https://guijiagi.com/categories/%E6%8A%80%E6%9C%AF%E5%B7%A5%E5%85%B7/Recent content in 技术工具 on AI 实战派 · 从技术到赚钱Hugozh-cn本站内容采用 CC BY-NC-SA 4.0 国际许可协议授权Fri, 23 Oct 2026 11:08:00 +0800https://guijiagi.com/posts/2026-10-23-68-ai-content-pipeline-scripts/Fri, 23 Oct 2026 11:08:00 +0800https://guijiagi.com/posts/2026-10-23-68-ai-content-pipeline-scripts/日更百篇的内容站，靠人力写写不动，靠脚本才转得起来。https://guijiagi.com/posts/2026-10-23-67-lora-finetune-beginners/Fri, 23 Oct 2026 11:07:00 +0800https://guijiagi.com/posts/2026-10-23-67-lora-finetune-beginners/LoRA微调是让通用模型学会你自己风格或领域知识的捷径，入门门槛比全量微调低得多，但坑也密集。https://guijiagi.com/posts/2026-10-23-66-vector-database-selection/Fri, 23 Oct 2026 11:06:00 +0800https://guijiagi.com/posts/2026-10-23-66-vector-database-selection/向量数据库选型，新手容易陷入选最强的陷阱。https://guijiagi.com/posts/2026-10-23-65-rag-retrieval-practice/Fri, 23 Oct 2026 11:05:00 +0800https://guijiagi.com/posts/2026-10-23-65-rag-retrieval-practice/RAG项目做不好，九成问题不在大模型，在检索。https://guijiagi.com/posts/2026-10-23-64-prompt-engineering-2026/Fri, 23 Oct 2026 11:04:00 +0800https://guijiagi.com/posts/2026-10-23-64-prompt-engineering-2026/提示词工程在2026年已经不是写一句精妙咒语的手艺活，而是系统工程。https://guijiagi.com/posts/2026-10-23-63-api-cost-optimization-tips/Fri, 23 Oct 2026 11:03:00 +0800https://guijiagi.com/posts/2026-10-23-63-api-cost-optimization-tips/AI应用跑起来后，最贵的不是开发，是API账单。https://guijiagi.com/posts/2026-10-23-62-ffmpeg-ai-video-batch/Fri, 23 Oct 2026 11:02:00 +0800https://guijiagi.com/posts/2026-10-23-62-ffmpeg-ai-video-batch/短视频批量生产里，最耗时间的不是剪片，是重复的格式转换、压字幕、抽帧、统一片尾。https://guijiagi.com/posts/2026-10-23-61-ollama-lmstudio-tool-compare/Fri, 23 Oct 2026 11:01:00 +0800https://guijiagi.com/posts/2026-10-23-61-ollama-lmstudio-tool-compare/本地跑大模型，新手最纠结的就是Ollama和LMStudio选哪个。https://guijiagi.com/posts/2026-10-23-60-local-llm-deploy-64g-guide/Fri, 23 Oct 2026 11:00:00 +0800https://guijiagi.com/posts/2026-10-23-60-local-llm-deploy-64g-guide/六十四G统一内存（或显存）跑本地大模型，属于能玩得很舒服的档位。https://guijiagi.com/posts/2026-10-23-59-comfyui-workflow-optimization/Fri, 23 Oct 2026 10:59:00 +0800https://guijiagi.com/posts/2026-10-23-59-comfyui-workflow-optimization/ComfyUI出图慢、爆显存，八成不是显卡不行，是工作流没调对。https://guijiagi.com/posts/2026-10-22-09-ai-pc-build-budget-config/Thu, 22 Oct 2026 12:54:00 +0800https://guijiagi.com/posts/2026-10-22-09-ai-pc-build-budget-config/把预算和用途告诉 AI 出配件清单、解释搭配逻辑、提醒兼容性问题的方法https://guijiagi.com/posts/2026-10-22-07-ai-zettelkasten-notes/Thu, 22 Oct 2026 12:42:00 +0800https://guijiagi.com/posts/2026-10-22-07-ai-zettelkasten-notes/把大段笔记拆成独立卡片、AI 帮忙提炼观点和建立关联的方法https://guijiagi.com/posts/2026-10-22-03-video-bilingual-subtitle-ai/Thu, 22 Oct 2026 12:18:00 +0800https://guijiagi.com/posts/2026-10-22-03-video-bilingual-subtitle-ai/本地视频自动生成字幕、翻译成双语、调整时间轴的完整流程，不用逐条手打https://guijiagi.com/posts/2026-10-22-02-audio-noise-reduction-record-cleanup/Thu, 22 Oct 2026 12:12:00 +0800https://guijiagi.com/posts/2026-10-22-02-audio-noise-reduction-record-cleanup/课录音、访谈录音的 