fix: set context window to 1024*128

app.py

@@ -25,12 +25,12 @@ if not os.path.exists(file_path):
         ):
             f.write(chunk)
 
-
+context_window = 1024 * 128
 llm = LlamaCppLLM(
     model_path=file_path,
     n_gpu_layers=-1,
-    # n_ctx=1024 * 128,
-    generation_kwargs={"max_new_tokens": 1024 * 128},
+    n_ctx=context_window,
+    generation_kwargs={"max_new_tokens": context_window},
 )
 task = ArgillaLabeller(llm=llm)
 task.load()
@@ -79,6 +79,7 @@ def process_records_gradio(records, example_records, fields, question):
 
         return json.dumps({"results": results}, indent=2)
     except Exception as e:
+        raise Exception(f"Error: {str(e)}")
         return f"Error: {str(e)}"