Update chain_setup.py

chain_setup.py

@@ -27,7 +27,14 @@ def load_llm():
         use_jinja=False,
         rope_freq_base=10000.0,
         rope_freq_scale=1.0,
-        use_mmap=True
+        use_mmap=True,
+        last_n_tokens_size=64,
+        echo=False,
+        # max_tokens= ,
+        repeat_penalty=1.1,
+        temperature=0.8,
+        top_k=40,
+        top_p=0.95
     )
 
     return llm