AI 降噪流程，怎么避免过度处理把人声也削掉https://guijiagi.com/posts/2026-10-22-07-browser-bookmark-ai-cleanup/Thu, 22 Oct 2026 11:42:00 +0800https://guijiagi.com/posts/2026-10-22-07-browser-bookmark-ai-cleanup/导出书签 HTML，让 AI 批量按主题归类并标出失效重复项，收藏夹从垃圾堆变索引https://guijiagi.com/posts/2026-10-22-03-immersive-translate-bilingual-reading/Thu, 22 Oct 2026 11:18:00 +0800https://guijiagi.com/posts/2026-10-22-03-immersive-translate-bilingual-reading/沉浸式翻译类插件的装法、双语对照显示设置、PDF 与视频字幕场景的用法，读资料不再开翻译窗口https://guijiagi.com/posts/2026-10-22-02-ai-excel-formula-cleanup-guide/Thu, 22 Oct 2026 11:12:00 +0800https://guijiagi.com/posts/2026-10-22-02-ai-excel-formula-cleanup-guide/把表头和脏数据特征描述给 AI，拿到清洗公式直接粘贴使用，含常见错误处理与核对方法https://guijiagi.com/posts/2026-10-21-03-duplicate-photo-cleaner-ai-family-album/Wed, 21 Oct 2026 11:18:00 +0800https://guijiagi.com/posts/2026-10-21-03-duplicate-photo-cleaner-ai-family-album/相似连拍、截图、转发图占空间的清理路径：自动分组工具先过一遍，AI 给挑选建议，人工最后拍板，不删原片先归档https://guijiagi.com/posts/2026-10-21-01-zotero-ai-plugin-literature-notes-tutorial/Wed, 21 Oct 2026 11:06:00 +0800https://guijiagi.com/posts/2026-10-21-01-zotero-ai-plugin-literature-notes-tutorial/免费文献管理工具配合 AI 插件做摘要、笔记与引用整理，解决读论文划了一堆却找不到、写作时翻不到原文的问题https://guijiagi.com/posts/2026-10-20-05-ai-python-local-utility-script-pyinstaller/Tue, 20 Oct 2026 16:30:00 +0800https://guijiagi.com/posts/2026-10-20-05-ai-python-local-utility-script-pyinstaller/什么需求值得写成脚本，怎样把输入输出讲清楚给 AI，调试缩小报错范围，以及打包成可执行文件后的检查清单https://guijiagi.com/posts/2026-10-20-03-labelstudio-local-image-annotation-workflow/Tue, 20 Oct 2026 16:18:00 +0800https://guijiagi.com/posts/2026-10-20-03-labelstudio-local-image-annotation-workflow/给自家项目或小团队搭建标注工程：安装部署、标签体系、标注员培训规则、质检抽检，最后导出训练可用格式https://guijiagi.com/posts/2026-10-20-02-tampermonkey-ai-web-table-markdown-script/Tue, 20 Oct 2026 16:12:00 +0800https://guijiagi.com/posts/2026-10-20-02-tampermonkey-ai-web-table-markdown-script/重复从网页复制表格整理笔记的替代方案：安装 Tampermonkey，让 AI 生成读取表格的脚本，失效选择器怎么排查，导出怎样归档https://guijiagi.com/posts/2026-10-20-01-trae-wechat-miniprogram-local-query-tutorial/Tue, 20 Oct 2026 16:06:00 +0800https://guijiagi.com/posts/2026-10-20-01-trae-wechat-miniprogram-local-query-tutorial/不会写前端代码也能做出第一个小程序：AI IDE 生成页面与逻辑，微信开发者工具预览调试，讲清练手项目的选择与上线细节https://guijiagi.com/posts/2026-10-20-05-everything-bulk-rename-folder-ai-rules/Tue, 20 Oct 2026 11:30:00 +0800https://guijiagi.com/posts/2026-10-20-05-everything-bulk-rename-folder-ai-rules/下载文件夹越堆越乱的清理流程，先用 Everything 按类型找齐，再让 AI 起草批量重命名脚本，按主题归档https://guijiagi.com/posts/2026-10-20-04-obs-studio-prompter-screen-recording-setup/Tue, 20 Oct 2026 11:24:00 +0800https://guijiagi.com/posts/2026-10-20-04-obs-studio-prompter-screen-recording-setup/真人录课与产品讲解的画面分区搭建方法，配合 AI 生成自然的滚动提词稿，避免背稿式生硬表达https://guijiagi.com/posts/2026-10-20-02-xnconvert-imagemagick-batch-image-ai/Tue, 20 Oct 2026 11:12:00 +0800https://guijiagi.com/posts/2026-10-20-02-xnconvert-imagemagick-batch-image-ai/自媒体配图、卖家素材、活动照片批量压缩与统一尺寸的两套方案，图形界面与命令行各取所长，附压缩参数取舍https://guijiagi.com/posts/2026-10-20-01-deepseek-mermaid-flowchart-excalidraw-workflow/Tue, 20 Oct 2026 11:06:00 +0800https://guijiagi.com/posts/2026-10-20-01-deepseek-mermaid-flowchart-excalidraw-workflow/把一段文字方案交给 AI 生成可渲染的 Mermaid 流程图，再导入 Excalidraw 改成手绘风格，汇报与教程配图不再求人https://guijiagi.com/posts/2026-10-17-06-claude-code-upwork-freelance/Sat, 17 Oct 2026 16:36:00 +0800https://guijiagi.com/posts/2026-10-17-06-claude-code-upwork-freelance/不会写代码也能接web开发外包？用Claude Code加Cursor做全栈交付，从报价到验收的完整流程https://guijiagi.com/posts/2026-10-17-05-ai-data-analysis-side-gig/Sat, 17 Oct 2026 16:30:00 +0800https://guijiagi.com/posts/2026-10-17-05-ai-data-analysis-side-gig/餐饮老板看不懂美团后台数据，你用ChatGPT写pandas脚本把报表自动化，一张月报收费三百https://guijiagi.com/posts/2026-10-17-06-cursor-freelance-business-site-cost/Sat, 17 Oct 2026 11:36:00 +0800https://guijiagi.com/posts/2026-10-17-06-cursor-freelance-business-site-cost/用 Cursor Pro 配合 Vercel 部署，从需求沟通到上线交付的完整成本拆解，单站报价多少才不亏https://guijiagi.com/posts/2026-10-15-12-cline-open-agent-runtime/Thu, 15 Oct 2026 22:06:00 +0800https://guijiagi.com/posts/2026-10-15-12-cline-open-agent-runtime/Cline 十月中旬主打「开放智能体」定位，宣称被超一千一百万开发者使用，同一套运行时可跑在编辑器、终端或嵌入自己的产品里。https://guijiagi.com/posts/2026-10-15-11-codex-cli-terminal-cadence/Thu, 15 Oct 2026 22:00:00 +0800https://guijiagi.com/posts/2026-10-15-11-codex-cli-terminal-cadence/OpenAI Codex 仓库十月十四日傍晚仍在发版，累计发布近一千五百个版本，这个跑在终端里的轻量编码智能体正以极高节奏迭代。https://guijiagi.com/posts/2026-10-15-10-serving-agentic-queue-routing/Thu, 15 Oct 2026 21:54:00 +0800https://guijiagi.com/posts/2026-10-15-10-serving-agentic-queue-routing/十月十四日的工程复盘指出，面向智能体的推理服务开始补多轮记忆、混合精度与更安全的 rollout，用排队论选副本带来约百分之十四吞吐提升与近五成八的 TTFT 下降。https://guijiagi.com/posts/2026-10-15-09-hf-spec-decoding-local-vram/Thu, 15 Oct 2026 21:48:00 +0800https://guijiagi.com/posts/2026-10-15-09-hf-spec-decoding-local-vram/Hugging Face 十月十四日发布直播复盘，拆解草稿模型选择、接受率与显存预算三者如何共同决定本地投机解码是否真的变快。https://guijiagi.com/posts/2026-10-15-08-vllm-nixl-push-kv-transfer/Thu, 15 Oct 2026 21:42:00 +0800https://guijiagi.com/posts/2026-10-15-08-vllm-nixl-push-kv-transfer/vLLM 团队十月十四日发博客，联合 AWS 与 Mistral 推出推式 KV 传输，用独立后台线程把 KV 写离引擎前向循环，改善分离式推理的 TTFT 与 TPOT。https://guijiagi.com/posts/2026-10-15-16-unsloth-oct-mac-speculative/Thu, 15 Oct 2026 20:30:00 +0800https://guijiagi.com/posts/2026-10-15-16-unsloth-oct-mac-speculative/Unsloth 十月十二日更新带来 Mac 投机解码、更快的 MoE GGUF 与文本扩散 LLM 训练，本地训练又快一截。https://guijiagi.com/posts/2026-10-15-15-pytorchcon-2026-vllm-sessions/Thu, 15 Oct 2026 20:24:00 +0800https://guijiagi.com/posts/2026-10-15-15-pytorchcon-2026-vllm-sessions/PyTorch 北美大会十月二十至二十一日在圣何塞举行，vLLM 覆盖 KV 缓存、分离式服务、MoE 与注意力等专场。https://guijiagi.com/posts/2026-10-15-14-baseten-metainfer-vibeqwen/Thu, 15 Oct 2026 20:18:00 +0800https://guijiagi.com/posts/2026-10-15-14-baseten-metainfer-vibeqwen/Baseten 十月初发布 MetaInfer，用 Claude Code 与 Fable 5 自动生成推理引擎，Qwen-3.6-35B 解码 1792 token 每秒。