Comment by ipsum2

I, and several others, had no problem running on PyTorch on AMD GPUs, with no code changes from CUDA. Check out MosaicML's blog posts: https://www.databricks.com/blog/training-llms-scale-amd-mi25...