Commit Graph

100 Commits

Author SHA1 Message Date
hiyouga
c47ab6c072 improve rlhf 2024-07-02 22:23:08 +08:00
hiyouga
4c296001c4 fix ppo callbacks 2024-07-02 17:34:56 +08:00
hiyouga
73280b7dc7 tiny fix 2024-07-01 05:43:17 +08:00
hiyouga
1856a08e87 add eval acc 2024-07-01 03:51:20 +08:00
hiyouga
2f4b89ace1 loose gemma2 attention 2024-06-29 01:42:14 +08:00
hiyouga
6f63050e1b add Gemma2 models 2024-06-28 01:26:50 +08:00
hiyouga
8baf3b22b0 refactor pissa, improve llamaboard 2024-06-28 01:04:24 +08:00
hzhaoy
677c86594e fix #4579 2024-06-27 13:49:57 +08:00
hiyouga
8d6cd69ac4 fix #4458 2024-06-26 19:52:35 +08:00
hiyouga
cc016461e6 fix #4379 2024-06-25 02:31:44 +08:00
hiyouga
095fab58d3 tiny fix about badam 2024-06-25 01:54:53 +08:00
hoshi-hiyouga
d0f953bf5b Merge pull request #4352 from Ledzy/main
[Enhancement] Support ZeRO-3 when using BAdam
2024-06-25 01:49:13 +08:00
hiyouga
e507e60638 update readme 2024-06-24 18:22:12 +08:00
hiyouga
4cff6a4ad5 fix templates 2024-06-19 17:44:05 +08:00
Jonery
5c2ff1b749 Cleaner integration. 2024-06-19 12:29:40 +08:00
Jonery
0f72aac8c9 Support distributed BAdam. 2024-06-18 12:27:47 +08:00
Jonery
ea1f3ba5e0 Merge remote-tracking branch 'upstream/main' 2024-06-17 18:44:51 +08:00
Jonery
33b4372778 adapt for badam with ds zero3 2024-06-17 18:18:10 +08:00
hiyouga
46093b5786 fix tol 2024-06-16 01:38:44 +08:00
hiyouga
8c1046d78a support pissa 2024-06-16 01:08:12 +08:00
hiyouga
38b6b0f52e tiny fix 2024-06-16 01:06:41 +08:00
hiyouga
d87108daa6 add license 2024-06-15 17:54:33 +08:00
hiyouga
78589cf90c fix #4295 2024-06-15 04:34:55 +08:00
hiyouga
b27269bd2b add test cases 2024-06-15 04:05:54 +08:00
hiyouga
6baafd4eb3 fix #4221 2024-06-13 02:48:21 +08:00
hiyouga
cf9f2d6c42 fix #4209
DeepSpeed ZeRO3 has inflight param error when calling model.eval()
2024-06-13 02:25:50 +08:00
hiyouga
2ed8270112 clean code 2024-06-13 01:58:16 +08:00
hiyouga
89f2bd8c8c fix #4198 2024-06-11 15:38:38 +08:00
hiyouga
f8d8690bf4 tiny fix 2024-06-07 05:19:21 +08:00
hiyouga
4489d73ac7 fix ppo trainer save zero3 model
accelerator.get_state_dict(ds_model) should be called at all ranks
2024-06-07 05:14:19 +08:00
hiyouga
2702d7e952 fix ppo in trl 0.8.6 2024-06-07 04:48:29 +08:00
hiyouga
f9e818d79c fix #4120 2024-06-07 04:18:05 +08:00
hiyouga
74f96efef9 rename files 2024-06-07 00:09:06 +08:00
hiyouga
76c61905b2 fix ppo+zero3 #3108 2024-06-06 23:30:07 +08:00
hiyouga
149610c636 fix ppo dataset bug #4012 2024-06-06 19:03:20 +08:00
hiyouga
fad2591e31 update trainers 2024-06-06 18:45:49 +08:00
hiyouga
67fe822324 fix #4090 2024-06-06 00:50:32 +08:00
hiyouga
f9a206509e remove gc warnings in DPO&KTO 2024-06-03 22:53:54 +08:00
hoshi-hiyouga
24499f40dc Update trainer.py 2024-06-03 22:08:38 +08:00
enji.zhou
34a2c5087a fix KTO Trainer Sampler 2024-06-03 21:32:38 +08:00
Uminosachi
14e97dc119 Set scheduler_specific_kwargs to get_scheduler 2024-05-31 13:45:39 +09:00
hiyouga
65cd8bdbdb 10x generate in ppo w/ zero3
https://github.com/huggingface/trl/pull/1483
2024-05-29 00:23:23 +08:00
hiyouga
7c8e01bb74 update dpo, kto trainer 2024-05-29 00:14:29 +08:00
hiyouga
900e1ea622 clean kto trainer 2024-05-28 21:43:26 +08:00
hiyouga
cb63b32986 support SimPO #3900 2024-05-26 23:46:33 +08:00
hiyouga
3a023bca2a refactor data preprocessing, fix mllm rlhf 2024-05-24 04:08:25 +08:00
hiyouga
a851056229 improve data process logger 2024-05-18 22:02:42 +08:00
hiyouga
c450ee87a3 improve KTO impl., replace datasets 2024-05-18 03:44:56 +08:00
enji.zhou
db1d5a4f51 add kto 2024-05-17 13:09:17 +08:00
hiyouga
308edbc426 rename package 2024-05-16 18:39:08 +08:00