GPU / CUDA / Docker 部署兼容性清单：从镜像到模型服务
用版本矩阵、镜像 digest、模型校验、GPU 可见性、安全边界和功能验收，把消费级 GPU 部署记录整理成可复用的兼容性案例。
Topic / 22 notes
AI & LLM
从模型调用到 Agent 系统：检索、工作流、推理服务、工具权限与生产化实践。
Recommended starting point
Recommended around AI & LLM
AI Agent 工程化学习地图：从模型调用到可控协作
把 RAG、LangChain、LangGraph、DeepAgents、OpenCode 和模型网关等历史文章整理成一条从 API 调用到可观测 Agent 系统的学习路线。
Read note →Ray + vLLM 离线多节点部署：从可复现安装到安全运维
把 Ray wheelhouse、GPU 容器、并行策略、私有网络、认证、健康检查和回滚拆开，避免把一条历史启动命令误当成生产保证。
Read note →LLM 推理服务上线清单：从单机验证到多节点运维
整合 vLLM、Ray、Docker、模型网关和离线部署笔记，给出模型服务从容量估算到灰度、监控与回滚的生产检查表。
Read note →GPU / CUDA / Docker 部署兼容性清单：从镜像到模型服务
用版本矩阵、镜像 digest、模型校验、GPU 可见性、安全边界和功能验收，把消费级 GPU 部署记录整理成可复用的兼容性案例。
Read note →Agent 沙箱与工具权限：从 Docker 到可审计执行边界
以威胁模型为中心整理 DeepAgents、OpenCode、Agent Swarm 和文件服务实践中的最小权限、网络、密钥、审批、预算与审计设计。
Read note →
The primary shelf
22 articles
Primary-topic AI & LLM notes
Ray + vLLM 离线多节点部署：从可复现安装到安全运维
把 Ray wheelhouse、GPU 容器、并行策略、私有网络、认证、健康检查和回滚拆开，避免把一条历史启动命令误当成生产保证。
Agent 沙箱与工具权限：从 Docker 到可审计执行边界
以威胁模型为中心整理 DeepAgents、OpenCode、Agent Swarm 和文件服务实践中的最小权限、网络、密钥、审批、预算与审计设计。
LLM 推理服务上线清单：从单机验证到多节点运维
整合 vLLM、Ray、Docker、模型网关和离线部署笔记，给出模型服务从容量估算到灰度、监控与回滚的生产检查表。
AI Agent 工程化学习地图：从模型调用到可控协作
把 RAG、LangChain、LangGraph、DeepAgents、OpenCode 和模型网关等历史文章整理成一条从 API 调用到可观测 Agent 系统的学习路线。
Agent Swarm: 部署一个自主协作的 AI 编程团队
Introduction to Agent Swarm, an open-source multi-agent AI operating system, with a step-by-step focus on offline deployment of its Docker-based Lead and Worker architecture.
从零到一：构建你的 LLM 应用开发知识体系
Learning roadmap for LLM application development, mapping RAG, LangChain, LangGraph, and DeepAgents from fundamentals to production-grade agent workflows.
从零搭建 vLLM 多节点推理集群：Docker 部署全流程实战
手把手教你用 Docker + Ray + vLLM 在多节点 GPU 集群上部署 70B+ 大模型推理服务，覆盖在线/离线安装、模型选型、性能调优与生产运维。
New-API 部署同一模型双进程双端口：实现负载均衡与高可用的完整指南
Deploy two New-API processes for the same model on separate ports to achieve load balancing and high availability, with complete Docker Compose configuration and tuning notes.
深入解析 OpenCode CLI：从架构设计到实战构建
OpenCode 是一个开源的 AI 编程助手，运行在终端中。本文深入解析其 TUI 架构、声明式 UI 渲染、Client-Server 分离设计，并手把手教你用 Bun + SolidJS + OpenTUI 构建一个类似的 AI CLI 工具。
Running DeepSeek-V4-Flash-0731 on 8x RTX 4090D with Docker
Step-by-step guide to building a Docker image and serving DeepSeek-V4-Flash-0731 on 8x RTX 4090D consumer GPUs using the vLLM SM89 fork.
Multi-Node LLM Serving: vLLM + Ray
End-to-end offline deployment of vLLM with Ray across two nodes in Docker, covering worker discovery, tensor-parallel inference, and automated health checks.
Multi-Node LLM Serving: Architecture, Frameworks & Best Practices
Overview of multi-node LLM serving architectures comparing vLLM, TensorRT-LLM, and SGLang, with deployment strategies for 70B+ models across GPU clusters.
MiniCPM5-1B Overview
Overview of MiniCPM5-1B, a 1B-parameter on-device language model from 面壁智能, with benchmarks, architecture details, and deployment considerations.
Deploying MiniCPM5-1B with llama.cpp
Offline deployment of MiniCPM5-1B using llama.cpp server in Docker, with Open-WebUI frontend and CPU-only inference configuration.
Deploying MiniCPM5-1B with Ollama
Offline deployment of MiniCPM5-1B using Ollama and Open-WebUI in Docker, covering image loading, model import, and CPU-only inference setup.
Yuxi Know
Setup and usage notes for Yuxi, an open-source knowledge base and knowledge graph agent platform for building searchable document collections.
AI Coding Agent: OpenCode
Overview of OpenCode, an open-source AI coding assistant with intelligent completion, multi-model support, and a CLI-based conversational workflow.
AI Model Hub: New API
Self-hosted gateway for unifying access to multiple AI providers, with usage tracking, key management, and OpenAI-compatible API endpoints.
Sentence Transformers: Sentence-BERT
Notes on using Sentence Transformers for semantic search and text similarity, covering embedding generation, reranking, and integration with vector stores.
DeepAgents 接入 DeepSeek 配置指南
Guide for integrating DeepSeek as the LLM backend for DeepAgents, covering uv package management, API key setup, and Windows verification.
基于DeepAgents的可私有化部署的沙箱方案
Private deployment of DeepAgents using sandbox backends for isolated code execution, with Docker-based sandboxes and custom backend configuration.
Keep exploring
Connected topics
- SystemsGPU, Linux, storage and networking
- DevOps & InfrastructureContainers, CI/CD, observability and rollback
- Security & NetworkingAgent sandbox, tool permissions and audit
- ProgrammingAgent gateways and service code