📄 中文摘要
在尝试在本地运行新的AI模型时，常常会遇到CUDA内存不足的错误，尤其是在使用双GPU设置（如NVIDIA RTX 3060）时。尽管总共有24GB的显存，但由于显存被分配到两个显卡上，许多AI工具的默认设置会导致系统崩溃。经过数月在Docker化Windows环境中的反复试验，开发出了一种“零崩溃管道”，该管道提供了从原始微调到使用Ollama、OpenClaw和ComfyUI实现智能体现实的具体蓝图。
24GB AI实验室：消费者硬件上全栈本地AI的生存指南
📄 English Summary
The 24GB AI Lab: A Survival Guide to Full-Stack Local AI on Consumer Hardware
Running a new AI model locally often leads to a 'CUDA Out of Memory' error, particularly with dual-GPU setups like the NVIDIA RTX 3060. While having a total of 24GB of VRAM, the physical split across two cards causes many AI tools' default settings to crash the system. After months of trial and error in a Dockerized Windows environment, a 'Zero-Crash Pipeline' has been developed. This pipeline serves as a blueprint for transitioning from raw fine-tuning to an agentic reality using Ollama, OpenClaw, and ComfyUI.