wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-03-30 21:43:43 +08:00
Code Issues Actions 13 Packages Projects Releases Wiki Activity

Fix QuantizedTensor weight restore failure causing progressive drift

Browse Source 
When weight_inplace_update=True and a set_func (e.g. fp8_scaled ops)
replaces the parameter with a new QuantizedTensor, the backup records
inplace_update=True.  On restore, unpatch_model uses copy_to_param
which calls param.data.copy_().  QuantizedTensor.__torch_dispatch__
routes copy_() through _dequant_and_fallback, which copies into a
temporary float tensor without updating the underlying quantized data.
The weight silently remains LoRA-patched after "restore", and the next
generation backs up the already-patched weight and applies LoRA again,
causing progressive quality degradation (washout).

Fix: when set_func is present, force backup_inplace=False so restore
uses set_attr_param (parameter replacement) instead of copy_to_param
(in-place copy).  set_attr_param correctly replaces the parameter
object, which works for both regular tensors and QuantizedTensors.

Fixes #11021
This commit is contained in:
Isaac Bender 2026-03-15 14:52:30 -07:00
parent 3814bf4454
commit 19eb196941
1 changed files with 9 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
comfy/model_patcher.py
View File

@ -692,7 +692,15 @@ class ModelPatcher:
inplace_update = self.weight_inplace_update or inplace_update
if key not in self.backup and not return_weight:
self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update)
# When set_func is present (e.g. QuantizedTensor/fp8_scaled ops), it replaces
# the parameter object rather than modifying it in-place. The restore path
# must therefore also replace the parameter (set_attr_param) instead of doing
# an in-place copy (copy_to_param), because QuantizedTensor.__torch_dispatch__
# routes copy_() through dequant-and-fallback which silently fails to update
# the underlying quantized data. Force inplace_update=False in the backup
# for these keys so unpatch_model uses set_attr_param for restoration.
backup_inplace = inplace_update if set_func is None else False
self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), backup_inplace)
temp_dtype = comfy.model_management.lora_compute_dtype(device_to)
if device_to is not None: