diff --git a/demo/app_januspro.py b/demo/app_januspro.py
index 702e58e..82cf6b0 100644
--- a/demo/app_januspro.py
+++ b/demo/app_januspro.py
@@ -22,7 +22,7 @@ vl_gpt = AutoModelForCausalLM.from_pretrained(model_path,
 if torch.cuda.is_available():
     vl_gpt = vl_gpt.to(torch.bfloat16).cuda()
 else:
-    vl_gpt = vl_gpt.to(torch.float16)
+    vl_gpt = vl_gpt.to('cpu')
 
 vl_chat_processor = VLChatProcessor.from_pretrained(model_path)
 tokenizer = vl_chat_processor.tokenizer
@@ -156,7 +156,7 @@ def generate_image(prompt,
                                                                    system_prompt='')
         text = text + vl_chat_processor.image_start_tag
         
-        input_ids = torch.LongTensor(tokenizer.encode(text))
+        input_ids = torch.LongTensor(tokenizer.encode(text)).to(cuda_device)
         output, patches = generate(input_ids,
                                    width // 16 * 16,
                                    height // 16 * 16,