move model inside generate query

app.py

@@ -19,20 +19,6 @@ view_name = "dataset_view"
 hf_api = HfApi()
 conn = duckdb.connect()
 
-llama = llama_cpp.Llama(
-    model_path="Hermes-2-Pro-Llama-3-8B-Q8_0.gguf",
-    n_gpu_layers=50,
-    chat_format="chatml",
-    n_ctx=2048,
-    verbose=False,
-    temperature=0.1,
-)
-
-create = instructor.patch(
-    create=llama.create_chat_completion_openai_v1,
-    mode=instructor.Mode.JSON_SCHEMA,
-)
-
 
 class OutputTypes(str, enum.Enum):
     TABLE = "table"
@@ -87,6 +73,20 @@ CREATE TABLE {} (
 
 @spaces.GPU
 def generate_query(dataset_id: str, query: str) -> dict:
+    llama = llama_cpp.Llama(
+        model_path="Hermes-2-Pro-Llama-3-8B-Q8_0.gguf",
+        n_gpu_layers=50,
+        chat_format="chatml",
+        n_ctx=2048,
+        verbose=False,
+        temperature=0.1,
+    )
+
+    create = instructor.patch(
+        create=llama.create_chat_completion_openai_v1,
+        mode=instructor.Mode.JSON_SCHEMA,
+    )
+
     ddl = get_dataset_ddl(dataset_id)
 
     system_prompt = f"""