It just needs a max abs selection to be perfect! :)
Extraltodeus opened this issue · comments
I used to do this:
def furthest_from_zero(tensors):
shape = tensors.shape
tensors = tensors.reshape(shape[0], -1)
tensors_abs = torch.abs(tensors)
max_abs_idx = torch.argmax(tensors_abs, dim=0)
result = tensors[max_abs_idx, torch.arange(tensors.shape[1])]
return result.reshape(shape[1:])
With .ckpt files to select the maximum absolute values from each models/layers. Now I'm a bit lost because I don't know how to do this with safetensors (I'm a self learner). Would you be able to integrate such merging method in your node? This would be really nice :)
edit: hey I just noticed that you made the WAS suite, that explains why this node is so elegantly written. Thank you for everything!
Thanks you.
Also I am just sending the ratios to ModelMergeBlocksNumber which does the merging. I think this would be something to bring up in base ConfyUI as a option for merging.
Thank you so much for your answer.
I guess I will have to find a way to make it happen :)