ai-twinkle/ultrascale-playbook-zh-tw

This project is an unofficial Traditional Chinese translation of The Ultra-Scale Playbook: Training LLMs on GPU Clusters, a guide originally published by the nanotron team at Hugging Face. The translation is maintained by the Twinkle AI Community in Taiwan, and the original authors are credited by name. Both the original and the translation are released under the Apache 2.0 license. The original playbook walks through how to train very large language models across many GPUs at once, starting from understanding memory use on a single GPU and building up to data, tensor, context, pipeline, and expert parallelism, along with GPU kernel optimization and mixed precision training. What sets this translated version apart is that every chapter has its own interactive lab, recreating the spirit of the original site's interactive diagrams with custom built components rather than a plain text translation. Examples listed include a memory calculator for a single GPU, a visualization of ZeRO memory sharding for data parallelism, animations of tensor parallelism splitting, a Ring Attention animation for context parallelism, a bubble scheduling simulator comparing two pipeline parallelism approaches, a Mixture of Experts routing visualization, a five dimensional parallelism configuration explorer, a step by step configuration wizard, animations of matrix multiplication tiling, and a floating point format explorer covering FP32, BF16, FP16, and FP8. The repository is organized into a folder of chapter by chapter translated markdown files, a static interactive website that can run entirely offline, the original English HTML source and bibliography kept for reference, the original source markdown, and a build script that turns the translated markdown into the website. Running it locally is a matter of starting a simple Python web server inside the webapp folder and rebuilding with the build script after editing any translation or component. The translation notes explain that the team worked from the original markdown as the primary source, cross checked against the original HTML for anything lost in format conversion, restored broken math notation and code blocks, and kept certain English abbreviations and terms untranslated, such as the parallelism strategy names and collective communication operation names, while reusing the original 134 illustrations rather than extracting them from a PDF.