Yuanhang Tang

I solve this problem just now. `device_map` should be set as `auto` otherwise the inference would be started on CPU. However, the 'half' precision is only implemented for GPU. Therefore,...