[patch_pipe] UnboundLocalError: local variable '_tmp' referenced before assignment

[gracopordeus]

I'm trying to work in kaggle notebooks.

File /opt/conda/lib/python3.10/site-packages/lora_diffusion/lora.py:1012, in patch_pipe(pipe, maybe_unet_path, token, r, patch_unet, patch_text, patch_ti, idempotent_token, unet_target_replace_module, text_target_replace_module)
   1010 elif maybe_unet_path.endswith(".safetensors"):
   1011     safeloras = safe_open(maybe_unet_path, framework="pt", device="cpu")
-> 1012     monkeypatch_or_replace_safeloras(pipe, safeloras)
   1013     tok_dict = parse_safeloras_embeds(safeloras)
   1014     if patch_ti:

File /opt/conda/lib/python3.10/site-packages/lora_diffusion/lora.py:809, in monkeypatch_or_replace_safeloras(models, safeloras)
    806     print(f"No model provided for {name}, contained in Lora")
    807     continue
--> 809 monkeypatch_or_replace_lora_extended(model, lora, target, ranks)

File /opt/conda/lib/python3.10/site-packages/lora_diffusion/lora.py:784, in monkeypatch_or_replace_lora_extended(model, loras, target_replace_module, r)
    781         _tmp.conv.bias = bias
    783 # switch the module
--> 784 _module._modules[name] = _tmp
    786 up_weight = loras.pop(0)
    787 down_weight = loras.pop(0)

UnboundLocalError: local variable '_tmp' referenced before assignment

This is the function called:

patch_pipe(
    pipe,
    os.getcwd()+"/lora/example_loras/lora_illust.safetensors",
    patch_text=True,
    patch_ti=True,
    patch_unet=True,
)

[justin-prnd]

If you can modify the source code this might be helpful. But if not, downgrade the version of diffusers to v0.16.0 where the Attention module uses nn.Linear rather than the LoRACompatibleLinear which causes the issue above.