Fix kohya FLUX CLIP text-encoder LoRA loading under transformers>=5 (#13984)

src/diffusers/loaders/lora_base.py

@@ -376,6 +376,14 @@ def _load_lora_into_text_encoder(
         # convert state dict
         state_dict = convert_state_dict_to_peft(state_dict)
 
+        # Under transformers>=5, CLIPTextModel was flattened and no longer has the
+        # `text_model.` wrapper module, so named_modules() returns unprefixed names
+        # like "encoder.layers.0.self_attn.q_proj". Kohya-sourced LoRA state dict
+        # keys still carry the "text_model." prefix from the old module layout.
+        # Strip it so that rank-key matching works regardless of the transformers version.
+        if not hasattr(text_encoder, "text_model"):
+            state_dict = {k.removeprefix("text_model."): v for k, v in state_dict.items()}
+
         for name, _ in text_encoder.named_modules():
             if name.endswith((".q_proj", ".k_proj", ".v_proj", ".out_proj", ".fc1", ".fc2")):
                 rank_key = f"{name}.lora_B.weight"

tests/lora/test_lora_layers_flux.py

@@ -256,6 +256,48 @@ def test_lora_expansion_works_for_extra_keys(self):
             "LoRA should lead to different results.",
         )
 
+    def test_kohya_clip_text_encoder_flattened_compat(self):
+        """Regression test for #13984: kohya FLUX CLIP text-encoder LoRA under transformers>=5.
+
+        Under transformers>=5, CLIPTextModel was flattened and no longer has the `text_model.`
+        wrapper module. Kohya-sourced state dict keys still carry the "text_model." prefix,
+        which caused an empty rank dict and IndexError. This test verifies the fix by passing
+        a synthetic kohya-style state dict with the stale prefix to a text encoder that
+        doesn't have the `text_model` submodule.
+        """
+        components, _, _ = self.get_dummy_components(FlowMatchEulerDiscreteScheduler)
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        # Simulate transformers>=5 flattened CLIPTextModel by removing the text_model
+        # wrapper module attribute. Under transformers>=5 named_modules() no longer
+        # includes this prefix.
+        if hasattr(pipe.text_encoder, "text_model"):
+            del pipe.text_encoder.text_model
+
+        # Build a synthetic kohya-style state dict whose text-encoder keys carry the stale
+        # "text_model." prefix that real kohya conversion produces.
+        state_dict = {}
+        with torch.no_grad():
+            for name, module in pipe.text_encoder.named_modules():
+                if name.endswith((".q_proj", ".k_proj", ".v_proj", ".out_proj", ".fc1", ".fc2")):
+                    # Kohya conversion emits keys with text_model. prefix
+                    kohya_key_b = f"text_encoder.text_model.{name}.lora_B.weight"
+                    kohya_key_a = f"text_encoder.text_model.{name}.lora_A.weight"
+                    state_dict[kohya_key_a] = torch.randn(2, module.weight.shape[0])
+                    state_dict[kohya_key_b] = torch.randn(module.weight.shape[1], 2)
+
+        # This should not raise IndexError (the pre-fix crash point)
+        pipe.load_lora_weights(state_dict)
+
+        self.assertTrue(
+            check_if_lora_correctly_set(pipe.text_encoder),
+            "LoRA should be correctly set in text encoder under transformers>=5 layout",
+        )
+
+        pipe.unload_lora_weights()
+
     @unittest.skip("Not supported in Flux.")
     def test_simple_inference_with_text_denoiser_block_scale(self):
         pass