Pytorch choose rocm or cuda depending on local settings or available gpu

[NSA-KEY]

I need to run the same PyTorch project on various machines, including those with CUDA GPUs and AMD GPUs without changing pyproject.toml.