ml-engineering

Machine Learning Engineering Open Book is a practical, experience-driven repository by Stas Bekman that documents methodologies, tools, and copy-paste commands for training and operating large language models (LLMs) and vision-language models (VLMs). The content draws on hands-on work with BLOOM-176B, IDEFICS-80B, and RAG systems at Contextual.AI. The repository is organized into seven parts: insights on AI engineering and cloud/GPU decisions; hardware guidance covering compute, storage, and networking; orchestration systems including SLURM; training and inference guides; development resources for debugging, testing, and troubleshooting; and miscellaneous resources such as LLM/VLM training logbooks. Practical tools include network benchmark scripts like all_reduce_bench.py, inter-node connectivity testing with torch-distributed-gpu-test.py, and accelerator TFLOPS measurement via mamf-finder.py. The project also provides comparison tables for accelerator performance and network speeds, a SKILL.md file for teaching AI agents, and downloadable PDF/EPUB ebook versions. The content is licensed under CC BY-SA 4.0 and welcomes community contributions and discussions.