* add support for other model types
dependent on future-peft-changes but with fallback to function now
* use encoding=utf8 for training format
* make shuffling optional
and describe dropout a bit more
* add eval_steps to control evaluation
* make callbacks not depend on globals
* make save steps controllable
* placeholder of initial loading-existing-model support
and var name cleanup
* save/load parameters
* last bit of cleanup
* remove `gptq_bits` ref as main branch removed that setting
* add higher_rank_limit option
2048 is basically unreachable due to VRAM, but i trained at 1536 with batch size = 1 on a 7B model.
Note that it's in the do_train input just to save as a parameter
* fix math on save_steps
a3eec62b50