[Feature Request]: Make Soft Inpainting Compatible With Mac Apple Silicon

[hsm207]

Is there an existing issue for this?

• I have searched the existing issues and checked the recent builds/commits

What would your feature do ?

When following the soft inpainting tutorial at stable diffusion art, the following error is thrown when doing the softinpainting:

TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead.

The reason is due to the following lines that cast an mps tensor into a float 64:

• stable-diffusion-webui/extensions-builtin/soft-inpainting/scripts/soft_inpainting.py

  Line 82 in 1564ffa

  current_magnitude = torch.norm(image_interp, p=2, dim=1, keepdim=True).to(torch.float64).add_(0.00001)

• stable-diffusion-webui/extensions-builtin/soft-inpainting/scripts/soft_inpainting.py

  Line 85 in 1564ffa

  a_magnitude = torch.norm(a, p=2, dim=1, keepdim=True).to(torch.float64).pow_(

• stable-diffusion-webui/extensions-builtin/soft-inpainting/scripts/soft_inpainting.py

  Line 87 in 1564ffa

  b_magnitude = torch.norm(b, p=2, dim=1, keepdim=True).to(torch.float64).pow_(

Changing to(torch.float64) to to(torch.float32) fixes the problem.

We should consider automating the casting to float64 or float32 depending on the users machine and warn user of the potential problems when casting can only be done to float32.

Proposed workflow

1. User starts the web ui
2. Web ui checks the architecture it is running on
3. If it is an aarm64, warn user
4. Any casting is done without user intervention behind the scenes

Additional information

Might have some relation to #13798 and #12907