kalle07
commented
1 week ago and btw you could check if its realy right RANK and Network Alpha and more importand if float and fp16 is set right for me Save precision Fp16, RANK 32, Alpha 4 -> filesize ~300MB on civitai ( idont know how they save?) RANK 32, Alpha 4 -> filesize ~38MB
its must be a bug ;) why is it such as slow from float to fp16
i use CUDA 16GB Vram (8GB used) on SDXL it was maybe one minute FLUX is 20min
and no check if new file name right:
2024-09-11 16:29:11 INFO resizing complete resize_lora.py:294 INFO saving model to: .safetensors resize_lora.py:356 Traceback (most recent call last): File "e:\kohya_ss\sd-scripts\networks\resize_lora.py", line 411, in
resize(args)
File "e:\kohya_ss\sd-scripts\networks\resize_lora.py", line 357, in resize
save_to_file(args.save_to, state_dict, save_dtype, metadata)
File "e:\kohya_ss\sd-scripts\networks\resize_lora.py", line 51, in save_to_file
torch.save(state_dict, file_name)
File "e:\kohya_ss\venv\lib\site-packages\torch\serialization.py", line 651, in save
with _open_zipfile_writer(f) as opened_zipfile:
File "e:\kohya_ss\venv\lib\site-packages\torch\serialization.py", line 525, in _open_zipfile_writer
return container(name_or_buffer)
File "e:\kohya_ss\venv\lib\site-packages\torch\serialization.py", line 496, in init
super().init(torch._C.PyTorchFileWriter(self.name))
RuntimeError: [enforce fail at inline_container.cc:642] . invalid file name: .safetensors
16:29:12-773619 INFO Done resizing...
and no Stop Button
i mean the resized lora works ...