up

Author: sayakpaul

setup.py

@@ -116,7 +116,8 @@
     "librosa",
     "numpy",
     "parameterized",
-    "peft>=0.16.1",
+    "peft>=0.15.0",
+    # "peft>=0.16.1",
     "protobuf>=3.20.3,<4",
     "pytest",
     "pytest-timeout",

src/diffusers/dependency_versions_table.py

@@ -23,7 +23,7 @@
     "librosa": "librosa",
     "numpy": "numpy",
     "parameterized": "parameterized",
-    "peft": "peft>=0.16.1",
+    "peft": "peft>=0.15.0",
     "protobuf": "protobuf>=3.20.3,<4",
     "pytest": "pytest",
     "pytest-timeout": "pytest-timeout",

tests/models/transformers/test_models_transformer_flux.py

@@ -172,6 +172,32 @@ def test_gradient_checkpointing_is_applied(self):
         expected_set = {"FluxTransformer2DModel"}
         super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
 
+    # The test exists for cases like
+    # https://github.com/huggingface/diffusers/issues/11874
+    def test_lora_exclude_modules(self):
+        from peft import LoraConfig, get_peft_model_state_dict, inject_adapter_in_model, set_peft_model_state_dict
+
+        lora_rank = 4
+        target_module = "single_transformer_blocks.0.proj_out"
+        adapter_name = "foo"
+        init_dict, _ = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        state_dict = model.state_dict()
+        target_mod_shape = state_dict[f"{target_module}.weight"].shape
+        lora_state_dict = {
+            f"{target_module}.lora_A.weight": torch.ones(lora_rank, target_mod_shape[1]) * 22,
+            f"{target_module}.lora_B.weight": torch.ones(target_mod_shape[0], lora_rank) * 33,
+        }
+        config = LoraConfig(
+            r=lora_rank, target_modules=["single_transformer_blocks.0.proj_out"], exclude_modules=["proj_out"]
+        )
+        inject_adapter_in_model(config, model, adapter_name=adapter_name, state_dict=lora_state_dict)
+        set_peft_model_state_dict(model, lora_state_dict, adapter_name)
+        retrieved_lora_state_dict = get_peft_model_state_dict(model, adapter_name=adapter_name)
+        assert (retrieved_lora_state_dict["single_transformer_blocks.0.proj_out.lora_A.weight"] == 22).all()
+        assert (retrieved_lora_state_dict["single_transformer_blocks.0.proj_out.lora_B.weight"] == 33).all()
+
 
 class FluxTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCase):
     model_class = FluxTransformer2DModel