> To leverage the capabilities of MI250X, ROCm enables the use of GPU matrix cores through its rocBLAS and MIOpen library implementations that, in turn,are leveraged by PyTorch.

- https://aclanthology.org/2023.emnlp-main.164.pdf

https://github.com/TurkuNLP/

They probably got a lot of hand holding from AMD.

Having access to enterprise GPUs on one of the biggest HPCs systems in Europe is probably enough.

AMDs bad rep in AI is mostly due to flaky support of its consumer GPUs.