Update app.py

app.py

@@ -71,7 +71,7 @@ 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()
-    past_key_values.past_key_values = PastKV(past_kv_to_device(model(tokenized.to(model.device)).past_key_values, 'cpu'))
+    past_key_values.past_key_values = PastKV(past_kv_to_device(model(tokenized.to(model.device)).past_key_values, model.device))
     return True
     
 
@@ -82,13 +82,14 @@ def generate(text, past_key_values):
         {'role': 'user', 'content': text}
     ]
     response = pipe(messages, 
-                    past_key_values=past_kv_to_device(past_key_values, model.device), 
+                    past_key_values=past_key_values, #past_kv_to_device(past_key_values, pipe.model.device), 
                     **generate_kwargs)[0]['generated_text']
     return response[-1]['content']
 
     
 if __name__ == "__main__":
-    set_past_key_values()
-    demo = gr.Interface(partial(generate, past_key_values=past_key_values.past_key_values), 
-                        inputs="textbox", outputs="textbox")
-    demo.launch()
+    with torch.no_grad():
+        set_past_key_values()
+        demo = gr.Interface(partial(generate, past_key_values=past_key_values.past_key_values), 
+                            inputs="textbox", outputs="textbox")
+        demo.launch()