Update app.py

app.py

@@ -298,7 +298,7 @@ def expand_prompt(prompt):
         outputs = model.generate(
             input_ids=input_ids,
             attention_mask=attention_mask,
-            max_new_tokens=512,
+            max_new_tokens=384,
             temperature=0.2,
             top_p=0.9,
             do_sample=True,
@@ -312,7 +312,7 @@ def expand_prompt(prompt):
         outputs_2 = model.generate(
             input_ids=input_ids_2,
             attention_mask=attention_mask_2,
-            max_new_tokens=1024,
+            max_new_tokens=512,
             temperature=0.2,
             top_p=0.9,
             do_sample=True,