Update app.py

app.py

@@ -9,7 +9,7 @@ from functools import lru_cache
 from threading import Thread
 
 # Constants
-MODEL_PATH = "Ozaii/Zephyrr"
+MODEL_PATH = "Ozaii/Zephyr"
 MAX_SEQ_LENGTH = 2048
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 MAX_GENERATION_TIME = 55  # Set to 55 seconds to give some buffer