https://guijiagi.com/posts/2026-10-15-13-outsystems-agent-experience-coding/Thu, 15 Oct 2026 20:12:00 +0800https://guijiagi.com/posts/2026-10-15-13-outsystems-agent-experience-coding/OutSystems 十月十二日在拉斯维加斯发布 Agent Experience，把 Mentor、Claude Code、Cursor、Codex、Kiro 接进低代码平台。https://guijiagi.com/posts/2026-10-15-12-bedrock-govcloud-gptoss-nemotron/Thu, 15 Oct 2026 20:06:00 +0800https://guijiagi.com/posts/2026-10-15-12-bedrock-govcloud-gptoss-nemotron/AWS GovCloud 十月十四日公告 GPT-OSS 20B 与 120B、GPT-5.4、Nemotron Nano 全系上线 Bedrock。https://guijiagi.com/posts/2026-10-15-11-comfyui-veo3lite-gptimage2-10bit/Thu, 15 Oct 2026 20:00:00 +0800https://guijiagi.com/posts/2026-10-15-11-comfyui-veo3lite-gptimage2-10bit/ComfyUI 更新日志十月十四日加入万相视频、Veo 3 Lite、GPT-Image 2 与核心节点十比特视频输入输出。https://guijiagi.com/posts/2026-10-15-12-tensorrtllm-nightly-cuda13-efa-oct15/Thu, 15 Oct 2026 16:06:00 +0800https://guijiagi.com/posts/2026-10-15-12-tensorrtllm-nightly-cuda13-efa-oct15/十月十五日 NGC 上架 tensorrtllm-runtime-nightly 新标签，基于 CUDA 13，带 EFA 变体，三家推理后端共用同一套 Dynamo 运行时。https://guijiagi.com/posts/2026-10-15-11-trendforce-2027-csp-capex-1.5t-ai-server/Thu, 15 Oct 2026 16:00:00 +0800https://guijiagi.com/posts/2026-10-15-11-trendforce-2027-csp-capex-1.5t-ai-server/TrendForce 预估 2026 年前九大 CSP 资本开支超 9570 亿美元年增 106%，2027 年破 1.5 万亿，AI 服务器占整体服务器产值 81%。https://guijiagi.com/posts/2026-10-15-10-google-75pct-ai-code-88pct-oversight/Thu, 15 Oct 2026 15:54:00 +0800https://guijiagi.com/posts/2026-10-15-10-google-75pct-ai-code-88pct-oversight/谷歌内部数据显示 AI 起草超 75% 生产代码，完全放手比例从 7.5% 升到 11.9%，Argon agent 跑 1389 次任务越复杂成功率越低。https://guijiagi.com/posts/2026-10-15-09-anthropic-scion-version-control-30k-merges/Thu, 15 Oct 2026 15:48:00 +0800https://guijiagi.com/posts/2026-10-15-09-anthropic-scion-version-control-30k-merges/AI 写代码把 Anthropic 单体仓库日合并推到 3 万次、日检出 2500 万次，Git 流量要超 600PB，七人团队五个月自研 Scion。https://guijiagi.com/posts/2026-10-15-08-vera-rubin-open-design-256cpu-800v/Thu, 15 Oct 2026 15:42:00 +0800https://guijiagi.com/posts/2026-10-15-08-vera-rubin-open-design-256cpu-800v/十月十五日英伟达向外部 OEM 开放下一代 AI 平台 Vera Rubin 服务器设计规范，推出 256 CPU 机架，直流电压提到 800 伏。https://guijiagi.com/posts/2026-10-15-16-zed-byok-gpt61-subagent/Thu, 15 Oct 2026 12:30:00 +0800https://guijiagi.com/posts/2026-10-15-16-zed-byok-gpt61-subagent/Zed 稳定版本周更新加入 GPT-6.1 Sol 的自带 Key 支持、子代理运行中实时切模型、子代理自动压缩与命令行 diff 分屏开关，本文实测新工作流。https://guijiagi.com/posts/2026-10-15-15-sglang-intel-xpu-install/Thu, 15 Oct 2026 12:24:00 +0800https://guijiagi.com/posts/2026-10-15-15-sglang-intel-xpu-install/SGLang Omni 的 Intel XPU 安装文档 10 月 14 日更新，给出从轮子安装到源码编译的完整路径，且刻意不钉死 sglang 版本，本文梳理安装步骤与常见报错。https://guijiagi.com/posts/2026-10-15-14-volcengine-las-sensitive-mask/Thu, 15 Oct 2026 12:18:00 +0800https://guijiagi.com/posts/2026-10-15-14-volcengine-las-sensitive-mask/火山引擎 AI 数据湖服务 10 月 10 日发布 1.6.0 版本，新增敏感内容打码算子，可自动识别图片视频中的敏感画面、文字与语音，并对敏感区域做高斯模糊或静音处理。