Update handler.py

handler.py

@@ -8,10 +8,11 @@ MAX_TOKENS=8192
 
 class EndpointHandler():
     def __init__(self, data):
-        self.model = Llama.from_pretrained("lmstudio-ai/gemma-2b-it-GGUF", filename="gemma-2b-it-q4_k_m.gguf", n_ctx=8192)
+        self.model = Llama.from_pretrained("lmstudio-ai/gemma-2b-it-GGUF", filename="gemma-2b-it-4_k_m.gguf", n_ctx=8192)
 
     def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
         args = gem.get_args_or_none(data)
+        fmat = "<startofturn>system\n{system_prompt} <endofturn>\n<startofturn>user\n{inputs} <endofturn>\n<startofturn>model"
         if not args[0]:
             return {
                 "status": args["status"],