Update chain_setup.py

chain_setup.py

@@ -23,14 +23,13 @@ def load_llm():
         chat_format="chatml",
         grammar="",
         streaming=True,
-        grammar_path="",  # ensure this file exists and is empty
+        grammar_path="empty.jinja",  # ensure this file exists and is empty
         use_jinja=False,
         rope_freq_base=10000.0,
         rope_freq_scale=1.0,
         use_mmap=True,
         last_n_tokens_size=64,
         echo=False,
-        # max_tokens= ,
         repeat_penalty=1.1,
         temperature=0.8,
         top_k=40,