<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[大猴子]]></title><description><![CDATA[寻梦不怕路远...]]></description><link>https://blog.youkuaiyun.com/lglglgl</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; lglglgl]]></copyright><item><title><![CDATA[11 | 海光 BW 异构加速卡 vLLM 推理引擎部署与性能评测]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/169066507</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/169066507</guid><author>lglglgl</author><pubDate>Sun, 25 Oct 2026 23:25:49 +0800</pubDate><description><![CDATA[海光 BW 异构加速卡 vLLM 推理引擎部署与性能评测报告，基于国家超算互联网 SCNet 平台（2 × 海光 BW，DTK 26.04，vLLM 0.11.0]]></description><category></category></item><item><title><![CDATA[10 | vLLM 定位与核心机制详解]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/168146935</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/168146935</guid><author>lglglgl</author><pubDate>Sat, 17 Oct 2026 06:01:37 +0800</pubDate><description><![CDATA[PagedAttention 主要是将 KV Cache 切分为固定大小的 Block 再按需分配，用非连续存储替代整块预留，消除碎片与过度占用，让同等显存可支撑更多并发请求。]]></description><category></category></item><item><title><![CDATA[09 | LangGraph 构建个人知识库对比 Qwen14B & GLM 效果]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/168042258</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/168042258</guid><author>lglglgl</author><pubDate>Fri, 16 Oct 2026 03:20:54 +0800</pubDate><description><![CDATA[用 LangGraph 搭建 RAG 知识库，以多版本简历为数据集，对比本地 Qwen 14B 与云端 GLM 的问答效果。两者检索层相同，GLM 引用更完整、结构更清晰，Qwen 胜在本地无网络依赖。]]></description><category></category></item><item><title><![CDATA[08 | 本地 RAG 模型对比Qwen 与 DeepSeek-R1-Distill（7B / 14B)]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/167891961</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/167891961</guid><author>lglglgl</author><pubDate>Wed, 14 Oct 2026 11:45:09 +0800</pubDate><description><![CDATA[在 MacBook Pro M4 环境下，模型从 7B 扩大到 14B，Qwen 和 DeepSeek 的平均生成耗时分别增加 2.54 倍和 2.18 倍。同等参数规模下，Qwen 的端到端响应速度约为 DeepSeek 的 4 倍，主要原因之一是 DeepSeek 输出更长。综合性能表现，Qwen 7B 更适合低延迟场景，Qwen 14B 则需要付出更高的推理成本；后续应针对小模型优化检索开销，针对大模型优化生成延迟与输出长度。]]></description><category></category></item><item><title><![CDATA[07 | FastAPI + llama.cpp 接入 Langfuse监控Prompt请求]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/167792992</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/167792992</guid><author>lglglgl</author><pubDate>Tue, 13 Oct 2026 11:40:30 +0800</pubDate><description><![CDATA[FastAPI + llama.cpp 服务接入 Langfuse Cloud,记录每次请求的 Trace 与 Token 用量,并用并发压测的 Prometheus 对账结果核对追踪是否完整。]]></description><category></category></item><item><title><![CDATA[06 | Prometheus打通FastAPI云原生数据采集]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/167704361</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/167704361</guid><author>lglglgl</author><pubDate>Tue, 13 Oct 2026 00:11:43 +0800</pubDate><description><![CDATA[Prometheus获取FastAPI云原生数据采集，FastAPI 在 Mac 的 Docker 里,Prometheus 在 Kubernetes 里。用手动 Endpoints 加 ServiceMonitor 打通采集,并验证成功。]]></description><category></category></item><item><title><![CDATA[05 | Jenkins 搭建 CD 流水线（部署到 Kubernetes、Trivy 漏洞扫描与推理冒烟测试）]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/167689217</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/167689217</guid><author>lglglgl</author><pubDate>Mon, 12 Oct 2026 13:36:23 +0800</pubDate><description><![CDATA[用 Jenkins 为 llm-serving 搭建 CD 流水线:按 IMAGE_TAG 通过 Helm 部署 FastAPI 与 llama.cpp,等待 rollout 就绪后,并行执行 Harbor Trivy 扫描和五项冒烟测试。第 #3 次构建 0 Critical,测试全部通过,约 50 秒 SUCCESS。]]></description><category></category></item><item><title><![CDATA[04 | 用 Jenkins 打通 CI自动构建镜像并推送到本地 Harbor]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/167590391</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/167590391</guid><author>lglglgl</author><pubDate>Mon, 12 Oct 2026 01:40:12 +0800</pubDate><description><![CDATA[在 Jenkins 里配置 Harbor 凭据,用 Pipeline script from SCM 从 GitHub 拉取代码,自动构建 llm-serving 和 llm-serving-llama 两个镜像,以构建号和 latest 双标签推送到本地 Harbor,并用 digest 一致来验收。]]></description><category></category></item><item><title><![CDATA[03 | 更新：Ollama 换成 llama.cpp构建更小的离线镜像]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/167586636</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/167586636</guid><author>lglglgl</author><pubDate>Sun, 11 Oct 2026 17:28:25 +0800</pubDate><description><![CDATA[用 llama.cpp 官方 server 镜像加上 Qwen_0.5b_Q4_K_M 量化的 GGUF 模型,打包出离线镜像 llm-serving-llama:0.1.0,取代体积庞大的 Ollama 镜像,同时保持模型名和 OpenAI 兼容接口不变。]]></description><category></category></item><item><title><![CDATA[Github | 初始化空白仓库到首次推送]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/167581051</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/167581051</guid><author>lglglgl</author><pubDate>Sun, 11 Oct 2026 13:29:22 +0800</pubDate><description><![CDATA[把本地项目变成一个 Git 仓库,并推送到 GitHub 上已经创建好的空白仓库。]]