My first impressions on ROCm and Strix Halo

A developer documents practical setup of AMD's Strix Halo APU with ROCm GPU compute stack on Ubuntu 24.04 LTS, including BIOS configuration, kernel tuning for shared CPU/GPU memory, and PyTorch integration. The guide demonstrates running Qwen 3.6 language model locally via llama.cpp with large context windows, validating ROCm's viability for local LLM inference workloads.