A regular TensorFlow installation, just the Python library, is an 184 MB wheel that unpacks to about 1.2 GB of stuff. I have no clue what mess goes in there, but it's a lot.
Still, if you're right that this package seems to take 18 GB disk size, something weird is going on.
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.
Still, if you're right that this package seems to take 18 GB disk size, something weird is going on.