GitLab

* modularize HFLM code

* pass through extra kwargs to AutoModel.from_pretrained call

* remove explicit model_kwargs

* rename gptq -> autogptq

* fix tokenizer pad token errors

* ensure model always respects device_map and autogptq's selected devices

* add a _get_config helper fn

* remove unlabled test sets

* add note to readme

BBH cot fewshot already has fewshot examples in the description. So num_fewshot needs to be set to 0 so that users won't mistakenly set other num_fewshot values.