Closed thojmr closed 1 year ago
You have to download and use the 32bit models.
Makes sense. It would be nice if it would just autocast to 32bit. Everyone that uses this script is going to run into the same problem.
This makes it work on CPU at least
w_a = torch.moveaxis(w_a, axis, 0).reshape((n, -1)).to(torch.float32)
w_b = torch.moveaxis(w_b, axis, 0).reshape((n, -1)).to(torch.float32)
Latest merge #7 should have fixed this
Hey there, trying out this merge implementation for the first time but running this consistent issue. Seems to happen for many different combinations of model merges. Both novel based and sd-v1-4 based models.
I should also mention that I am using the "device id" pull request https://github.com/ogkalu2/Merge-Stable-Diffusion-models-without-distortion/pull/4, but even without that I get the same error.