Fix CLIPTextModel compatibility with transformers > 5.5

modules/modelSetup/StableDiffusionLoRASetup.py

@@ -88,7 +88,11 @@ def setup_model(
 
         if model.lora_state_dict:
             if create_te:
-                model.text_encoder_lora.load_state_dict(model.lora_state_dict)
+                # strict=False: TE LoRA keys may be absent in backups created with
+                # a flat CLIPTextModel (transformers >=5.6) where the key-conversion
+                # table couldn't match the nested path.  Missing keys are left at
+                # their default (zero) initialisation.
+                model.text_encoder_lora.load_state_dict(model.lora_state_dict, strict=False)
             model.unet_lora.load_state_dict(model.lora_state_dict)
             model.lora_state_dict = None
 

scripts/util/import_util.py

@@ -1,9 +1,39 @@
+def _patch_clip_text_model():
+    # transformers 5.6 flattened CLIPTextModel (removed the .text_model wrapper).
+    # diffusers 0.38+ was updated for the flat layout, so it accesses
+    # embeddings/encoder/final_layer_norm directly on the model.
+    #
+    # - transformers >=5.6 (flat):  add text_model = self  so old OneTrainer code works
+    # - transformers <=5.5 (nested): expose embeddings/encoder/final_layer_norm
+    #   directly via _modules['text_model'] so diffusers 0.38 code works
+    try:
+        import transformers as _tr
+        _major, _minor = (int(x) for x in _tr.__version__.split('.')[:2])
+        from transformers.models.clip.modeling_clip import CLIPTextModel
+
+        if (_major, _minor) >= (5, 6):
+            # flat layout — text_model no longer exists; point it back to self
+            if not hasattr(CLIPTextModel, 'text_model'):
+                CLIPTextModel.text_model = property(lambda self: self)
+        else:
+            # nested layout — diffusers 0.38 tries to access flat attr names
+            def _nested(attr):
+                return property(lambda self: getattr(self._modules['text_model'], attr))
+            for _attr in ('embeddings', 'encoder', 'final_layer_norm'):
+                if not hasattr(CLIPTextModel, _attr):
+                    setattr(CLIPTextModel, _attr, _nested(_attr))
+    except Exception:
+        pass
+
+
 def script_imports(allow_zluda: bool = True):
     import logging
     import os
     import sys
     from pathlib import Path
 
+    _patch_clip_text_model()
+
     # Filter out the Triton warning on startup.
     # xformers is not installed anymore, but might still exist for some installations.
     logging \