Commit History
fix distributed devices (#612)
2fe95cd
unverified
Maxime
commited on
support to disable exllama for gptq (#604)
faecff9
unverified
Delete duplicate lines (#606)
aa656e0
unverified
improve handling for empty text on the tokenization step (#502)
1eebbd0
unverified
Fix for check with cfg and merge_lora (#600)
62a7741
unverified
minor tweaks to simplify (#597)
31b9e0c
unverified
btlm and falcon monkey patches for flash attn (#566)
6b9b229
unverified
add bf16 check (#587)
131afdb
unverified
Feat(data): Allow loading local csv and text (#594)
00dce35
unverified
gather/broadcast the max value of the packing efficiency automatically (#463)
b15b19e
unverified
don't add position_ids for evals (#591)
ab534d7
unverified
optionally configure sample packing for evals (#589)
21ec195
unverified
make phi training work with Loras (#588)
62eaee7
unverified
set fsdp state dict (#584)
be75668
unverified
Jan Philipp Harries
Jan Philipp Harries
commited on
pop block_cls since it's not an actual kwarg
aeec7c4
don't resize embeddings if it's already large enough (#577)
3607882
unverified
Support Sample packing for phi arch (#586)
12a2dbb
unverified
support custom field for completion from yml (#580)
f7a2263
unverified
prevent cli functions from getting fired on import (#581)
8dcd40a
unverified
E2e device cuda (#575)
2414673
unverified
mypy wandb ignore (#572)
c6d870b
unverified
remove columns after tokenizing for pretraining (#571)
1157950
unverified
fix save_steps so it doesn't get duplicated (#567)
3fbde76
unverified
Model parallel (#538)
f6060a6
unverified
improve how we setup eval/save strategies and steps (#547)
36e53c7
unverified
gracefully handle length feature used for group by (#565)
e7aa7b1
unverified
add optimization for group-by-len (#563)
e5bb22a
unverified
fix wandb so mypy doesn't complain (#562)
bf08044
unverified
Add training callback to send predictions to WandB table (#521)
5b67ea9
unverified
Fix pretraining with iterable/streaming Dataset (#556)
2f586d1
unverified
Jan Philipp Harries
Jan Philipp Harries
commited on
fix for quant config from model (#540)
a94f9cb
unverified
workaround for md5 variations (#533)
0b4cf5b
unverified
Early stopping metric (#537)
e30f1e3
unverified
recommend padding when using sample packing (#531)
3437149
unverified
log rank too (#527)
245c5c4
unverified
misc fixes/improvements (#513)
a546ca2
unverified
Add support for GPTQ using native transformers/peft (#468)
3355706
unverified
Merge pull request #520 from bdashore3/sharegpt-fixes
daa4fac
unverified
reorg a bit
fc8766e
use flash_attn rmsnorm when available (#526)
72a6fe1
unverified
use flash_attn xentropy when available (#525)
5fe30b1
unverified
move is_llama_derived_model into normalize_config (#524)
44454ae
unverified
No gather single gpu (#523)
09f1543
unverified
Prompters: ShareGPT: Allow for custom system prompts
995557b
fix: bad dtype for full finetune (#504)
1991946
unverified
log supervised token count (#448)
7710e81
unverified
Debug tokenization output: Add ability to output text only (no tokens), and/or specify num samples to see (#511)
48434be
unverified
Tom Jobbins
commited on