Remove hardcoded move to device in forward.

hf_model.py

@@ -65,7 +65,6 @@ class CondViTForEmbedding(PreTrainedModel):
                 convert_to_tensor=True,
                 convert_to_numpy=False,
             )
-            text_embeddings = text_embeddings.to(pixel_values.device)
         else:
             text_embeddings = None
         return self.condvit(imgs=pixel_values, c=text_embeddings)