还是cxx11abiTRUE，使用如下方式判断。]]></description><category></category></item><item><title><![CDATA[大模型调用常用参数解析]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/153784608</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/153784608</guid><author>sslfk</author><pubDate>Thu, 23 Oct 2025 16:41:00 +0800</pubDate><description><![CDATA[frequency_penalty的值。]]></description><category></category></item><item><title><![CDATA[大模型训练-训练数据量与训练周期（epoch）的关系]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/153313543</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/153313543</guid><author>sslfk</author><pubDate>Wed, 15 Oct 2025 09:53:53 +0800</pubDate><description><![CDATA[本文总结了不同数据量下的训练周期推荐。对于1k-5k的小数据集，建议1-3个epoch（全参）或3-10个（LoRA），需采用早停机制防过拟合。5k-20k中等规模数据可适当增加epoch至2-5（全参）或5-10（LoRA）。超过20k的大数据集，1-3个epoch（全参）或2-5个（LoRA）即可，避免过拟合且迭代收益有限。核心原则是根据数据规模调整训练轮次，大数据集快速迭代，小数据集谨慎训练。]]></description><category></category></item><item><title><![CDATA[基于llama_factory的qwen3全参微调]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/152268280</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/152268280</guid><author>sslfk</author><pubDate>Mon, 29 Sep 2025 14:26:28 +0800</pubDate><description><![CDATA[【代码】基于llama_factory的qwen3全参微调。]]></description><category></category></item><item><title><![CDATA[基于autoawq进行qwen3 的awq量化]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/151425579</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/151425579</guid><author>sslfk</author><pubDate>Wed, 10 Sep 2025 20:40:23 +0800</pubDate><description><![CDATA[awq量化 精度降低6个点。推理耗时降低从0.447s降低到0.4s。在llamafactory环境中，安装。]]></description><category></category></item><item><title><![CDATA[大模型量化方法及qwen3的AutoGPTQ量化]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/151189845</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/151189845</guid><author>sslfk</author><pubDate>Thu, 04 Sep 2025 17:08:33 +0800</pubDate><description><![CDATA[处理完后的 msg 应该是[{"role":"user","content":""},{"role":"assistant","content":""}] 类似的，可以让chat_template解析，同时带有回复答案的。data_path = "模型训练数据路径，注意格式"基于llama_factory的训练环境，安装量化,quant_path = "模型量化保存路径"model_path = "模型路径"]]></description><category></category></item><item><title><![CDATA[大模型时代个人提效+创新工具集锦]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/147134943</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/147134943</guid><author>sslfk</author><pubDate>Fri, 11 Apr 2025 09:35:25 +0800</pubDate><description><![CDATA[基于大模型的应用提效方法]]></description><category></category></item><item><title><![CDATA[elasticsearch异步查询]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/145701337</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/145701337</guid><author>sslfk</author><pubDate>Tue, 18 Feb 2025 11:18:41 +0800</pubDate><description><![CDATA[elasticsearch异步检索]]></description><category></category></item><item><title><![CDATA[vllm的部署和使用]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/144848681</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/144848681</guid><author>sslfk</author><pubDate>Tue, 31 Dec 2024 17:41:52 +0800</pubDate><description><![CDATA[qwen2.5模型下载。]]></description><category></category></item><item><title><![CDATA[ONNX Runtime gpu版本安装]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/144836450</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/144836450</guid><author>sslfk</author><pubDate>Mon, 30 Dec 2024 23:20:25 +0800</pubDate><description><![CDATA[ONNX Runtime版本与cudatoolkit版本对应关系：NVIDIA - CUDA | onnxruntimeonnx runtime发的版本：Releases · microsoft/onnxruntimeonnx runtime 官网：ONNX Runtime | Homeonnx和onnx runtime版本对应关系：Compatibility | onnxruntime]]></description><category></category></item><item><title><![CDATA[python 服务的docker镜像制作和启动]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/144243574</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/144243574</guid><author>sslfk</author><pubDate>Wed, 04 Dec 2024 16:32:08 +0800</pubDate><description><![CDATA[dockerfile 制作。]]></description><category></category></item><item><title><![CDATA[milvus和attu客户端的安装]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/143893889</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/143893889</guid><author>sslfk</author><pubDate>Tue, 19 Nov 2024 19:52:14 +0800</pubDate><description><![CDATA[milvus的安装（待验证）：attu客户端的安装：attu的容器启动：]]></description><category></category></item><item><title><![CDATA[基于bce-embedding-base_v1模型将torch模型转换为onnx模型]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/143674304</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/143674304</guid><author>sslfk</author><pubDate>Mon, 11 Nov 2024 10:23:00 +0800</pubDate><description><![CDATA[确定cuda和onnxruntime的版本对应关系：https://blog.51cto.com/u_15962038/12360310。]]></description><category></category></item><item><title><![CDATA[大模型prompt相关]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/142494437</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/142494437</guid><author>sslfk</author><pubDate>Tue, 24 Sep 2024 18:54:28 +0800</pubDate><description><![CDATA[n\n从给定的文本中，我们可以抽取以下知识图谱三元组：\n\n```json\n[\n    {\"subject\": \"刘德华\", \"predicate\": \"老婆\", \"object\": \"朱丽倩\"}\n]```\n\n文本输入为：北京是中国的首都\n\n从给定的文本中，我们可以抽取以下知识图谱三元组：\n\n。\n\n文本输入：\n\n我很烦，今天的事情太多了。\n\n分类为：\n\n消极\n\n文本输入：\n\n我今天感觉很开心。\n\n分类为：\n\n。]]></description><category></category></item><item><title><![CDATA[typora使用和激活]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/142459240</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/142459240</guid><author>sslfk</author><pubDate>Mon, 23 Sep 2024 15:16:53 +0800</pubDate><description><![CDATA[此时打开就会发现提示已经激活成功，但是每次打开都会提示这个激活，并且左下角会提示未激活。右键——打开方式——使用记事本打开。]]></description><category></category></item><item><title><![CDATA[python 单例模式实现]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/142366622</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/142366622</guid><author>sslfk</author><pubDate>Thu, 19 Sep 2024 18:02:18 +0800</pubDate><description><![CDATA[【代码】python 单例模式实现。]]></description><category></category></item><item><title><![CDATA[gradio 组件相关技巧]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/141612065</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/141612065</guid><author>sslfk</author><pubDate>Tue, 27 Aug 2024 20:47:18 +0800</pubDate><description><![CDATA[【代码】gradio 组件相关技巧。]]></description><category></category></item><item><title><![CDATA[langchain 基于ES的数据向量化存储和检索]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/138585799</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/138585799</guid><author>sslfk</author><pubDate>Wed, 08 May 2024 21:06:45 +0800</pubDate><description><![CDATA[3、多语言模型：BAAI/bge-m3 支持的输入长度]]></description><category></category></item><item><title><![CDATA[elasticsearch 关于向量化检索]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/137925767</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/137925767</guid><author>sslfk</author><pubDate>Thu, 18 Apr 2024 16:54:50 +0800</pubDate><description><![CDATA[向量检索，RAG]]></description><category></category></item><item><title><![CDATA[语义相似性计算]]></title><link>https://blog.youkuaiyun.com/sslfk/article/details/137512584</link><guid>https://blog.youkuaiyun.com/sslfk/article/details/137512584</guid><author>sslfk</author><pubDate>Mon, 08 Apr 2024 16:38:07 +0800</pubDate><description><![CDATA[语义匹配]]></description><category></category></item></channel></rss>