Update app.py

app.py

@@ -61,7 +61,7 @@ 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)
+    return tuple((k.detach().numpy().tolist(), v.detach().numpy().tolist()) for k, v in past_kv)
 
 @spaces.GPU
 def set_past_key_values():
@@ -72,7 +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(detach_past_kv(model(tokenized.to(model.device)).past_key_values))
+    return detach_past_kv(model(tokenized.to(model.device)).past_key_values)
     
 
 @spaces.GPU
@@ -90,7 +90,7 @@ def generate(text, past_key_values):
 if __name__ == "__main__":
     with torch.no_grad():
         past_key_values = set_past_key_values()
-        print(f'{past_key_values.past_key_values=}')
+        print(f'{past_key_values=}')
         demo = gr.Interface(generate, 
                             inputs=["textbox", gr.State(past_key_values)], outputs="textbox")
         demo.launch()