Update app.py

app.py

@@ -111,7 +111,7 @@ def engine(input_embeds):
             batch_input_embeds = input_embeds[start:start+batch_size]
             outputs = m.generate(
                 inputs_embeds = batch_input_embeds.to("cuda"),
-                max_new_tokens = 16,
+                max_new_tokens = 512,
                 do_sample = True,
                 temperature = 0.6,
                 top_p = 0.9,