LLM engineering by the notebook: from PyTorch loops to GPU kernels

What it does

This is a Chinese-language tutorial lab built as Jupyter notebooks that treats LLM algorithms and systems as a stack of practice exercises. It runs from PyTorch implementations of training and inference all the way down to Triton kernels and CUDA optimization, organized into five progressive parts and cross-cutting topics like quantization and parallelism. Each unit pairs theory with runnable code, problem sets, and validation.

The interesting bit

The project borrows its structure from coding interview prep: every topic is split into a problem block, an answer block, and test cases. It also forces a hardware-aware discipline, requiring learners to back up claims with GPU benchmarks and profiling traces instead of hand-waving.

Key highlights

• Five-part arc: Python/PyTorch basics → hardware and math → PyTorch algorithms → Triton kernels → CUDA/system optimization (Part 04 unfinished, Part 05 reserved).
• Cross-cutting “topic routes” let you jump straight to fine-tuning, inference optimization, memory tuning, or profiling without reading linearly.
• Notebook-first format with problem/answer/test sections, designed to be modified and rerun.
• Explicit emphasis on evidence over slides: benchmarks, profiling, and real GPU experiments are core to the methodology.
• README is bilingual, but the educational content itself is primarily in Chinese.

Caveats

• Part 04 (CUDA C++ and system optimization) is under construction, and Part 05 is just a placeholder, so the deepest low-level material is still incomplete.
• The “Operators & Compilers” topic route is also flagged as in-progress, meaning graph-level optimization content is still thin.

Verdict

Ideal for Mandarin-speaking developers who want to move from reading LLM papers to rewriting attention kernels and measuring their throughput. Not useful if you are looking for a packaged library or a quick English-language reference.