https://guijiagi.com/posts/2026-10-15-13-claude-haiku-4-5-deprecation-migration/Thu, 15 Oct 2026 12:12:00 +0800https://guijiagi.com/posts/2026-10-15-13-claude-haiku-4-5-deprecation-migration/Anthropic 弃用表显示 claude-haiku-4-5-20251001 的退役评估下限为 2026 年 10 月 15 日，模型当前仍可用，但调用方现在就该规划降级路由。https://guijiagi.com/posts/2026-10-15-12-ollama-fly-agent-cloud/Thu, 15 Oct 2026 12:06:00 +0800https://guijiagi.com/posts/2026-10-15-12-ollama-fly-agent-cloud/Ollama 10 月 13 日发布博客宣布，AI 智能体现在可以在 Fly.io 的云电脑中安全使用 Ollama 的云端模型，本地 Agent 与云模型之间第一次有了托管级通道。https://guijiagi.com/posts/2026-10-15-11-vllm-adaptive-verification-online-estimator/Thu, 15 Oct 2026 12:00:00 +0800https://guijiagi.com/posts/2026-10-15-11-vllm-adaptive-verification-online-estimator/vLLM 团队 10 月 14 日发布技术博客，推出在线拟合的接受率估计器，让没有置信头的草稿模型也能跑自适应投机解码，本文拆解原理与部署要点。https://guijiagi.com/posts/2026-10-10-18-cascadep-async-moe-prefill/Sat, 10 Oct 2026 16:42:00 +0800https://guijiagi.com/posts/2026-10-10-18-cascadep-async-moe-prefill/CascadeEP 用异步专家并行、streamFFN 批处理、机会式权重拉取三招，治 MoE prefill 阶段注意力副本之间的长尾失衡。https://guijiagi.com/posts/2026-10-10-16-jevman-pacman-decision-benchmark/Sat, 10 Oct 2026 16:30:00 +0800https://guijiagi.com/posts/2026-10-10-16-jevman-pacman-decision-benchmark/Opper 十月九日发布 JEVMAN，让 Jev、Clef、GPT-6 Luna 等决策模型各玩 100 局吃豆人，按均分、延迟、成本排名，2 秒超时回退规则。https://guijiagi.com/posts/2026-10-10-12-pytorch-spyre-native-device/Sat, 10 Oct 2026 16:06:00 +0800https://guijiagi.com/posts/2026-10-10-12-pytorch-spyre-native-device/十月六日 PyTorch 官方博客详解 torch-spyre，用 PrivateUse1 给 Spyre 真实设备身份，让张量常驻、流、编译器抽象全部原生对接。https://guijiagi.com/posts/2026-10-10-07-architect-liquid-inference-auction/Sat, 10 Oct 2026 15:36:00 +0800https://guijiagi.com/posts/2026-10-10-07-architect-liquid-inference-auction/十月七日发布的 Liquid Inference 是个对每个请求跑现场拍卖的 LLM 路由，供应商实时竞价，买方按满足约束的最低出价成交。https://guijiagi.com/posts/2026-10-10-03-pplx-embed-v2-late-colbert/Sat, 10 Oct 2026 15:12:00 +0800https://guijiagi.com/posts/2026-10-10-03-pplx-embed-v2-late-colbert/pplx-embed-v2-late 用 0.6B 边缘模型跑查询、9B 模型建索引，做免 OCR 的 PDF 页图检索，MADQA 拿到 92.4。https://guijiagi.com/posts/2026-10-10-16-llm-openai-decisions-plugin-jev/Sat, 10 Oct 2026 12:30:00 +0800https://guijiagi.com/posts/2026-10-10-16-llm-openai-decisions-plugin-jev/Simon Willison 用 GPT-6 Astra 读文档自动写了 llm-openai-decisions 插件，让终端用户直接调用 OpenAI 的 Decisions API。https://guijiagi.com/posts/2026-10-10-15-vllm-cpu-zendnn-amd-epyc/Sat, 10 Oct 2026 12:24:00 +0800https://guijiagi.com/posts/2026-10-10-15-vllm-cpu-zendnn-amd-epyc/红帽 AI 3.4 引入 vLLM-CPU 的 ZenDNN 后端，为 AMD EPYC 处理器提供优化内核，让 CPU 推理性能逼近入门级 GPU。https://guijiagi.com/posts/2026-10-10-12-redhat-ai-inference-35-cpu-deploy/Sat, 10 Oct 2026 12:06:00 +0800https://guijiagi.com/posts/2026-10-10-12-redhat-ai-inference-35-cpu-deploy/红帽十月六日发布完整教程，用容器在 x86 CPU 上部署大语言模型推理，配套 cpueval CLI 做自动化性能基准。