微调自己的 LLM，最烦的往往不是训练，而是配环境。
![微调自己的 LLM，最烦的往往不是训练，而是配环境。 Soup 直接把流程压成三步： pip install "soup-cli[train]" soup init --template chat soup train 一个 YAML…](https://pbs.twimg.com/media/HSlC_jtakAA-H8X.png?name=large)
This is a dev post classified by Jev as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
微调自己的 LLM，最烦的往往不是训练，而是配环境。 Soup 直接把流程压成三步： pip install "soup-cli[train]" soup init --template chat soup train 一个 YAML 管完训练配置，聊天、代码、工具调用、推理、视觉、RLHF 等模板也都有。 更狠的是低显存方案：4GB 显存也能微调 8B 模型，通过 layer streaming 把冻结参数逐层送进 GPU，峰值约 3.32GB，官方测试最高约 119 tokens/s。 训练完还能导出 GGUF，直接丢进 Ollama，也能上传 HuggingFace。 🔗 https://github.com/MakazhanAlpamys/Soup
Posted by lumxss (24.7k followers) 2 h ago · 1 likes · 374 views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: soup.
More dev work like this
- 💥什么鬼！本地开源 Jev 居然杀到了 0 输出 Token，自带三档模型直接落地结构化判断！ — @NFT_Chen
- AI Agent System for Deep Learning Model Migration from TensorFlow to JAX! #BigData… — @gp_pulipaka
- Vectorizing the Math for LLM! #BigData #Analytics #DataScience #AI #MachineLearning #IoT… — @gp_pulipaka
- Hermes Wingtips #79: `agent[.]disabled_toolsets` — @witcheer
- Give your coding agent a box — not your whole machine. — @DanKornas
- it's awesome to see the reception here 🔥 — @jerryjliu0
- 🌙 中秋福利续上！OpenCode 官方再延长一周！DeepSeek V4.1 Flash 的 4 倍额度已延至 9 月 27 日，完美覆盖中秋！ — @NFT_Chen
- 🚨 Step 5 Preview — @chetaslua
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 18.9k posts from 4.9k X accounts over the last 21 days, 2.1k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-21 06:12 UTC. Full method.