Defer llama_cpp import into @spaces.GPU path (ZeroGPU: CUDA libs only exist inside the GPU call)

app.py

@@ -19,7 +19,6 @@ _load_env()
 
 import gradio as gr
 from huggingface_hub import hf_hub_download
-from llama_cpp import Llama
 
 try:
     import spaces
@@ -94,6 +93,7 @@ def model_path():
 def get_llm():
     global _llm
     if _llm is None:
+        from llama_cpp import Llama
         _llm = Llama(
             model_path=model_path(),
             n_gpu_layers=N_GPU_LAYERS,