fix: move to gpu when running

app.py

@@ -8,18 +8,10 @@ from sentence_transformers import SentenceTransformer
 import gradio as gr
 from threading import Thread
 
-# FAISS Index auf GPU laden
+# FAISS Index laden (vorerst auf CPU)
 print("Loading FAISS index...")
 cpu_index = faiss.read_index("vector_db/index.faiss")
-
-# Prüfen ob GPU verfügbar ist und Index auf GPU verschieben
-if torch.cuda.is_available():
-    print("Moving FAISS to GPU...")
-    res = faiss.StandardGpuResources()
-    index = faiss.index_cpu_to_gpu(res, 0, cpu_index)
-else:
-    print("GPU not available, using CPU index")
-    index = cpu_index
+gpu_index = None  # GPU Index wird später erstellt
 
 with open("vector_db/chunks.pkl", "rb") as f:
     chunks = pickle.load(f)
@@ -55,8 +47,21 @@ SYS = (
 
 @spaces.GPU()
 def retrieve(q, k=3):
+    global gpu_index, cpu_index
+    
+    # GPU Index beim ersten Aufruf erstellen
+    if gpu_index is None:
+        try:
+            print("Creating GPU index...")
+            res = faiss.StandardGpuResources()
+            gpu_index = faiss.index_cpu_to_gpu(res, 0, cpu_index)
+            print("FAISS index successfully moved to GPU")
+        except Exception as e:
+            print(f"Failed to move FAISS to GPU: {e}")
+            gpu_index = cpu_index  # Fallback to CPU
+    
     emb = ST.encode(q)
-    D, I = index.search(np.array([emb], dtype="float32"), k)
+    D, I = gpu_index.search(np.array([emb], dtype="float32"), k)
     docs, file_sources = [], []
     for i in I[0]:
         chunk = chunks[i]