Update app.py

app.py

@@ -45,11 +45,6 @@ def cached_vision_process(image, max_crops, num_tokens):
 def answer_question(image, question, max_crops, num_tokens):
     prompt = f"""<|im_start|>user
 <image>
-{question}<|im_end|>
-<|im_start|>assistant
-"""
-    replace_prompt = f"""<|im_start|>user
-
 {question}<|im_end|>
 <|im_start|>assistant
 """
@@ -74,7 +69,7 @@ def answer_question(image, question, max_crops, num_tokens):
         buffer += new_text
         print(buffer)
         if len(buffer) > 1:
-            yield buffer[len(replace_prompt) :]
+            yield buffer[len(prompt) :]
 
 
 with gr.Blocks() as demo: