The default PyTorch 1.12 binary doesn't work with CUDA 11.6, it seems.
This is why there's this bodge:
https://github.com/replicate/cog-stable-diffusion/blob/bbedf972a628f7e3befcf8acb36efdb95a9bebc6/cog.yaml#L5-L7
Looking at the PyTorch website, looks like selecting 11.6 you need that argument:
But for 11.7 you don't need the argument:
We should represent this in the compatibility tables.
The default PyTorch 1.12 binary doesn't work with CUDA 11.6, it seems.
This is why there's this bodge:
https://github.com/replicate/cog-stable-diffusion/blob/bbedf972a628f7e3befcf8acb36efdb95a9bebc6/cog.yaml#L5-L7
Looking at the PyTorch website, looks like selecting 11.6 you need that argument:
But for 11.7 you don't need the argument:
We should represent this in the compatibility tables.