the following fix is required for apple silicon, note the additional .float(), else the forge script will give this error
File "/Applications/Data/Packages/Stable Diffusion WebUI Forge/extensions/sd-forge-layerdiffuse/lib_layerdiffusion/models.py", line 277, in estimate_augmented
median = torch.median(result.cpu(), dim=0).values
RuntimeError: "median_out" not implemented for 'Half'
in lib_layerdiffusion/models.py
if self.load_device == torch.device("mps"):
'''
In case that apple silicon devices would crash when calling torch.median() on tensors
in gpu vram with dimensions higher than 4, we move it to cpu, call torch.median()
and then move the result back to gpu.
'''
result_cpu = result.cpu().float() # Convert to float32 on CPU
median = torch.median(result_cpu, dim=0).values
median = median.to(device=self.load_device, dtype=self.dtype)
else:
median = torch.median(result, dim=0).values
the following fix is required for apple silicon, note the additional .float(), else the forge script will give this error
in lib_layerdiffusion/models.py