code does not work with older graphics cards

[xvdp]

Code uses dtype=torch.bfloat16 and torch.backends.cuda.matmul.allow_tf32 = True, only available in the latest graphics cards.
Error is easy to detect when loading in free colab, which typically gives you Tesla generation.

train_stage_B.py
train_stage_C.py
würstchen-stage-B.ipynb
würstchen-stage-C.ipynb 

as well as the file linked with 'open in colab' which is not one of the the ipynbs

There is another problem with the "open in colab" is that the free colab graphics cards are typically smaller.
files affected