One repo to hoard every LLM link you’ll never finish reading

What it does The repo is a manually curated Chinese-language directory of LLM resources—blog posts, papers, and code walkthroughs—organized across the full stack: Transformer algorithms, distributed training, fine-tuning, RLHF alignment, inference optimization, model compression, benchmarks, and down to CUDA and AI chips.

The interesting bit It treats the LLM lifecycle as one continuous syllabus, mixing arXiv surveys with Zhihu explainers and WeChat articles. For Mandarin-reading developers, it is a rare attempt to map everything from nanoGPT source dives to TensorRT-LLM architecture in a single flat file.

Key highlights

• Coverage spans the whole pipeline: pre-training math, PEFT, RLHF/DPO/ORPO papers, speculative decoding, KV cache optimization, vLLM source analysis, and quantization toolkits
• Heavy tilt toward Chinese-language sources—Zhihu columns, WeChat official account posts, and Baidu forums—alongside English references from Hugging Face and EleutherAI
• Includes infrastructure layers often skipped by algorithm-focused lists: AI chips, CUDA, AI compilers, and LLMOps
• Maintainers explicitly invite community contributions to expand the hoard

Caveats

• It is purely a link aggregator; the README contains no runnable code, tools, or original implementations
• Nearly all linked content is Chinese-language; English-only readers will get little from the curated articles
• Flat, manually maintained structure with no search or tagging beyond emoji section headers

Verdict A useful starting point for Mandarin-speaking engineers and researchers who want a broad map of the LLM landscape; skip it if you need executable frameworks or an English-first reading list.