Leetmonkey In Action via Inference

app.py

@@ -33,7 +33,7 @@ def download_model(model_name):
 model_path = download_model(MODEL_NAME)
 llm = Llama(
     model_path=model_path,
-    n_ctx=512,
+    n_ctx=1024,
     n_threads=8,
     n_gpu_layers=1,  # Use all available GPU layers
     verbose=False,