fix(training_utils): wrap device in list for DiffusionPipeline (#12178)

- Modify offload_models function to handle DiffusionPipeline correctly
- Ensure compatibility with both single and multiple module inputs

Co-authored-by: Sayak Paul <[email protected]>

Author: MengAiDev

src/diffusers/training_utils.py

@@ -339,7 +339,8 @@ def offload_models(
             original_devices = [next(m.parameters()).device for m in modules]
         else:
             assert len(modules) == 1
-            original_devices = modules[0].device
+            # For DiffusionPipeline, wrap the device in a list to make it iterable
+            original_devices = [modules[0].device]
         # move to target device
         for m in modules:
             m.to(device)