[core] reuse `AttentionMixin` for compatible classes (#12463)

* remove attn_processors property

* more

* up

* up more.

* up

* add AttentionMixin to AuraFlow.

* up

* up

* up

* up

src/diffusers/models/autoencoders/autoencoder_kl.py

@@ -11,7 +11,7 @@
 # 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.
-from typing import Dict, Optional, Tuple, Union
+from typing import Optional, Tuple, Union
 
 import torch
 import torch.nn as nn
@@ -21,11 +21,11 @@ from ...loaders import PeftAdapterMixin
 from ...loaders.single_file_model import FromOriginalModelMixin
 from ...utils import deprecate
 from ...utils.accelerate_utils import apply_forward_hook
+from ..attention import AttentionMixin
 from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
     Attention,
-    AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
     FusedAttnProcessor2_0,
@@ -35,7 +35,9 @@ from ..modeling_utils import ModelMixin
 from .vae import AutoencoderMixin, Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder
 
 
-class AutoencoderKL(ModelMixin, AutoencoderMixin, ConfigMixin, FromOriginalModelMixin, PeftAdapterMixin):
+class AutoencoderKL(
+    ModelMixin, AttentionMixin, AutoencoderMixin, ConfigMixin, FromOriginalModelMixin, PeftAdapterMixin
+):
     r"""
     A VAE model with KL loss for encoding images into latents and decoding latent representations into images.
 
@@ -138,66 +140,6 @@ class AutoencoderKL(ModelMixin, AutoencoderMixin, ConfigMixin, FromOriginalModel
         self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
         self.tile_overlap_factor = 0.25
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
     def set_default_attn_processor(self):
         """

src/diffusers/models/autoencoders/autoencoder_kl_flux2.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import math
-from typing import Dict, Optional, Tuple, Union
+from typing import Optional, Tuple, Union
 
 import torch
 import torch.nn as nn
@@ -22,11 +22,11 @@ from ...loaders import PeftAdapterMixin
 from ...loaders.single_file_model import FromOriginalModelMixin
 from ...utils import deprecate
 from ...utils.accelerate_utils import apply_forward_hook
+from ..attention import AttentionMixin
 from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
     Attention,
-    AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
     FusedAttnProcessor2_0,
@@ -36,7 +36,9 @@ from ..modeling_utils import ModelMixin
 from .vae import AutoencoderMixin, Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder
 
 
-class AutoencoderKLFlux2(ModelMixin, AutoencoderMixin, ConfigMixin, FromOriginalModelMixin, PeftAdapterMixin):
+class AutoencoderKLFlux2(
+    ModelMixin, AutoencoderMixin, AttentionMixin, ConfigMixin, FromOriginalModelMixin, PeftAdapterMixin
+):
     r"""
     A VAE model with KL loss for encoding images into latents and decoding latent representations into images.
 
@@ -154,66 +156,6 @@ class AutoencoderKLFlux2(ModelMixin, AutoencoderMixin, ConfigMixin, FromOriginal
         self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
         self.tile_overlap_factor = 0.25
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
     def set_default_attn_processor(self):
         """

src/diffusers/models/autoencoders/autoencoder_kl_temporal_decoder.py

@@ -12,14 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import itertools
-from typing import Dict, Optional, Tuple, Union
+from typing import Optional, Tuple, Union
 
 import torch
 import torch.nn as nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils.accelerate_utils import apply_forward_hook
-from ..attention_processor import CROSS_ATTENTION_PROCESSORS, AttentionProcessor, AttnProcessor
+from ..attention import AttentionMixin
+from ..attention_processor import CROSS_ATTENTION_PROCESSORS, AttnProcessor
 from ..modeling_outputs import AutoencoderKLOutput
 from ..modeling_utils import ModelMixin
 from ..unets.unet_3d_blocks import MidBlockTemporalDecoder, UpBlockTemporalDecoder
@@ -135,7 +136,7 @@ class TemporalDecoder(nn.Module):
         return sample
 
 
-class AutoencoderKLTemporalDecoder(ModelMixin, AutoencoderMixin, ConfigMixin):
+class AutoencoderKLTemporalDecoder(ModelMixin, AttentionMixin, AutoencoderMixin, ConfigMixin):
     r"""
     A VAE model with KL loss for encoding images into latents and decoding latent representations into images.
 
