update to llama3.1

app.py

@@ -8,7 +8,7 @@ from langchain_core.output_parsers import StrOutputParser
 
 MAX_TOKENS = os.getenv("MAX_TOKENS")
 if MAX_TOKENS is None:
-    MAX_TOKENS = "8192"
+    MAX_TOKENS = "131072"
 
 MAX_TOKENS = int(MAX_TOKENS)
 
@@ -46,7 +46,7 @@ def load_models():
 
     models = [obj['id']
               for obj in result['data'] if obj['active']]
-    models.sort(key=lambda x: not x.startswith('llama3-70b'))
+    models.sort(key=lambda x: not x.startswith('llama-3.1'))
     return models