diff --git a/src/diffusers/pipelines/transformers_loading_utils.py b/src/diffusers/pipelines/transformers_loading_utils.py
index c750292f59..6f17c2a8cc 100644
--- a/src/diffusers/pipelines/transformers_loading_utils.py
+++ b/src/diffusers/pipelines/transformers_loading_utils.py
@@ -112,10 +112,14 @@ def _load_transformers_model_from_dduf(
                 tensors = safetensors.torch.load(mmap)
                 # Update the state dictionary with tensors
                 state_dict.update(tensors)
-            return cls.from_pretrained(
+            model = cls.from_pretrained(
                 pretrained_model_name_or_path=None,
                 config=config,
                 generation_config=generation_config,
                 state_dict=state_dict,
                 **kwargs,
             )
+            # Models loaded via from_pretrained are in eval mode by default,
+            # but we need to preserve training mode for consistency with non-DDUF loading
+            model.train()
+            return model