@@ -202,66 +203,6 @@ class AutoencoderKLTemporalDecoder(ModelMixin, AutoencoderMixin, ConfigMixin):
 
         self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1)
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     def set_default_attn_processor(self):
         """
         Disables custom attention processors and sets the default attention implementation.

src/diffusers/models/autoencoders/consistency_decoder_vae.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from dataclasses import dataclass
-from typing import Dict, Optional, Tuple, Union
+from typing import Optional, Tuple, Union
 
 import torch
 import torch.nn.functional as F
@@ -23,10 +23,10 @@ from ...schedulers import ConsistencyDecoderScheduler
 from ...utils import BaseOutput
 from ...utils.accelerate_utils import apply_forward_hook
 from ...utils.torch_utils import randn_tensor
+from ..attention import AttentionMixin
 from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
-    AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
 )
@@ -49,7 +49,7 @@ class ConsistencyDecoderVAEOutput(BaseOutput):
     latent_dist: "DiagonalGaussianDistribution"
 
 
-class ConsistencyDecoderVAE(ModelMixin, AutoencoderMixin, ConfigMixin):
+class ConsistencyDecoderVAE(ModelMixin, AttentionMixin, AutoencoderMixin, ConfigMixin):
     r"""
     The consistency decoder used with DALL-E 3.
 
@@ -167,66 +167,6 @@ class ConsistencyDecoderVAE(ModelMixin, AutoencoderMixin, ConfigMixin):
         self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
         self.tile_overlap_factor = 0.25
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
     def set_default_attn_processor(self):
         """

src/diffusers/models/controlnets/controlnet.py

@@ -21,10 +21,10 @@ from torch.nn import functional as F
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders.single_file_model import FromOriginalModelMixin
 from ...utils import BaseOutput, logging
+from ..attention import AttentionMixin
 from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
-    AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
 )
@@ -106,7 +106,7 @@ class ControlNetConditioningEmbedding(nn.Module):
         return embedding
 
 
-class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+class ControlNetModel(ModelMixin, AttentionMixin, ConfigMixin, FromOriginalModelMixin):
     """
     A ControlNet model.
 
@@ -515,66 +515,6 @@ class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
 
         return controlnet
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
     def set_default_attn_processor(self):
         """

src/diffusers/models/controlnets/controlnet_flux.py

@@ -21,7 +21,7 @@ import torch.nn as nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import PeftAdapterMixin
 from ...utils import USE_PEFT_BACKEND, BaseOutput, logging, scale_lora_layers, unscale_lora_layers
-from ..attention_processor import AttentionProcessor
+from ..attention import AttentionMixin
 from ..controlnets.controlnet import ControlNetConditioningEmbedding, zero_module
 from ..embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
 from ..modeling_outputs import Transformer2DModelOutput
@@ -38,7 +38,7 @@ class FluxControlNetOutput(BaseOutput):
     controlnet_single_block_samples: Tuple[torch.Tensor]
 
 
-class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
+class FluxControlNetModel(ModelMixin, AttentionMixin, ConfigMixin, PeftAdapterMixin):
     _supports_gradient_checkpointing = True
 
     @register_to_config
@@ -118,66 +118,6 @@ class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
 
         self.gradient_checkpointing = False
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self):
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     @classmethod
     def from_transformer(
         cls,

src/diffusers/models/controlnets/controlnet_qwenimage.py

@@ -21,7 +21,7 @@ import torch.nn as nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
 from ...utils import USE_PEFT_BACKEND, BaseOutput, logging, scale_lora_layers, unscale_lora_layers
-from ..attention_processor import AttentionProcessor
+from ..attention import AttentionMixin
 from ..cache_utils import CacheMixin
 from ..controlnets.controlnet import zero_module
 from ..modeling_outputs import Transformer2DModelOutput
@@ -42,7 +42,9 @@ class QwenImageControlNetOutput(BaseOutput):
     controlnet_block_samples: Tuple[torch.Tensor]
 
