Update app.py

app.py

@@ -108,7 +108,7 @@ def respond(
             result = local_generate(
                 prompt,
                 generation_config=generator_conf,
-                max_new_tokens=max_tokens,
+                max_new_tokens=tokens_inc,
                 do_sample=True,
                 top_p=top_p,
                 repetition_penalty=1.2,