[Fix] Fix OOM when qlora converting

[fanqiNO1]

If the LLM is too big to be loaded in a single GPU, we need device_map = 'auto' to avoid OOM.

According to the issue #715.