Update app.py

app.py

@@ -14,8 +14,8 @@ DESCRIPTION = "# Arcee Model Testing - current test: arcee-ai/SEC-Calme-7B-Instr
 if not torch.cuda.is_available():
     DESCRIPTION += "\n<p>Running on CPU 🥶 This demo does not work on CPU.</p>"
 
-MAX_MAX_NEW_TOKENS = 2048
-DEFAULT_MAX_NEW_TOKENS = 1024
+MAX_MAX_NEW_TOKENS = 4096
+DEFAULT_MAX_NEW_TOKENS = 4096
 MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
 
 if torch.cuda.is_available():
@@ -28,7 +28,7 @@ if torch.cuda.is_available():
 def generate(
     message: str,
     chat_history: list[tuple[str, str]],
-    max_new_tokens: int = 1024,
+    max_new_tokens: int = 4096,
     temperature: float = 0.6,
     top_p: float = 0.9,
     top_k: int = 50,