third commit - add spaces.GPU

app.py

@@ -36,6 +36,8 @@ from langchain_huggingface import HuggingFaceEmbeddings
 from tqdm import tqdm
 import joblib
 
+import spaces
+
 # Token streamer for generation
 class TokenStreamer(BaseStreamer):
     def __init__(self, skip_prompt: bool = False, timeout=None):
@@ -170,6 +172,7 @@ def query_vector_store(vector_store: FAISS, query, k=4, relevance_threshold=0.8)
     context = [doc.page_content for doc in similar_docs]
     return context
 
+@spaces.GPU
 class ModelWorker:
     def __init__(self, model_path, device='cuda'):
         self.device = device