[LoRA, Xformers] Fix xformers lora (#5201)

* fix xformers lora

* improve

* fix

src/diffusers/models/attention_processor.py

@@ -310,19 +310,16 @@ class Attention(nn.Module):
 
         self.set_processor(processor)
 
-    def set_processor(self, processor: "AttnProcessor"):
-        if (
-            hasattr(self, "processor")
-            and not isinstance(processor, LORA_ATTENTION_PROCESSORS)
-            and self.to_q.lora_layer is not None
-        ):
+    def set_processor(self, processor: "AttnProcessor", _remove_lora=False):
+        if hasattr(self, "processor") and _remove_lora and self.to_q.lora_layer is not None:
             deprecate(
                 "set_processor to offload LoRA",
                 "0.26.0",
-                "In detail, removing LoRA layers via calling `set_processor` or `set_default_attn_processor` is deprecated. Please make sure to call `pipe.unload_lora_weights()` instead.",
+                "In detail, removing LoRA layers via calling `set_default_attn_processor` is deprecated. Please make sure to call `pipe.unload_lora_weights()` instead.",
             )
             # TODO(Patrick, Sayak) - this can be deprecated once PEFT LoRA integration is complete
             # We need to remove all LoRA layers
+            # Don't forget to remove ALL `_remove_lora` from the codebase
             for module in self.modules():
                 if hasattr(module, "set_lora_layer"):
                     module.set_lora_layer(None)

src/diffusers/models/autoencoder_kl.py

@@ -196,7 +196,9 @@ class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
         return processors
 
     # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
         r"""
         Sets the attention processor to use to compute attention.
 
@@ -220,9 +222,9 @@ class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
         def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
             if hasattr(module, "set_processor"):
                 if not isinstance(processor, dict):
-                    module.set_processor(processor)
+                    module.set_processor(processor, _remove_lora=_remove_lora)
                 else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
 
             for sub_name, child in module.named_children():
                 fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
@@ -244,7 +246,7 @@ class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
                 f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
             )
 
-        self.set_attn_processor(processor)
+        self.set_attn_processor(processor, _remove_lora=True)
 
     @apply_forward_hook
     def encode(self, x: torch.FloatTensor, return_dict: bool = True) -> AutoencoderKLOutput:

src/diffusers/models/controlnet.py

@@ -517,7 +517,9 @@ class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalControlnetMixin):
         return processors
 
     # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
         r"""
         Sets the attention processor to use to compute attention.
 
@@ -541,9 +543,9 @@ class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalControlnetMixin):
         def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
             if hasattr(module, "set_processor"):
                 if not isinstance(processor, dict):
-                    module.set_processor(processor)
+                    module.set_processor(processor, _remove_lora=_remove_lora)
                 else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
 
             for sub_name, child in module.named_children():
                 fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
@@ -565,7 +567,7 @@ class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalControlnetMixin):
                 f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
             )
 
-        self.set_attn_processor(processor)
+        self.set_attn_processor(processor, _remove_lora=True)
 
     # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attention_slice
     def set_attention_slice(self, slice_size):

src/diffusers/models/prior_transformer.py

@@ -192,7 +192,9 @@ class PriorTransformer(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
         return processors
 
     # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
         r"""
         Sets the attention processor to use to compute attention.
 
@@ -216,9 +218,9 @@ class PriorTransformer(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
         def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
             if hasattr(module, "set_processor"):
                 if not isinstance(processor, dict):
-                    module.set_processor(processor)
+                    module.set_processor(processor, _remove_lora=_remove_lora)
                 else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
 
             for sub_name, child in module.named_children():
                 fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
@@ -240,7 +242,7 @@ class PriorTransformer(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
                 f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
             )
 
-        self.set_attn_processor(processor)
+        self.set_attn_processor(processor, _remove_lora=True)
 
     def forward(
         self,

src/diffusers/models/unet_2d_condition.py

@@ -613,7 +613,9 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
 
         return processors
 
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
         r"""
         Sets the attention processor to use to compute attention.
 
@@ -637,9 +639,9 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
             if hasattr(module, "set_processor"):
                 if not isinstance(processor, dict):
-                    module.set_processor(processor)
+                    module.set_processor(processor, _remove_lora=_remove_lora)
                 else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
 
             for sub_name, child in module.named_children():
                 fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
@@ -660,7 +662,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
                 f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
             )
 
-        self.set_attn_processor(processor)
+        self.set_attn_processor(processor, _remove_lora=True)
 
     def set_attention_slice(self, slice_size):
         r"""

src/diffusers/models/unet_3d_condition.py

@@ -366,7 +366,9 @@ class UNet3DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
             fn_recursive_set_attention_slice(module, reversed_slice_size)
 
     # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
         r"""
         Sets the attention processor to use to compute attention.
 
@@ -390,9 +392,9 @@ class UNet3DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
             if hasattr(module, "set_processor"):
                 if not isinstance(processor, dict):
-                    module.set_processor(processor)
+                    module.set_processor(processor, _remove_lora=_remove_lora)
                 else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
 
             for sub_name, child in module.named_children():
                 fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
@@ -454,7 +456,7 @@ class UNet3DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
                 f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
             )
 
-        self.set_attn_processor(processor)
+        self.set_attn_processor(processor, _remove_lora=True)
 
     def _set_gradient_checkpointing(self, module, value=False):
         if isinstance(module, (CrossAttnDownBlock3D, DownBlock3D, CrossAttnUpBlock3D, UpBlock3D)):

src/diffusers/pipelines/audioldm2/modeling_audioldm2.py

@@ -538,7 +538,9 @@ class AudioLDM2UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoad
         return processors
 
     # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
         r"""
         Sets the attention processor to use to compute attention.
 
@@ -562,9 +564,9 @@ class AudioLDM2UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoad
         def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
             if hasattr(module, "set_processor"):
                 if not isinstance(processor, dict):
-                    module.set_processor(processor)
+                    module.set_processor(processor, _remove_lora=_remove_lora)
                 else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
 
             for sub_name, child in module.named_children():
                 fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
@@ -586,7 +588,7 @@ class AudioLDM2UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoad
                 f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
             )
 
-        self.set_attn_processor(processor)
+        self.set_attn_processor(processor, _remove_lora=True)
 
     # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attention_slice
     def set_attention_slice(self, slice_size):

src/diffusers/pipelines/versatile_diffusion/modeling_text_unet.py

@@ -820,7 +820,9 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
 
         return processors
 
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
         r"""
         Sets the attention processor to use to compute attention.
 
@@ -844,9 +846,9 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
         def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
             if hasattr(module, "set_processor"):
                 if not isinstance(processor, dict):
-                    module.set_processor(processor)
+                    module.set_processor(processor, _remove_lora=_remove_lora)
                 else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
 
             for sub_name, child in module.named_children():
                 fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
@@ -868,7 +870,7 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
                 f" {next(iter(self.attn_processors.values()))}"
             )
 
-        self.set_attn_processor(processor)
+        self.set_attn_processor(processor, _remove_lora=True)
 
     def set_attention_slice(self, slice_size):
         r"""

tests/lora/test_lora_layers_old_backend.py

@@ -1786,7 +1786,7 @@ class UNet3DConditionModelTests(unittest.TestCase):
         with torch.no_grad():
             sample = model(**inputs_dict, cross_attention_kwargs={"scale": 0.0}).sample
 
-        model.set_attn_processor(AttnProcessor())
+        model.set_default_attn_processor()
 
         with torch.no_grad():
             new_sample = model(**inputs_dict).sample