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

in hugging face model loglikelihood_tokens and greedy_util functions
when batch-size is set to auto

This adds a bunch of simple annotations suggested by https://github.com/JelleZijlstra/autotyping.

`max_length` was misdefined as a tuple.

I discovered that the accuracy of all models (e.g., llama7b, llama13b, starcoder) in the 'gsm8k-cot' task was 0%. After a thorough investigation, I realized that the generated text for each question was causing an early stop, preventing the 'regex_pattern' from finding any answers. This issue was caused by an incorrect assignment of the 'primary_until' variable in the 'greedy_until' function. Specifically, 'primary_until' should be a list of strings instead of a single string, as the 'stop_sequences' parameter in the 'stop_sequences_criteria' function requires a List[str]. Once I assigned 'primary_until' to '[until[0]]', the accuracy of llama7b in the 'gsm8k-cot' task increased to 1.67%.