https://guijiagi.com/posts/2026-10-10-11-llm-036-cli-gpt6-sol-luna-single-turn/Sat, 10 Oct 2026 12:00:00 +0800https://guijiagi.com/posts/2026-10-10-11-llm-036-cli-gpt6-sol-luna-single-turn/llm 0.36 新增 gpt-6-sol 和 gpt-6-luna 模型 ID，插件可声明 supports_conversation=False 适配单轮模型。https://guijiagi.com/posts/2026-10-10-06-jetbrains-mellum21-coding-agent-model/Sat, 10 Oct 2026 11:30:00 +0800https://guijiagi.com/posts/2026-10-10-06-jetbrains-mellum21-coding-agent-model/Mellum 2.1 十月八日发布，12B 参数，Apache 2.0，专为路由、问答和子代理场景设计，用强化学习在真实编码环境中训练。https://guijiagi.com/posts/2026-10-10-04-saluki-27b-2bit-gguf-local-tool-calling/Sat, 10 Oct 2026 11:18:00 +0800https://guijiagi.com/posts/2026-10-10-04-saluki-27b-2bit-gguf-local-tool-calling/Conway Research 发布 Saluki 27B，2-bit GGUF 仅 7.89GB，Apache 2.0 协议，在工具调用基准上居然超过原版 Qwen3.8-27B。https://guijiagi.com/posts/2026-10-09-12-tilert-pluggable-decode/Fri, 09 Oct 2026 16:06:00 +0800https://guijiagi.com/posts/2026-10-09-12-tilert-pluggable-decode/一旦 prefill 与 decode 分开，decode 侧就变成可替换模块。vLLM x TileRT 示范了延迟敏感场景下的专用 decode 设计。https://guijiagi.com/posts/2026-10-09-11-fp8-kv-cache-capacity/Fri, 09 Oct 2026 16:00:00 +0800https://guijiagi.com/posts/2026-10-09-11-fp8-kv-cache-capacity/vLLM V1 已支持 FP8 KV Cache。把 KV 从 BF16 压到 FP8，显存占用直接减半，但要付出精度代价。本文给换算公式与取舍建议。https://guijiagi.com/posts/2026-10-09-10-qwen35-gdn-25k-tps/Fri, 09 Oct 2026 15:54:00 +0800https://guijiagi.com/posts/2026-10-09-10-qwen35-gdn-25k-tps/Qwen3.5 混合全注意力与 Gated Delta Network 层，vLLM 在 Blackwell 上单 GPU 冲到 25K TPS。难点在异构状态的跨 worker 传输。https://guijiagi.com/posts/2026-10-09-09-llmd-prefix-aware-routing/Fri, 09 Oct 2026 15:48:00 +0800https://guijiagi.com/posts/2026-10-09-09-llmd-prefix-aware-routing/llm-d 读取 vLLM 的 KV-cache 事件做 prefix-aware 路由，并用全局索引与 CPU、SSD 分层卸载管理分布式 KV 缓存。https://guijiagi.com/posts/2026-10-09-08-vllm-pd-disaggregation/Fri, 09 Oct 2026 15:42:00 +0800https://guijiagi.com/posts/2026-10-09-08-vllm-pd-disaggregation/vLLM 把计算密集的 prefill 与带宽密集的 decode 拆到不同实例，分别调优 TTFT 与 ITL。本文讲清什么时候该拆、拆了怎么调度。https://guijiagi.com/posts/2026-10-09-16-imatrix-gguf-quant/Fri, 09 Oct 2026 12:30:00 +0800https://guijiagi.com/posts/2026-10-09-16-imatrix-gguf-quant/同样是 4bit，带 importance matrix 的量化明显更稳。本文记录 imatrix 数据采集、量化与质量回测的完整流程。https://guijiagi.com/posts/2026-10-09-15-portkey-gateway-fallback/Fri, 09 Oct 2026 12:24:00 +0800https://guijiagi.com/posts/2026-10-09-15-portkey-gateway-fallback/单家模型挂了或涨价怎么办。本文用 Portkey 在前面加一层网关，挂多家供应商，自动 fallback 与重试，上层代码零改动。https://guijiagi.com/posts/2026-10-09-14-deepeval-pytest-ci/Fri, 09 Oct 2026 12:18:00 +0800https://guijiagi.com/posts/2026-10-09-14-deepeval-pytest-ci/提示词一改就翻车，是因为没有回归集。本文用 DeepEval 把评测做成 pytest 用例，接进 CI，每次改动自动跑黄金集。https://guijiagi.com/posts/2026-10-09-13-tgi-canary-rolling/Fri, 09 Oct 2026 12:12:00 +0800https://guijiagi.com/posts/2026-10-09-13-tgi-canary-rolling/换模型最怕全量重启引发雪崩。本文用 HuggingFace TGI 加一层轻量网关，实现新模型金丝雀发布与滚动回滚。https://guijiagi.com/posts/2026-10-09-12-exl2-tabbyapi-multigpu/Fri, 09 Oct 2026 12:06:00 +0800https://guijiagi.com/posts/2026-10-09-12-exl2-tabbyapi-multigpu/高并发推理框架之外，ExLlamaV2 配 TabbyAPI 是本地多卡跑长上下文的另一套选择。