observantdistressed commited on
Commit
07b1e48
·
1 Parent(s): 429d0f4

Fix LoRA load failure for Wan-native diff_b format files

Browse files

Production log: 'iGoon - Blink_Front_Doggystyle_I2V' and 'Ultimate
Deepthroat' failed to load with 'TorchaoLoraLinear does not support
lora_bias yet', and generation proceeded WITHOUT their weights while the
prompt tags still claimed they were active.

Root cause: these LoRAs are saved in the Wan-native diff_b format (447
'diffusion_model.*.diff_b' learned-bias keys per file). diffusers'
_convert_non_diffusers_wan_lora_to_diffusers maps each diff_b key to
'lora_B.bias', which flips lora_bias=True, which peft's TorchaoLoraLinear
hard-rejects on torchao-quantized layers. The previous strip only removed
literal 'lora_B...bias' keys from already-converted dicts, missing the
Wan-native form entirely.

Fix: _strip_lora_bias now also drops keys ending in '.diff_b'. Verified
against the real production files (447/447 bias terms stripped, weights
preserved) and added a mock-pipe test case for the Wan-native format.

Regression: 5/5 mock-pipe cases, 18/18 auto-activation matrix, py_compile
all pass.

Files changed (1) hide show
  1. lora_loader.py +13 -8
lora_loader.py CHANGED
@@ -1008,17 +1008,22 @@ def download_lora(group_name):
1008
 
1009
 
1010
  def _strip_lora_bias(state_dict):
1011
- """Drop lora_B.*.bias keys that peft's TorchaoLoraLinear cannot support.
1012
-
1013
- diffusers auto-detects lora_bias=True purely from the presence of
1014
- "lora_B" ... ".bias" state-dict keys (src/diffusers/utils/peft_utils.py),
1015
- and peft's TorchaoLoraLinear.__init__ hard-rejects lora_bias=True on
1016
- torchao-quantized layers. Removing the keys makes lora_bias=False and the
1017
- adapter loads fine; only the (small) learned bias delta is discarded.
 
 
 
 
 
1018
  """
1019
  return {
1020
  k: v for k, v in state_dict.items()
1021
- if not ("lora_B" in k and k.endswith(".bias"))
1022
  }
1023
 
1024
 
 
1008
 
1009
 
1010
  def _strip_lora_bias(state_dict):
1011
+ """Drop keys that would make peft's TorchaoLoraLinear reject the adapter.
1012
+
1013
+ Two sources of a learned bias term:
1014
+ 1. Already-converted dicts: keys containing "lora_B" ending in ".bias"
1015
+ (diffusers' lora_bias auto-detect looks for these in
1016
+ src/diffusers/utils/peft_utils.py).
1017
+ 2. Wan-native dicts: keys ending in ".diff_b". diffusers'
1018
+ _convert_non_diffusers_wan_lora_to_diffusers maps every diff_b key to a
1019
+ "lora_B.bias" key, tripping the same lora_bias=True path.
1020
+
1021
+ Both are dropped so lora_bias stays False; only the (small) learned bias
1022
+ delta is discarded.
1023
  """
1024
  return {
1025
  k: v for k, v in state_dict.items()
1026
+ if not (k.endswith(".diff_b") or ("lora_B" in k and k.endswith(".bias")))
1027
  }
1028
 
1029