torch.npu.set_device_limit(device_id, cube_num, vec_num)
aclrtSetDeviceResLimit(cube=4)
aclrtSetDeviceResLimit(vector=8)