本文记录其启动参数、显存占用与分工。https://guijiagi.com/posts/2026-10-09-11-tensorrtllm-blackwell-build/Fri, 09 Oct 2026 12:00:00 +0800https://guijiagi.com/posts/2026-10-09-11-tensorrtllm-blackwell-build/GB300 的 $0.123 单价背后是 TensorRT-LLM 编译出的 engine。本文记录在 Blackwell 上 build engine 的流程、关键参数与常见编译错误。https://guijiagi.com/posts/2026-10-08-12-moe-wideep-expert-parallel/Thu, 08 Oct 2026 16:06:00 +0800https://guijiagi.com/posts/2026-10-08-12-moe-wideep-expert-parallel/MoE 模型 decode 阶段被专家权重加载时间卡住。NVIDIA 的 WideEP 把专家跨 NVL72 节点分片，是 2026 年 MoE 部署的关键工程。https://guijiagi.com/posts/2026-10-08-11-nvidia-dynamo-disaggregation/Thu, 08 Oct 2026 16:00:00 +0800https://guijiagi.com/posts/2026-10-08-11-nvidia-dynamo-disaggregation/Dynamo 1.0 四月发布、六月迭代到 v1.2，把 prefill 与 decode 拆到不同资源池。本文解释这个早就被讨论的概念为什么今年才真正可用。https://guijiagi.com/posts/2026-10-08-10-kv-cache-compression-2026/Thu, 08 Oct 2026 15:54:00 +0800https://guijiagi.com/posts/2026-10-08-10-kv-cache-compression-2026/长上下文与 Agent 并发把 KV cache 推成推理成本第二大项。本文梳理 2026 年可用的 KV 压缩组合拳，不重复 FlashAttention 旧账。https://guijiagi.com/posts/2026-10-08-09-nvfp4-quantization-blackwell/Thu, 08 Oct 2026 15:48:00 +0800https://guijiagi.com/posts/2026-10-08-09-nvfp4-quantization-blackwell/Blackwell 之后，FP4 从后训练量化技巧变成硬件原生格式。本文梳理 NVFP4、MXFP4 与 DeepGEMM 这条新量化路线。https://guijiagi.com/posts/2026-10-08-08-sglang-radix-prefix-cache/Thu, 08 Oct 2026 15:42:00 +0800https://guijiagi.com/posts/2026-10-08-08-sglang-radix-prefix-cache/SGLang 用基数树做 KV 前缀缓存，并在 DeepSeek-V4、Qwen3.8 上做到 day-0 支持。本文拆解这套机制为什么对多轮 Agent 至关重要。https://guijiagi.com/posts/2026-10-08-whisper-subtitle-pipeline/Thu, 08 Oct 2026 12:54:00 +0800https://guijiagi.com/posts/2026-10-08-whisper-subtitle-pipeline/想要一批视频的文字稿，不用手工听写。yt-dlp 下载、faster-whisper 转写，串起来就是一条本地可复现的字幕流水线。https://guijiagi.com/posts/2026-10-08-promptfoo-prompt-regression/Thu, 08 Oct 2026 12:30:00 +0800https://guijiagi.com/posts/2026-10-08-promptfoo-prompt-regression/改一行提示词，回答质量可能悄悄崩。promptfoo 把提示词评测做成可重复跑的命令，让你像测代码一样测提示词。https://guijiagi.com/posts/2026-10-08-openwebui-local-team-chat/Thu, 08 Oct 2026 12:24:00 +0800https://guijiagi.com/posts/2026-10-08-openwebui-local-team-chat/成员各自装 ChatGPT、各自充会员，是最乱的用法。Open WebUI 一个网页，背后同时接本地模型和云端 API，团队共用。https://guijiagi.com/posts/2026-10-08-litellm-gateway-proxy/Thu, 08 Oct 2026 12:00:00 +0800https://guijiagi.com/posts/2026-10-08-litellm-gateway-proxy/团队同时接 OpenAI、DeepSeek、智谱、本地模型时，各自适配是噩梦。LiteLLM 用一套 OpenAI 协议把它们统一成一个入口。https://guijiagi.com/posts/2026-10-08-langfuse-agent-tracing/Thu, 08 Oct 2026 11:54:00 +0800https://guijiagi.com/posts/2026-10-08-langfuse-agent-tracing/Agent 出了错，最难的是复现「它当时看到了什么、调了哪个工具」。Langfuse 这类开源观测工具能把每一步轨迹完整录下来。https://guijiagi.com/posts/2026-10-08-ai-code-review-github-actions/Thu, 08 Oct 2026 11:06:00 +0800https://guijiagi.com/posts/2026-10-08-ai-code-review-github-actions/用 GitHub Actions 加一个 PR 自动评审工作流，调大模型 API 对 diff 逐条点评。