Early stopping metric (#537)

No gather single gpu (#523)

Changed Bench Eval to report metrics correctly by split. Added total accuracy and renamed previously used bench_accuracy to bench_average_accuracy. (#512)

add eval benchmark callback (#441)

ReLoRA implementation (with quantization) (#322)

improve GPU logging to break out pytorch cache and system mem

log GPU memory usage

more gpt-neox long ctx fixes

fix bettertransformers save, force it to skip after saving correctly in callback

more tweaks to do pre-training with bettertransformers

Apply isort then black

Lint callbacks.py

black formatting

Add callback save peft_model on_save