Update app.py

app.py

@@ -45,6 +45,11 @@ 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
 """
@@ -65,11 +70,15 @@ def answer_question(image, question, max_crops, num_tokens):
     thread.start()
 
     buffer = ""
+    output_started = False
     for new_text in streamer:
+        if not output_started:
+            if "<assistant>" in new_text:
+                output_started = True
+            continue
         buffer += new_text
-        print(buffer)
         if len(buffer) > 1:
-            yield buffer[len(prompt) :]
+            yield buffer
 
 
 with gr.Blocks() as demo: