The community is basically doing this now with the kickflips in driver/rocm/etc "stuff". It's gotten so good and well-embraced that even AMD is backing some of these community efforts.

Do you have some pointers to these efforts, please?

halogen. made by a user, one of the early, best model serving things for a strixhalo. amd community engaging person saw it was well liked, and sent that developer 2 strix halos to help them develop it further.

ive seen that same amd community manager engage on reddit and actively pull in changes into lemonade server, for better ways to serve up models best on the strixhalo hardware. so they're accepting community help where it helps.

Is there something wrong with PyTorch? I’ve heard people complain a lot about their GPUs. I’ve only used MI250x in an HPC, but PyTorch worked fine.

It's due to ROCm, think equivalent of CUDA lower level drivers, which is well problematic.

There's huge upswing in using Vulkan Compute which works across NVIDIA, AMD, and Intel (and even others), and which is apparently much easier and stable to run on AMD