 
-class QwenImageControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin):
+class QwenImageControlNetModel(
+    ModelMixin, AttentionMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin
+):
     _supports_gradient_checkpointing = True
 
     @register_to_config
@@ -92,66 +94,6 @@ class QwenImageControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOr
 
         self.gradient_checkpointing = False
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self):
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     @classmethod
     def from_transformer(
         cls,

src/diffusers/models/controlnets/controlnet_sana.py

@@ -21,7 +21,7 @@ from torch import nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import PeftAdapterMixin
 from ...utils import USE_PEFT_BACKEND, BaseOutput, logging, scale_lora_layers, unscale_lora_layers
-from ..attention_processor import AttentionProcessor
+from ..attention import AttentionMixin
 from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
@@ -38,7 +38,7 @@ class SanaControlNetOutput(BaseOutput):
     controlnet_block_samples: Tuple[torch.Tensor]
 
 
-class SanaControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
+class SanaControlNetModel(ModelMixin, AttentionMixin, ConfigMixin, PeftAdapterMixin):
     _supports_gradient_checkpointing = True
     _no_split_modules = ["SanaTransformerBlock", "PatchEmbed"]
     _skip_layerwise_casting_patterns = ["patch_embed", "norm"]
@@ -117,66 +117,6 @@ class SanaControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
 
         self.gradient_checkpointing = False
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     def forward(
         self,
         hidden_states: torch.Tensor,

src/diffusers/models/controlnets/controlnet_sd3.py

@@ -22,8 +22,8 @@ import torch.nn as nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
 from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
-from ..attention import JointTransformerBlock
-from ..attention_processor import Attention, AttentionProcessor, FusedJointAttnProcessor2_0
+from ..attention import AttentionMixin, JointTransformerBlock
+from ..attention_processor import Attention, FusedJointAttnProcessor2_0
 from ..embeddings import CombinedTimestepTextProjEmbeddings, PatchEmbed
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
@@ -39,7 +39,7 @@ class SD3ControlNetOutput(BaseOutput):
     controlnet_block_samples: Tuple[torch.Tensor]
 
 
-class SD3ControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
+class SD3ControlNetModel(ModelMixin, AttentionMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
     r"""
     ControlNet model for [Stable Diffusion 3](https://huggingface.co/papers/2403.03206).
 
@@ -204,66 +204,6 @@ class SD3ControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginal
         for module in self.children():
             fn_recursive_feed_forward(module, chunk_size, dim)
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     # Copied from diffusers.models.transformers.transformer_sd3.SD3Transformer2DModel.fuse_qkv_projections
     def fuse_qkv_projections(self):
         """

src/diffusers/models/controlnets/controlnet_sparsectrl.py

@@ -22,10 +22,10 @@ from torch.nn import functional as F
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin
 from ...utils import BaseOutput, logging
+from ..attention import AttentionMixin
 from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
-    AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
 )
@@ -93,7 +93,7 @@ class SparseControlNetConditioningEmbedding(nn.Module):
         return embedding
 
 
-class SparseControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+class SparseControlNetModel(ModelMixin, AttentionMixin, ConfigMixin, FromOriginalModelMixin):
     """
     A SparseControlNet model as described in [SparseCtrl: Adding Sparse Controls to Text-to-Video Diffusion
     Models](https://huggingface.co/papers/2311.16933).
@@ -448,66 +448,6 @@ class SparseControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
 
         return controlnet
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
     def set_default_attn_processor(self):
         """

src/diffusers/models/controlnets/controlnet_union.py

@@ -19,10 +19,10 @@ from torch import nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders.single_file_model import FromOriginalModelMixin
 from ...utils import logging
+from ..attention import AttentionMixin
 from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
-    AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
 )
@@ -81,7 +81,7 @@ class ResidualAttentionBlock(nn.Module):
         return x
 
 
-class ControlNetUnionModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+class ControlNetUnionModel(ModelMixin, AttentionMixin, ConfigMixin, FromOriginalModelMixin):
     """
     A ControlNetUnion model.
 
