How to tune llama.cpp for different ROCm gpus?

[ccbadd]

I have a server with 2 MI100s and 2 W6800s. When I use them for inferencing the W6800s are nearly twice the speed of the MI100s. Most of the devs that have worked on llama.cpp and the like seem to have access to an RX6800 or similar so everything is tuned for those cards. The MI100s should be a lot faster and I have seen mention of tuning for specific cards. How do we go about tuning for the MI's, or any other card? Such an absolute waste of recourses.