GitLab

* fix: passing max_length to vllm engine args

* feat: add `max_model_len`

* chore: lint

* doc_to_decontamination_query can use function

* add option for doc_to_decontamination_query to follow doc_to_text

* added documentation for doc_to_decontamination_query

* adjust description

* format

* Additional process for doc_to_choice

* doc_to_choice can also parse a string

* 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.