@@ -455,66 +455,6 @@ class ControlNetUnionModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
 
         return controlnet
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
     def set_default_attn_processor(self):
         """

src/diffusers/models/controlnets/controlnet_xs.py

@@ -21,11 +21,11 @@ from torch import Tensor, nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils import BaseOutput, logging
 from ...utils.torch_utils import apply_freeu
+from ..attention import AttentionMixin
 from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
     Attention,
-    AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
     FusedAttnProcessor2_0,
@@ -241,7 +241,7 @@ def get_up_block_adapter(
     return UpBlockControlNetXSAdapter(ctrl_to_base=nn.ModuleList(ctrl_to_base))
 
 
-class ControlNetXSAdapter(ModelMixin, ConfigMixin):
+class ControlNetXSAdapter(ModelMixin, AttentionMixin, ConfigMixin):
     r"""
     A `ControlNetXSAdapter` model. To use it, pass it into a `UNetControlNetXSModel` (together with a
     `UNet2DConditionModel` base model).
@@ -508,7 +508,7 @@ class ControlNetXSAdapter(ModelMixin, ConfigMixin):
         )
 
 
-class UNetControlNetXSModel(ModelMixin, ConfigMixin):
+class UNetControlNetXSModel(ModelMixin, AttentionMixin, ConfigMixin):
     r"""
     A UNet fused with a ControlNet-XS adapter model
 
@@ -868,66 +868,6 @@ class UNetControlNetXSModel(ModelMixin, ConfigMixin):
         for u in self.up_blocks:
             u.freeze_base_params()
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
     def set_default_attn_processor(self):
         """

src/diffusers/models/transformers/auraflow_transformer_2d.py

@@ -23,9 +23,9 @@ from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
 from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention import AttentionMixin
 from ..attention_processor import (
     Attention,
-    AttentionProcessor,
     AuraFlowAttnProcessor2_0,
     FusedAuraFlowAttnProcessor2_0,
 )
@@ -275,7 +275,7 @@ class AuraFlowJointTransformerBlock(nn.Module):
         return encoder_hidden_states, hidden_states
 
 
-class AuraFlowTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
+class AuraFlowTransformer2DModel(ModelMixin, AttentionMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
     r"""
     A 2D Transformer model as introduced in AuraFlow (https://blog.fal.ai/auraflow/).
 
@@ -365,66 +365,6 @@ class AuraFlowTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, From
 
         self.gradient_checkpointing = False
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedAuraFlowAttnProcessor2_0
     def fuse_qkv_projections(self):
         """

src/diffusers/models/transformers/cogvideox_transformer_3d.py

@@ -22,8 +22,8 @@ from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import PeftAdapterMixin
 from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import maybe_allow_in_graph
-from ..attention import Attention, FeedForward
-from ..attention_processor import AttentionProcessor, CogVideoXAttnProcessor2_0, FusedCogVideoXAttnProcessor2_0
+from ..attention import Attention, AttentionMixin, FeedForward
+from ..attention_processor import CogVideoXAttnProcessor2_0, FusedCogVideoXAttnProcessor2_0
 from ..cache_utils import CacheMixin
 from ..embeddings import CogVideoXPatchEmbed, TimestepEmbedding, Timesteps
 from ..modeling_outputs import Transformer2DModelOutput
@@ -157,7 +157,7 @@ class CogVideoXBlock(nn.Module):
         return hidden_states, encoder_hidden_states
 
 
-class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, CacheMixin):
+class CogVideoXTransformer3DModel(ModelMixin, AttentionMixin, ConfigMixin, PeftAdapterMixin, CacheMixin):
     """
     A Transformer model for video-like data in [CogVideoX](https://github.com/THUDM/CogVideo).
 
@@ -331,66 +331,6 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, Cac
 
         self.gradient_checkpointing = False
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedCogVideoXAttnProcessor2_0
     def fuse_qkv_projections(self):
         """

src/diffusers/models/transformers/consisid_transformer_3d.py

@@ -22,8 +22,8 @@ from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import PeftAdapterMixin
 from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import maybe_allow_in_graph
-from ..attention import Attention, FeedForward
-from ..attention_processor import AttentionProcessor, CogVideoXAttnProcessor2_0
+from ..attention import Attention, AttentionMixin, FeedForward
+from ..attention_processor import CogVideoXAttnProcessor2_0
 from ..embeddings import CogVideoXPatchEmbed, TimestepEmbedding, Timesteps
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
@@ -348,7 +348,7 @@ class ConsisIDBlock(nn.Module):
         return hidden_states, encoder_hidden_states
 
 
-class ConsisIDTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
+class ConsisIDTransformer3DModel(ModelMixin, AttentionMixin, ConfigMixin, PeftAdapterMixin):
     """
     A Transformer model for video-like data in [ConsisID](https://github.com/PKU-YuanGroup/ConsisID).
 
@@ -620,66 +620,6 @@ class ConsisIDTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
             ]
         )
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     def forward(
         self,
         hidden_states: torch.Tensor,

src/diffusers/models/transformers/hunyuan_transformer_2d.py

@@ -11,7 +11,7 @@
 # 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.
-from typing import Dict, Optional, Union
+from typing import Optional
 
 import torch
 from torch import nn
@@ -19,8 +19,8 @@ from torch import nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils import logging
 from ...utils.torch_utils import maybe_allow_in_graph
-from ..attention import FeedForward
-from ..attention_processor import Attention, AttentionProcessor, FusedHunyuanAttnProcessor2_0, HunyuanAttnProcessor2_0
+from ..attention import AttentionMixin, FeedForward
+from ..attention_processor import Attention, FusedHunyuanAttnProcessor2_0, HunyuanAttnProcessor2_0
 from ..embeddings import (
     HunyuanCombinedTimestepTextSizeStyleEmbedding,
     PatchEmbed,
@@ -200,7 +200,7 @@ class HunyuanDiTBlock(nn.Module):
         return hidden_states
 
 
-class HunyuanDiT2DModel(ModelMixin, ConfigMixin):
+class HunyuanDiT2DModel(ModelMixin, AttentionMixin, ConfigMixin):
     """
     HunYuanDiT: Diffusion model with a Transformer backbone.
 
@@ -350,66 +350,6 @@ class HunyuanDiT2DModel(ModelMixin, ConfigMixin):
         if self.original_attn_processors is not None:
             self.set_attn_processor(self.original_attn_processors)
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     def set_default_attn_processor(self):
         """
         Disables custom attention processors and sets the default attention implementation.

src/diffusers/models/transformers/pixart_transformer_2d.py

@@ -11,15 +11,15 @@
 # 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.
-from typing import Any, Dict, Optional, Union
+from typing import Any, Dict, Optional
 
 import torch
 from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils import logging
-from ..attention import BasicTransformerBlock
-from ..attention_processor import Attention, AttentionProcessor, AttnProcessor, FusedAttnProcessor2_0
+from ..attention import AttentionMixin, BasicTransformerBlock
+from ..attention_processor import Attention, AttnProcessor, FusedAttnProcessor2_0
 from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
@@ -29,7 +29,7 @@ from ..normalization import AdaLayerNormSingle
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class PixArtTransformer2DModel(ModelMixin, ConfigMixin):
+class PixArtTransformer2DModel(ModelMixin, AttentionMixin, ConfigMixin):
     r"""
     A 2D Transformer model as introduced in PixArt family of models (https://huggingface.co/papers/2310.00426,
     https://huggingface.co/papers/2403.04692).
@@ -184,66 +184,6 @@ class PixArtTransformer2DModel(ModelMixin, ConfigMixin):
                 in_features=self.config.caption_channels, hidden_size=self.inner_dim
             )
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     def set_default_attn_processor(self):
         """
         Disables custom attention processors and sets the default attention implementation.

src/diffusers/models/transformers/prior_transformer.py

 from dataclasses import dataclass
-from typing import Dict, Optional, Union
+from typing import Optional, Union
 
 import torch
 import torch.nn.functional as F
@@ -8,11 +8,10 @@ from torch import nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin
 from ...utils import BaseOutput
-from ..attention import BasicTransformerBlock
+from ..attention import AttentionMixin, BasicTransformerBlock
 from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
-    AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
 )
@@ -33,7 +32,7 @@ class PriorTransformerOutput(BaseOutput):
     predicted_image_embedding: torch.Tensor
 
 
-class PriorTransformer(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin):
+class PriorTransformer(ModelMixin, AttentionMixin, ConfigMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin):
     """
     A Prior Transformer model.
 
@@ -166,66 +165,6 @@ class PriorTransformer(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, Pef
         self.clip_mean = nn.Parameter(torch.zeros(1, clip_embed_dim))
         self.clip_std = nn.Parameter(torch.zeros(1, clip_embed_dim))
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
     def set_default_attn_processor(self):
         """

src/diffusers/models/transformers/sana_transformer.py

@@ -21,9 +21,9 @@ from torch import nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
 from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from ..attention import AttentionMixin
 from ..attention_processor import (
     Attention,
-    AttentionProcessor,
     SanaLinearAttnProcessor2_0,
 )
 from ..embeddings import PatchEmbed, PixArtAlphaTextProjection, TimestepEmbedding, Timesteps
@@ -289,7 +289,7 @@ class SanaTransformerBlock(nn.Module):
         return hidden_states
 
 
-class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
+class SanaTransformer2DModel(ModelMixin, AttentionMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
     r"""
     A 2D Transformer model introduced in [Sana](https://huggingface.co/papers/2410.10629) family of models.
 
@@ -414,66 +414,6 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
 
         self.gradient_checkpointing = False
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     def forward(
         self,
         hidden_states: torch.Tensor,

src/diffusers/models/transformers/stable_audio_transformer.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 
-from typing import Dict, Optional, Union
+from typing import Optional, Union
 
 import numpy as np
 import torch
@@ -22,8 +22,8 @@ import torch.nn as nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils import logging
 from ...utils.torch_utils import maybe_allow_in_graph
-from ..attention import FeedForward
-from ..attention_processor import Attention, AttentionProcessor, StableAudioAttnProcessor2_0
+from ..attention import AttentionMixin, FeedForward
+from ..attention_processor import Attention, StableAudioAttnProcessor2_0
 from ..modeling_utils import ModelMixin
 from ..transformers.transformer_2d import Transformer2DModelOutput
 
@@ -182,7 +182,7 @@ class StableAudioDiTBlock(nn.Module):
         return hidden_states
 
 
-class StableAudioDiTModel(ModelMixin, ConfigMixin):
+class StableAudioDiTModel(ModelMixin, AttentionMixin, ConfigMixin):
     """
     The Diffusion Transformer model introduced in Stable Audio.
 
@@ -274,66 +274,6 @@ class StableAudioDiTModel(ModelMixin, ConfigMixin):
 
         self.gradient_checkpointing = False
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        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)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
     # Copied from diffusers.models.transformers.hunyuan_transformer_2d.HunyuanDiT2DModel.set_default_attn_processor with Hunyuan->StableAudio
     def set_default_attn_processor(self):
         """