fix(peft_utils): raise actionable ValueError when rank_dict is empty in get_peft_kwargs

src/diffusers/utils/peft_utils.py

@@ -155,6 +155,31 @@ def get_peft_kwargs(
 ):
     rank_pattern = {}
     alpha_pattern = {}
+    if not rank_dict:
+        # rank_dict is populated by the caller by walking the model's named_modules
+        # and probing the state_dict for `{name}.lora_B.weight` keys (see e.g.
+        # `_load_lora_into_text_encoder` and `load_lora_into_unet`). When the
+        # state_dict keys do not match that pattern (typically because of a
+        # missing or extra prefix on the saved keys, an adapter-name infix such
+        # as `.default_0.` between `lora_B` and `weight`, or a non-diffusers
+        # serialization format that was not converted upstream), `rank_dict`
+        # ends up empty and we would crash here with a cryptic IndexError on
+        # `list(rank_dict.values())[0]`. Surface the actual problem instead so
+        # the caller can debug the key mismatch. See issue #3238 on huggingface/peft
+        # (the original report was filed against peft, but the failure path is
+        # this function in diffusers).
+        n_keys = len(peft_state_dict) if peft_state_dict is not None else 0
+        sample_keys = list(peft_state_dict.keys())[:3] if peft_state_dict else []
+        raise ValueError(
+            "Could not extract LoRA rank: `rank_dict` is empty. This means none of the "
+            "expected `{module_name}.lora_B.weight` keys were found in the state_dict. "
+            "Usual causes: the saved keys carry an extra or missing prefix versus the "
+            "target model (e.g. `text_model.encoder.*` vs `encoder.*`); the keys carry "
+            "an adapter-name infix such as `.default_0.` between `lora_B` and `weight`; "
+            "or the state_dict was saved in a format that diffusers does not yet "
+            "convert. "
+            f"State dict has {n_keys} keys; first 3: {sample_keys}."
+        )
     r = lora_alpha = list(rank_dict.values())[0]
 
     if len(set(rank_dict.values())) > 1:

tests/others/test_peft_utils.py

@@ -0,0 +1,85 @@
+# coding=utf-8
+# Copyright 2025 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+from diffusers.utils.peft_utils import get_peft_kwargs
+
+
+class GetPeftKwargsTest(unittest.TestCase):
+    """Tests for diffusers.utils.peft_utils.get_peft_kwargs."""
+
+    def test_empty_rank_dict_raises_actionable_value_error(self):
+        """Regression for huggingface/peft#3238 (failure path is in diffusers,
+        not peft). When the caller's rank-discovery loop produces an empty
+        `rank_dict` (typical when state_dict keys carry an extra/missing
+        prefix or an adapter-name infix that the loop did not match), we
+        used to crash with a cryptic IndexError on
+        `list(rank_dict.values())[0]`. Now we raise a `ValueError` whose
+        message names the underlying mismatch and shows a few state_dict
+        keys so the user can diagnose.
+        """
+        peft_state_dict = {
+            "encoder.layers.0.self_attn.k_proj.lora_A.default_0.weight": object(),
+            "encoder.layers.0.self_attn.k_proj.lora_B.default_0.weight": object(),
+            "encoder.layers.0.self_attn.q_proj.lora_A.default_0.weight": object(),
+        }
+        with self.assertRaises(ValueError) as cm:
+            get_peft_kwargs(
+                rank_dict={},
+                network_alpha_dict=None,
+                peft_state_dict=peft_state_dict,
+                is_unet=False,
+            )
+        message = str(cm.exception)
+        self.assertIn("`rank_dict` is empty", message)
+        self.assertIn("lora_B.weight", message)
+        # The message includes a sample of the state_dict so the user can spot
+        # the prefix/infix mismatch from the error alone.
+        self.assertIn("State dict has 3 keys", message)
+
+    def test_empty_rank_dict_with_none_state_dict_is_safe(self):
+        """The diagnostic message should not crash on a None peft_state_dict."""
+        with self.assertRaises(ValueError) as cm:
+            get_peft_kwargs(
+                rank_dict={},
+                network_alpha_dict=None,
+                peft_state_dict=None,
+                is_unet=True,
+            )
+        self.assertIn("State dict has 0 keys", str(cm.exception))
+
+    def test_non_empty_rank_dict_unchanged(self):
+        """The fast-path (rank_dict populated as before) must remain
+        functionally identical. Smoke-test that get_peft_kwargs returns the
+        expected keys for a minimal one-module rank_dict.
+        """
+        rank_dict = {"q_proj.lora_B.weight": 4}
+        peft_state_dict = {
+            "q_proj.lora_A.weight": object(),
+            "q_proj.lora_B.weight": object(),
+        }
+        kwargs = get_peft_kwargs(
+            rank_dict=rank_dict,
+            network_alpha_dict=None,
+            peft_state_dict=peft_state_dict,
+            is_unet=True,
+        )
+        self.assertEqual(kwargs["r"], 4)
+        self.assertEqual(kwargs["lora_alpha"], 4)
+        self.assertEqual(kwargs["rank_pattern"], {})
+        self.assertIn("q_proj", kwargs["target_modules"])
+        self.assertFalse(kwargs["use_dora"])
+        self.assertFalse(kwargs["lora_bias"])