Exploring efficient open-source LLM training and inference on AMD ROCm, with a focus on portability, performance, and reproducible experiments.