set gpu layers to 50

app.py

@@ -21,7 +21,7 @@ conn = duckdb.connect()
 
 llama = llama_cpp.Llama(
     model_path="Hermes-2-Pro-Llama-3-8B-Q8_0.gguf",
-    n_gpu_layers=-1,
+    n_gpu_layers=50,
     chat_format="chatml",
     n_ctx=2048,
     verbose=False,