Update app.py

app.py

@@ -57,10 +57,11 @@ start_messages = [
 
 # past_key_values = PastKV()
 
-
 def past_kv_to_device(past_kv, device):
     return tuple((k.to(device).detach(), v.to(device).detach()) for k, v in past_kv)
 
+def detach_past_kv(past_kv):
+    return tuple((k.detach(), v.detach()) for k, v in past_kv)
 
 @spaces.GPU
 def set_past_key_values():
@@ -71,7 +72,7 @@ def set_past_key_values():
     test_messages = [*start_messages, {'role': 'user', 'content': 'Hello World!'}]
     tokenized_test = tokenizer.apply_chat_template(test_messages, return_tensors='pt')    
     assert (tokenized_test[:, :tokenized.shape[1]] == tokenized).all().cpu().item()
-    return gr.State(model(tokenized.to(model.device)).past_key_values)
+    return gr.State(detach_past_kv(model(tokenized.to(model.device)).past_key_values))
     
 
 @spaces.GPU