23 Commits

Author SHA1 Message Date
hiyouga
3595d98b4c fix #4609
unwrap_model_for_generation(reward_model) is necessary for zero3 training


Former-commit-id: 8845e94f917b503bbee0604d7290efea7260a30c
2024-07-03 19:45:51 +08:00
hiyouga
c9e9beee4e tiny fix
Former-commit-id: 71cdf8956e1640a1f3e5f6a4b86d28db70e72041
2024-07-02 23:06:13 +08:00
hiyouga
ea2d3f6c18 remove rlhf support for chatglm2&3
Former-commit-id: 821bb6660e57c29ebf6ac482e78dd2efb8d72437
2024-07-02 23:03:17 +08:00
hiyouga
4828bed837 upcast logits
Former-commit-id: c13ae2df19ed4cdc849bef55d04225e1a98c19b5
2024-07-02 22:32:05 +08:00
hiyouga
cc31014002 improve rlhf
Former-commit-id: c47ab6c07287fb260ea49b8b7af46bdd416f88f7
2024-07-02 22:23:08 +08:00
hiyouga
08296f4092 fix ppo callbacks
Former-commit-id: 4c296001c4b77b814e4bd6cb4049a279718cb775
2024-07-02 17:34:56 +08:00
hiyouga
835f0578c2 refactor pissa, improve llamaboard
Former-commit-id: 8baf3b22b0fb9624807d809832f097301982d192
2024-06-28 01:04:24 +08:00
hiyouga
28e613efd0 fix #4458
Former-commit-id: 8d6cd69ac43afd4bd7c14bd02b0061455827ac9e
2024-06-26 19:52:35 +08:00
hiyouga
a225b5a70c tiny fix about badam
Former-commit-id: 095fab58d3692607c9e78747b4218ae1abcf5aaf
2024-06-25 01:54:53 +08:00
hoshi-hiyouga
fe6ef6400c Merge pull request #4352 from Ledzy/main
[Enhancement] Support ZeRO-3 when using BAdam

Former-commit-id: d0f953bf5bdbfd49acc82ff055bd54889241761a
2024-06-25 01:49:13 +08:00
hiyouga
7be502c5c5 update readme
Former-commit-id: e507e60638b2e8c66f24805b3b28f6b9f98f5924
2024-06-24 18:22:12 +08:00
Jonery
c779899f7b Cleaner integration.
Former-commit-id: 5c2ff1b749a265dd3c979189ec491d8ac911a6f6
2024-06-19 12:29:40 +08:00
Jonery
3a5eacb4cf Support distributed BAdam.
Former-commit-id: 0f72aac8c9227e33ad20d2b1641b1c9faae16a5f
2024-06-18 12:27:47 +08:00
hiyouga
c0c6b8075a tiny fix
Former-commit-id: 38b6b0f52edeb8ba45aa03b415b3c0c1b0e0c1e4
2024-06-16 01:06:41 +08:00
hiyouga
2946153cea add license
Former-commit-id: d87108daa68bd40174b262be1ca65fe6e1b7ab56
2024-06-15 17:54:33 +08:00
hiyouga
a3f4925c2c add test cases
Former-commit-id: b27269bd2b52fb9d43cde8a8b7f293099b0127a2
2024-06-15 04:05:54 +08:00
hiyouga
81ed4d8abf fix #4209
DeepSpeed ZeRO3 has inflight param error when calling model.eval()


Former-commit-id: cf9f2d6c42b5a37038c9eededbb767eae6a3f67d
2024-06-13 02:25:50 +08:00
hiyouga
4f3c89a6eb fix ppo trainer save zero3 model
accelerator.get_state_dict(ds_model) should be called at all ranks


Former-commit-id: 4489d73ac75c8dbc002fc16c854148994d432c3a
2024-06-07 05:14:19 +08:00
hiyouga
f76d427332 fix ppo in trl 0.8.6
Former-commit-id: 2702d7e952523b584d67c8901888b492d4a79b14
2024-06-07 04:48:29 +08:00
hiyouga
8da149ba40 rename files
Former-commit-id: 74f96efef9bcd63f65d0190c901ff9be54ccd350
2024-06-07 00:09:06 +08:00
hiyouga
368695483d fix ppo+zero3 #3108
Former-commit-id: 76c61905b20f69fac5c7a6c4ea9450bf33d3b1f2
2024-06-06 23:30:07 +08:00
hiyouga
468d0e7ed1 10x generate in ppo w/ zero3
https://github.com/huggingface/trl/pull/1483

Former-commit-id: 65cd8bdbdbe1b19250ecd813aeb72c8e00ef2f9c
2024-05-29 00:23:23 +08:00
hiyouga
cae823ddf0 rename package
Former-commit-id: 308edbc4260d45907b4a9d3a45ec21d83e48aacb
2024-05-16 18:39:08 +08:00