enable cuda, cpu takes forever

app.py

@@ -40,16 +40,16 @@ client.add(collection_name="recipes",
 model_name = "LeoLM/leo-hessianai-13b-chat"
 
 last_messages = []
-#@spaces.GPU
+@spaces.GPU
 def load_model():
     ankerbot_model = AutoModelForCausalLM.from_pretrained(
         model_name,
-        device_map="cpu",
+        device_map="cuda:0",
         torch_dtype=torch.float16,
         use_cache=True,
         offload_folder="../offload"
     )
-
+    ankerbot_model.gradient_checkpointing_enable()
     ankerbot_tokenizer = AutoTokenizer.from_pretrained(model_name,
         torch_dtype=torch.float16,
         truncation=True,
@@ -60,7 +60,7 @@ def load_model():
     
 _model_cache = None
 
-#@spaces.GPU
+@spaces.GPU
 def get_model():
     global _model_cache
     if _model_cache is None:
@@ -69,7 +69,7 @@ def get_model():
         _model_cache = load_model()
     return _model_cache
 
-#@spaces.GPU
+@spaces.GPU
 def generate_response(query, context, prompts, max_tokens, temperature, top_p, generator):
     system_message_support = f"""<|im_start|>system
         Rolle: Du bist der KI-Assistent für Kundenservice, der im Namen des Unternehmens und Gewürzmanufaktur Ankerkraut handelt und Antworten aus der Ich-Perspektive, basierend auf den bereitgestellten Informationen gibt.
@@ -154,7 +154,7 @@ def search_qdrant_with_context(query_text, collection_name, top_k=3):
     print("Retrieved Text ", retrieved_texts)
 
     return retrieved_texts
-#@spaces.GPU
+@spaces.GPU
 def respond(
     query,
     history: list[tuple[str, str]],
@@ -186,6 +186,7 @@ def respond(
     if len(last_messages) > 5:
         last_messages.pop(0)
     last_messages.append(full_conv)
+    print(last_messages)
     return answer
 
 """