Update app.py

app.py

@@ -30,8 +30,8 @@ import ast
 import html
 
 # Constants for text generation
-MAX_MAX_NEW_TOKENS = 8192
-DEFAULT_MAX_NEW_TOKENS = 4096
+MAX_MAX_NEW_TOKENS = 5120
+DEFAULT_MAX_NEW_TOKENS = 3072
 MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")