Update chain_setup.py

chain_setup.py

@@ -21,9 +21,9 @@ def load_llm():
         n_ctx=2048,
         n_batch=512,
         chat_format="chatml",
-        grammar="",
+        grammar=None,
         streaming=True,
-        grammar_path="empty.jinja",  # ensure this file exists and is empty
+        grammar_path=None,  # ensure this file exists and is empty
         use_jinja=False,
         rope_freq_base=10000.0,
         rope_freq_scale=1.0,
@@ -33,7 +33,21 @@ def load_llm():
         repeat_penalty=1.1,
         temperature=0.8,
         top_k=40,
-        top_p=0.95
+        top_p=0.95,
+        logprobs=None,
+        callback_manager=None,
+        custom_get_token_ids = None,
+        logprobs= None,
+        lora_base = None,
+        lora_path = None,
+        max_tokens = 256,
+        metadata= None,
+        n_gpu_layers= None,
+        n_threads= None,
+        stop=[],
+        suffix= None,
+        tags = None
+
     )
 
     return llm