本文给出可直接复制的 yml 与提示词要点。https://guijiagi.com/posts/2026-10-07-gptq-awq-quantization-deep-dive/Wed, 07 Oct 2026 18:54:00 +0800https://guijiagi.com/posts/2026-10-07-gptq-awq-quantization-deep-dive/同样是把权重压到 4bit，GPTQ 用二阶误差补偿，AWQ 保护显著权重。拆解两派原理、质量差异与选型建议。https://guijiagi.com/posts/2026-10-07-flashattention-kv-cache-deep-dive/Wed, 07 Oct 2026 18:48:00 +0800https://guijiagi.com/posts/2026-10-07-flashattention-kv-cache-deep-dive/注意力是 O(N²) 的显存怪兽，KV Cache 又随并发与长度膨胀。拆解 FlashAttention 怎么省显存，以及长上下文的成本账。https://guijiagi.com/posts/2026-10-07-prompt-structured-output-json/Wed, 07 Oct 2026 18:42:00 +0800https://guijiagi.com/posts/2026-10-07-prompt-structured-output-json/模型偶尔吐出的 JSON 缺括号、带多余文字。从 prompt 约束到 JSON mode 再到语法约束，三层手段把格式锁死。https://guijiagi.com/posts/2026-10-07-ollama-lmstudio-local-tips/Wed, 07 Oct 2026 18:36:00 +0800https://guijiagi.com/posts/2026-10-07-ollama-lmstudio-local-tips/Ollama 偏命令行、LM Studio 偏图形界面。从模型选型、上下文长度到 API 调用，讲清两个本地工具的分工与坑。https://guijiagi.com/posts/2026-10-07-speculative-decoding-speedup/Wed, 07 Oct 2026 18:30:00 +0800https://guijiagi.com/posts/2026-10-07-speculative-decoding-speedup/draft 模型先猜、主模型批量验证，吞吐能提 1.5 到 3 倍。但词表不匹配会反噬。手把手讲清怎么开、怎么调。https://guijiagi.com/posts/2026-10-07-office-ai-automation-toolchain/Wed, 07 Oct 2026 12:42:00 +0800https://guijiagi.com/posts/2026-10-07-office-ai-automation-toolchain/每周花三小时整理报表、回邮件、归档文件？这些重复动作才是脚本和 AI 该接管的。三个切入点，照做就能省时间。https://guijiagi.com/posts/2026-10-07-amd-roc-local-llm-deploy/Wed, 07 Oct 2026 12:12:00 +0800https://guijiagi.com/posts/2026-10-07-amd-roc-local-llm-deploy/AMD 卡不是不能跑大模型，是 ROCm 版本和显卡型号要对得上。这篇讲清支持列表、安装坑和 Ollama/LM Studio 的调用方式。https://guijiagi.com/posts/2026-10-07-ollama-common-commands-api/Wed, 07 Oct 2026 12:06:00 +0800https://guijiagi.com/posts/2026-10-07-ollama-common-commands-api/ollama run 一条命令跑模型，pull 拉取，serve 起服务。这篇把高频命令、Modelfile 自定义和常见报错一次整理。https://guijiagi.com/posts/2026-10-07-lm-studio-gui-local-model-api/Wed, 07 Oct 2026 12:00:00 +0800https://guijiagi.com/posts/2026-10-07-lm-studio-gui-local-model-api/不想敲命令？LM Studio 把下载、量化选择、对话、开 OpenAI 兼容 API 全做成了图形界面。新手本地跑模型的最短路径。https://guijiagi.com/posts/2026-10-07-claude-desktop-mcp-config-troubleshoot/Wed, 07 Oct 2026 11:42:00 +0800https://guijiagi.com/posts/2026-10-07-claude-desktop-mcp-config-troubleshoot/配置文件写错一个字符，MCP 工具就静默消失。这篇把 json 配置模板、日志排查路径和几个开箱即用的官方服务器一次讲清。https://guijiagi.com/posts/2026-10-07-vector-database-selection-guide/Wed, 07 Oct 2026 11:06:00 +0800https://guijiagi.com/posts/2026-10-07-vector-database-selection-guide/百万级以内用 pgvector 或 Chroma 就够，Qdrant 适合生产混合检索，Milvus 留给亿级规模。选错库，运维成本比省下的查询时间还贵。https://guijiagi.com/posts/25-gpt-live-1-full-duplex-voice-api/Mon, 14 Sep 2026 00:00:00 +0000https://guijiagi.com/posts/25-gpt-live-1-full-duplex-voice-api/2026年9月10日，OpenAI发布GPT-Live-1语音对话API。 如果你只看新闻标题，"OpenAI又发新模型"似乎并不稀奇。但只要亲自用过它的Demo，就会立刻意识到这次不一样：你说话的时候，它在听；它说话的时候，你可以随时打断；你停顿半秒，它会接话；你说"等一下"，它真的会等。