code

app.py

@@ -50,7 +50,7 @@ def respond(
         temperature=temperature,
         streamer=streamer,
         top_p=top_p,
-        return_full_text=True,
+        return_full_text=False,
         do_sample=True,
     )
     
@@ -60,9 +60,7 @@ def respond(
     thread.start()
     # Print the generated text in real-time
     for new_text in streamer:
-        response += new_text
-        response = tokenizer.decode(tokenizer.encode(response), skip_special_tokens=True)
-    yield response
+        yield tokenizer.decode(tokenizer.encode(new_text), skip_special_tokens=True)
     
 with gr.Blocks() as demo:
     with gr.Row():