feat: implement apply_lora_scale to remove boilerplate.

src/diffusers/models/transformers/transformer_flux.py

@@ -22,7 +22,7 @@ import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FluxTransformer2DLoadersMixin, FromOriginalModelMixin, PeftAdapterMixin
-from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from ...utils import apply_lora_scale, logging
 from ...utils.torch_utils import maybe_allow_in_graph
 from .._modeling_parallel import ContextParallelInput, ContextParallelOutput
 from ..attention import AttentionMixin, AttentionModuleMixin, FeedForward
@@ -634,6 +634,7 @@ class FluxTransformer2DModel(
 
         self.gradient_checkpointing = False
 
+    @apply_lora_scale("joint_attention_kwargs")
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -675,20 +676,6 @@ class FluxTransformer2DModel(
             If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
             `tuple` where the first element is the sample tensor.
         """
-        if joint_attention_kwargs is not None:
-            joint_attention_kwargs = joint_attention_kwargs.copy()
-            lora_scale = joint_attention_kwargs.pop("scale", 1.0)
-        else:
-            lora_scale = 1.0
-
-        if USE_PEFT_BACKEND:
-            # weight the lora layers by setting `lora_scale` for each PEFT layer
-            scale_lora_layers(self, lora_scale)
-        else:
-            if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
-                logger.warning(
-                    "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
-                )
 
         hidden_states = self.x_embedder(hidden_states)
 
@@ -785,10 +772,6 @@ class FluxTransformer2DModel(
         hidden_states = self.norm_out(hidden_states, temb)
         output = self.proj_out(hidden_states)
 
-        if USE_PEFT_BACKEND:
-            # remove `lora_scale` from each PEFT layer
-            unscale_lora_layers(self, lora_scale)
-
         if not return_dict:
             return (output,)
 

src/diffusers/pipelines/pipeline_utils.py

@@ -111,7 +111,7 @@ LIBRARIES = []
 for library in LOADABLE_CLASSES:
     LIBRARIES.append(library)
 
-SUPPORTED_DEVICE_MAP = ["balanced"] + [get_device(), "cpu"]
+SUPPORTED_DEVICE_MAP = ["balanced"] + [get_device()]
 
 logger = logging.get_logger(__name__)
 
@@ -467,7 +467,8 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         pipeline_is_sequentially_offloaded = any(
             module_is_sequentially_offloaded(module) for _, module in self.components.items()
         )
-        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
+
+        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline which doesn't allow explicit device placement using `to()`. You can call `reset_device_map()` to remove the existing device map from the pipeline."
@@ -1186,7 +1187,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         """
         self._maybe_raise_error_if_group_offload_active(raise_error=True)
 
-        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
+        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_model_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_model_cpu_offload()`."
@@ -1310,7 +1311,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
         self.remove_all_hooks()
 
-        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
+        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_sequential_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_sequential_cpu_offload()`."
@@ -2199,21 +2200,6 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 return True
         return False
 
-    def _is_pipeline_device_mapped(self):
-        # We support passing `device_map="cuda"`, for example. This is helpful, in case
-        # users want to pass `device_map="cpu"` when initializing a pipeline. This explicit declaration is desirable
-        # in limited VRAM environments because quantized models often initialize directly on the accelerator.
-        device_map = self.hf_device_map
-        is_device_type_map = False
-        if isinstance(device_map, str):
-            try:
-                torch.device(device_map)
-                is_device_type_map = True
-            except RuntimeError:
-                pass
-
-        return not is_device_type_map and isinstance(device_map, dict) and len(device_map) > 1
-
 
 class StableDiffusionMixin:
     r"""

src/diffusers/utils/__init__.py

@@ -130,6 +130,7 @@ from .loading_utils import get_module_from_name, get_submodule_by_name, load_ima
 from .logging import get_logger
 from .outputs import BaseOutput
 from .peft_utils import (
+    apply_lora_scale,
     check_peft_version,
     delete_adapter_layers,
     get_adapter_name,

src/diffusers/utils/peft_utils.py

@@ -16,6 +16,7 @@ PEFT utilities: Utilities related to peft library
 """
 
 import collections
+import functools
 import importlib
 from typing import Optional
 
@@ -275,6 +276,59 @@ def set_weights_and_activate_adapters(model, adapter_names, weights):
                 module.set_scale(adapter_name, get_module_weight(weight, module_name))
 
 
+def apply_lora_scale(kwargs_name: str = "joint_attention_kwargs"):
+    """
+    Decorator to automatically handle LoRA layer scaling/unscaling in forward methods.
+
+    This decorator extracts the `lora_scale` from the specified kwargs parameter, applies scaling before the forward
+    pass, and ensures unscaling happens after, even if an exception occurs.
+
+    Args:
+        kwargs_name (`str`, defaults to `"joint_attention_kwargs"`):
+            The name of the keyword argument that contains the LoRA scale. Common values include
+            "joint_attention_kwargs", "attention_kwargs", "cross_attention_kwargs", etc.
+    """
+
+    def decorator(forward_fn):
+        @functools.wraps(forward_fn)
+        def wrapper(self, *args, **kwargs):
+            from . import USE_PEFT_BACKEND
+
+            lora_scale = 1.0
+            attention_kwargs = kwargs.get(kwargs_name)
+
+            if attention_kwargs is not None:
+                attention_kwargs = attention_kwargs.copy()
+                kwargs[kwargs_name] = attention_kwargs
+                lora_scale = attention_kwargs.pop("scale", 1.0)
+            else:
+                if (
+                    not USE_PEFT_BACKEND
+                    and attention_kwargs is not None
+                    and attention_kwargs.get("scale", None) is not None
+                ):
+                    logger.warning(
+                        f"Passing `scale` via `{kwargs_name}` when not using the PEFT backend is ineffective."
+                    )
+
+            # Apply LoRA scaling if using PEFT backend
+            if USE_PEFT_BACKEND:
+                scale_lora_layers(self, lora_scale)
+
+            try:
+                # Execute the forward pass
+                result = forward_fn(self, *args, **kwargs)
+                return result
+            finally:
+                # Always unscale, even if forward pass raises an exception
+                if USE_PEFT_BACKEND:
+                    unscale_lora_layers(self, lora_scale)
+
+        return wrapper
+
+    return decorator
+
+
 def check_peft_version(min_version: str) -> None:
     r"""
     Checks if the version of PEFT is compatible.