Cannot extract Lora from Safetensors model trained in Kohya, model also produces NaNs in Automatic1111

Everything goes allright until SVD comes in:

`calculating by svd
0%| | 0/278 [00:00
svd(args)
File "C:\AI\kohya_ss\networks\extract_lora_from_models.py", line 116, in svd
U, S, Vh = torch.linalg.svd(mat)
torch._C._LinAlgError: linalg.svd: The algorithm failed to converge because the input matrix is ill-conditioned or has too many repeated singular values (error code: 319).`

Changing the precision doesn't help, neither does changing to cpu.

Start at networks/extract_lora_from_models.py, especially the svd entry point and torch.linalg.svd call at line 116. Reproduce the Kohya-trained Safetensors extraction failure and inspect the matrix values and convergence warning; done means the extraction completes without the reported SVD error.

Written by the indexing model from the issue text.