There's a lot of kernels that are specialized for particular sets of input parameters and tuned for improved performance on specific hardware, which makes the libraries a couple hundred megabytes per architecture. The ROCm libraries are huge because they are fat binaries containing native machine code for ~13 different GPU architectures.

He’s not wrong. I did a new arch install to try and get a 7800XT working with ROCm and PyTorch and was concussed on how I ran out of space but saw that ROCm was 18GB.