Update handler.py

handler.py

@@ -5,7 +5,7 @@ MAX_TOKENS=8192
 
 class EndpointHandler():
     def __init__(self, data):
-        self.model = Llama.from_pretrained("MrOvkill/gemma-2-inference-endpoint-GGUF", filename="gemma-2b.q8_0.gguf")
+        self.model = Llama.from_pretrained("lmstudio-ai/gemma-2b-it-GGUF", filename="gemma-2b.q8_0.gguf", n_ctx=8192, cache_dir="./", n_gpu_layers=99)
 
     def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
         inputs = data.pop("inputs", "")