过去这一年我们团队陆陆续续做了几个 AI Service——大多是拿 FastAPI 包一个模型,再对外出 API。做到第三四个的时候我意识到一个尴尬的事实:每个项目长得都差不多,但每次起新项目还是在从零开始复制粘贴,而且不同的人写出来的目录结构、配置方式、metrics 命名都不太一样。一个新人接手第二个服务的时候,几乎要把第一个服务的认知全部重学一遍。 这篇文章想聊的不是某个具体框架,而是把这几个项目踩过的坑、迭代下来的设计、留下来的规范集中讲一遍——也算是给”维护一个模型 API 服务”这件事做一次复盘。
I’ve been overwhelmed by security work lately, so let me switch to something lighter to relax my mind. Python 3.14 has officially introduced a new mechanism called Tail Call Interpreter (Made by Ken Jin), which is undoubtedly another major milestone that lays the foundation for the future.
During the National Day holiday, I worked on a project to use wasmtime to execute CPython virtual machine compiled into WASM/WASI bytecode, and extend it with Host Functions implemented in Python on the host side. I’d like to clarify again that this is just a personal project I wanted to work on, without any validation in production environments, just for fun (XDDD