gather/broadcast the max value of the packing efficiency automatically (#463)

don't add position_ids for evals (#591)

optionally configure sample packing for evals (#589)

make phi training work with Loras (#588)

set fsdp state dict (#584)

pop block_cls since it's not an actual kwarg

don't resize embeddings if it's already large enough (#577)

Support Sample packing for phi arch (#586)

support custom field for completion from yml (#580)

prevent cli functions from getting fired on import (#581)

refactor scripts/finetune.py into new cli modules (#550)

E2e device cuda (#575)

mypy wandb ignore (#572)

remove columns after tokenizing for pretraining (#571)

fix save_steps so it doesn't get duplicated (#567)

Model parallel (#538)

let hf trainer handle torch compile (#516)

improve how we setup eval/save strategies and steps (#547)

gracefully handle length feature used for group by (#565)

add optimization for group-by-len (#563)

fix wandb so mypy doesn't complain (#562)

Add training callback to send predictions to WandB table (#521)

Fix pretraining with iterable/streaming Dataset (#556)

fix for quant config from model (#540)

workaround for md5 variations (#533)

Early stopping metric (#537)

recommend padding when using sample packing (#531)

log rank too (#527)

misc fixes/improvements (#513)

Add support for GPTQ using native transformers/peft (#468)

Merge pull request #520 from bdashore3/sharegpt-fixes

reorg a bit

use flash_attn rmsnorm when available (#526)

use flash_attn xentropy when available (#525)

move is_llama_derived_model into normalize_config (#524)

No gather single gpu (#523)

Prompters: ShareGPT: Allow for custom system prompts

fix: bad dtype for full finetune (#504)

log supervised token count (#448)

Debug tokenization output: Add ability to output text only (no tokens), and/or specify num samples to see (#511)

Added advanced DDP args (#515)

split train from other cli options (#503)

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

drop empty tokenized rows too (#509)

add eval benchmark callback (#441)

support for datasets with multiple names (#480)

Refactor train cfg cli (#499)

use math.ceil instead of round /cc #498

pad_to_worst_case_seq_len boolean, for testing memory limits (#498)

simplify linear layer locator