// the find
liguodongiot/llm-action
本项目旨在分享大模型相关技术原理以及实战经验（大模型工程化、大模型应用落地）
A Chinese-language knowledge base covering the full LLM engineering stack: training, fine-tuning, inference optimization, quantization, and deployment on domestic AI chips. It's primarily a curated collection of blog posts and tutorials from the author's Zhihu/CSDN writing, with accompanying code. The target audience is Chinese engineers building or deploying large models in production.
The coverage of Ascend/Huawei NPU adaptation is unusually thorough — most English resources ignore this entirely, so for teams working under US chip sanctions this is genuinely hard to find elsewhere. The fine-tuning section walks through real experiments from 6B to 65B with actual code (LoRA, QLoRA, P-Tuning v2, GaLore), not just theory. The inference optimization section links to nano-vLLM and mini-SGLang — stripped-down reimplementations specifically for reading source code, which is a better learning tool than the production versions. The quantization coverage is comprehensive: GPTQ, AWQ, SmoothQuant, ZeroQuant, FP8, FP6, FP4 — most are linked to original analysis rather than just pointing at library docs.
The repo is essentially a personal blog index with some code attached — the actual content lives on Zhihu, Juejin, and CSDN behind Chinese platforms, so you can't use it offline or fork it meaningfully. Many sections are marked '待更新' (to be updated) and have been for a long time — sequence parallelism coverage, data engineering pipeline details, and the AI compiler deep-dives are skeletons. The code samples are scattered Jupyter notebooks without a coherent project structure or dependency management, so running them requires hunting down the right environment from the prose. If you don't read Chinese, roughly 95% of the value is inaccessible.