></description><category></category></item><item><title><![CDATA[03 | 离线部署 FastAPI + Ollama(qwen2:0.5b):从 Docker 联调到 Kubernetes 验收]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/167398721</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/167398721</guid><author>lglglgl</author><pubDate>Sat, 10 Oct 2026 11:54:21 +0800</pubDate><description><![CDATA[把一个本地 LLM 推理服务拆成两个独立组件FastAPI 服务和Ollama 服务,并最终跑在 Kubernetes 上]]></description><category></category></item><item><title><![CDATA[02 | Kubernetes 部署 FastAPI LLM Serving 并接入 Ollama]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/167284877</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/167284877</guid><author>lglglgl</author><pubDate>Fri, 09 Oct 2026 01:25:12 +0800</pubDate><description><![CDATA[本文基于Kubernetes部署FastAPI LLM服务，通过Helm安装并配置Service与Ingress，实现外部访问。FastAPI容器通过host.docker.internal调用MacBook本地Ollama服务，使用Qwen3:8b模型。测试验证了Pod内健康检查、环境变量、Ollama连接及网络链路畅通，完整打通从客户端到LLM推理的调用路径。]]></description><category></category></item><item><title><![CDATA[01|LLM Serving：FastAPI + Ollama 的 LLM 推理服务与性能测试]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/167279112</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/167279112</guid><author>lglglgl</author><pubDate>Thu, 08 Oct 2026 18:30:23 +0800</pubDate><description><![CDATA[基于MacBook Pro M4+24G，验证了LLM服务端完整链路：通过Ollama加载Qwen、Llama等模型，FastAPI实现OpenAI兼容API，支持流式推理与健康检查。测试显示5次请求成功率100%，平均延迟17.32秒，性能稳定。后续计划集成Prometheus、Grafana与OpenTelemetry，监控TTFT、P95、Tokens/s等关键指标，提升可观测性。]]></description><category></category></item><item><title><![CDATA[【记录】MAC本地微调大模型（MLX + Qwen2.5）并利用Ollama接入项目实战]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/153418419</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/153418419</guid><author>lglglgl</author><pubDate>Thu, 16 Oct 2025 21:18:17 +0800</pubDate><description><![CDATA[参考和感谢：https://juejin.cn/post/7499137821425549363(qwen05b) (base) igwanglyang@MacbookProdahouzicn Qwen2.5-Sex % source venv/bin/activate(venv) (base) igwanglyang@MacbookProdahouzicn Qwen2.5-Sex % pip3 install huggingface_hub mlx-lm transformers torch numpy ]]></description><category></category></item><item><title><![CDATA[AWS EC2设置用户名密码登录]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/144062135</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/144062135</guid><author>lglglgl</author><pubDate>Tue, 26 Nov 2024 16:38:26 +0800</pubDate><description><![CDATA[通过用户名和密码登录到AWS EC2实例。请确保在生产环境中使用安全的密码，并考虑使用密钥对认证以提高安全性。]]></description><category></category></item><item><title><![CDATA[Rabin加解密算法(python3)]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/134479714</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/134479714</guid><author>lglglgl</author><pubDate>Sat, 18 Nov 2023 16:20:50 +0800</pubDate><description><![CDATA[【代码】Rabin加解密算法(python3)]]></description><category></category></item><item><title><![CDATA[USM选课（20230912）]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/133169443</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/133169443</guid><author>lglglgl</author><pubDate>Fri, 22 Sep 2023 14:20:50 +0800</pubDate><description><![CDATA[#
Course Code
Course Name
课程
Semester
Type
Unit
Method
Date/Time




1
CCS528&nbsp;Confirmed
INFORMATION SECURITY AND CRYPTOGRAPHY
信息安全和密码学
231
ELECTIVE (E)
4
WEB
12/09/2023 18:53 PM


2
CCS591&nbsp;Confirmed
RESEARCH AND EMPIRICAL METHODS IN COMPUTER ]]></description><category></category></item><item><title><![CDATA[centos7 默认路由顺序调整（IPV4_ROUTE_METRIC）]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/132059328</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/132059328</guid><author>lglglgl</author><pubDate>Wed, 02 Aug 2023 12:24:45 +0800</pubDate><description><![CDATA[将公网的网卡路由优先级设计为最高即可。]]></description><category></category></item><item><title><![CDATA[[异常处理]mysql8 SELinux is preventing from name_bind]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/131781587</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/131781587</guid><author>lglglgl</author><pubDate>Tue, 18 Jul 2023 10:33:45 +0800</pubDate><description><![CDATA[my.cnf 配置lower_case_table_names=1导致无法启动。SELinux安全机制禁止了mysql的其他端口绑定，导致无法通信。]]></description><category></category></item><item><title><![CDATA[MySQL两主一从部署（keepalived）]]></title><link>https://blog.youkuaiyun.com/lglglgl/article/details/131287956</link><guid>https://blog.youkuaiyun.com/lglglgl/article/details/131287956</guid><author>lglglgl</author><pubDate>Mon, 19 Jun 2023 14:57:58 +0800</pubDate><description><![CDATA[操作系统：Centos 7.9数据库：MySQL 5.7.41。+keepalived]]></description><category></category></item></channel></rss>