git/InvokeAI
1
0
Fork 0
mirror of https://github.com/invoke-ai/InvokeAI synced 2026-04-28 09:41:54 +02:00
Code Issues Packages Projects Releases Wiki Activity

fix: SDXL DoRA LoRA fails with enable_partial_loading=true

Browse Source 
cast_to_device returns plain torch.Tensor instead of torch.nn.Parameter,
causing _aggregate_patch_parameters to replace valid weights with meta
device dummies, falsely triggering DoRA's quantization guard.

Fixes invoke-ai/InvokeAI#8624
This commit is contained in:
Alexander Eichhorn 2026-04-17 15:23:12 +02:00
parent 7eaf1d5bd0
commit a0a87212a0
1 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
invokeai/backend/model_manager/load/model_cache/torch_module_autocast/custom_modules/custom_module_mixin.py
View File

@ -49,7 +49,10 @@ class CustomModuleMixin:
# parameters. But, of course, any sub-layers that need to access the actual values of the parameters will fail.
for param_name in orig_params.keys():
param = orig_params[param_name]
if type(param) is torch.nn.Parameter and type(param.data) is torch.Tensor:
if type(param) is torch.Tensor:
# Plain tensor (e.g. after cast_to_device moved a Parameter to another device).
pass
elif type(param) is torch.nn.Parameter and type(param.data) is torch.Tensor:
pass
elif type(param) is GGMLTensor:
# Move to device and dequantize here. Doing it in the patch layer can result in redundant casts /