Update app.py

app.py

@@ -21,7 +21,7 @@ DESCRIPTION = """\
 """
 
 MAX_MAX_NEW_TOKENS = 4096
-DEFAULT_MAX_NEW_TOKENS = 128
+DEFAULT_MAX_NEW_TOKENS = 512
 MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "3840"))
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
@@ -42,7 +42,7 @@ def generate(
     message: str,
     chat_history: list[tuple[str, str]],
     system_prompt: str,
-    max_new_tokens: int = 128,
+    max_new_tokens: int = 512,
     temperature: float = 0.6,
     top_p: float = 0.9,
     top_k: int = 50,