Did a 30 runs of llama-bench to find optimal settings for my use case (Frigate and HomeAssistant) on my MI60 32gb VRAM GPU - two models tested Gemma4 and Qwen3.6 - Figured I'd share in case it helps anyone else

I'm running llama.cpp using this docker container: https://github.com/mixa3607/ML-gfx906 (it's just a lot easier than building from source, which I was doing previously). The MI60 (or MI50) are just a real pain in the be