91 Commits

Author SHA1 Message Date
statelesshz
95384eefe8 support export model on Ascend NPU
Former-commit-id: b3e41c6d49728e239b61d8cd9b3603c8dc877549
2023-09-20 10:26:02 +08:00
hiyouga
4e6e42bb55 add tests.cal_flops.py
Former-commit-id: 469f859161dec0e34f4cc849f20e43d442680b5c
2023-09-16 23:40:41 +08:00
hiyouga
2a75b10015 fix #913
Former-commit-id: 0b5f970c05c524670d66c810e9f081a52a1fb5e6
2023-09-15 20:58:28 +08:00
hiyouga
632fff02e0 fix #887
Former-commit-id: 8857e4560219c4052bdb7c7dc1a014a5f5fd0163
2023-09-14 17:56:58 +08:00
mmbwf
6ded1725d9 Update utils.py
Fix parameters load error.

Former-commit-id: 30fb721f1222d1b56a2712519960a63655c20360
2023-09-14 15:38:04 +08:00
hiyouga
c8780205bc fix ppo save model
Former-commit-id: 7ba57d5b1469cd0de0bb391b915bedec97b20ebd
2023-09-12 16:25:29 +08:00
hiyouga
4e86462bad fix #762 #814
Former-commit-id: d4be857e23c74ed65e06903e19da6f18f15d9e30
2023-09-12 16:10:10 +08:00
hiyouga
33bab0e7c1 update flashattn, fix ppo save model
Former-commit-id: 0fbece85a70222e5262a2295203de07ffe648fda
2023-09-11 17:25:36 +08:00
hiyouga
6a71361a54 remove PeftTrainer
Former-commit-id: b218c271edfb07006ddc34b1aca404088de6c528
2023-09-10 22:23:23 +08:00
hiyouga
8ab5566dc0 support FlashAttention2
Former-commit-id: d8aa1404bee9842f3e4cd037ad8d66c85470ac37
2023-09-10 20:43:56 +08:00
hiyouga
85aa16f6c6 fix #850
Former-commit-id: 815b92e698562bfae6eb9a6fa1b612a05d43ed67
2023-09-10 14:22:03 +08:00
hiyouga
f865d0bd51 fix lora target
Former-commit-id: a51b7c98acc599de5ed2eaeeebe7b184105722c5
2023-09-09 17:04:45 +08:00
hiyouga
c818a7ff60 support lora target auto find
Former-commit-id: bca1a247bcef51dced59655c8a14c197569367ca
2023-09-09 15:38:37 +08:00
hiyouga
c6265e6969 fix chatglm2 tokenizer
Former-commit-id: d8d82ca281811c20c89cc03dd00f69735515d6cf
2023-09-09 13:50:29 +08:00
hiyouga
43a20c67d4 fix bug in DPO data collator
Former-commit-id: 90bd085ae4c2775f1e82e045ab0157a451774082
2023-09-08 20:45:07 +08:00
hiyouga
405df0f63d fix #761
Former-commit-id: b34797a845dec8f6daea59e3c353b8a8f8830100
2023-09-08 20:22:18 +08:00
hiyouga
9ed4bb63d4 change to right-padding, update reward score #803
Former-commit-id: 8ea32e4046d75ddfa9517669e9de9f48fea720c6
2023-09-08 20:04:31 +08:00
hiyouga
f225a71445 update requirements
Former-commit-id: f5351c18e15cd26fe628ed364eaa4ecd49874596
2023-09-07 19:26:25 +08:00
hiyouga
091326dc9f fix #818
Former-commit-id: 5a9970dbef3a6975ce5ec6ac2bef19182c75b662
2023-09-07 19:19:53 +08:00
hiyouga
5030f05126 add deepspeed check in PPO training
Former-commit-id: ed1c2c5557bb2714c3341294f0ea86f6496d4b0c
2023-09-07 19:12:40 +08:00
hiyouga
e6fa0229f4 fix #809
Former-commit-id: e2bf7c3badbd5d2fd513ca7a00bd74d9c0d62d07
2023-09-07 19:04:32 +08:00
hiyouga
f74b980650 fix baichuan templates
Former-commit-id: 85b1f6632a752029dabdaed87c58986deb3a6b1d
2023-09-07 18:54:14 +08:00
hiyouga
a4fd976048 refactor dataset_attr, add eos in pt, fix #757
Former-commit-id: a9d1fb72f791ae57a4d12f4e3a7e2abccf6a7077
2023-09-01 19:00:45 +08:00
codemayq
c955d9267c add dataset stage check
Former-commit-id: f7fdc088d49564f7d436fd445e7e1987a9a00a0b
2023-08-30 16:23:08 +08:00
hiyouga
0d0232479f fix import error
Former-commit-id: 2de1a7610a78e41680970b9f308741f98df489fa
2023-08-23 20:45:03 +08:00
hiyouga
38080233a5 fix #649
Former-commit-id: 57146c101f3e8f688b016a44c85e8ad5d1b6f938
2023-08-23 20:21:15 +08:00
hiyouga
d05d535a58 fix #617
Former-commit-id: 5235b15c9181f2b68f7d6caa9a6324b8570d3d0c
2023-08-21 18:16:11 +08:00
hiyouga
570ccc3618 fix ppo trainer #551
Former-commit-id: 0676497104eccc8a737d27890eabf1ca8713c235
2023-08-20 14:07:11 +08:00
hiyouga
acaac6df9e Release v0.1.7
Former-commit-id: 9c9009f49fdff83a29b83d8f97eb5c99e2574256
2023-08-18 17:21:27 +08:00
hiyouga
9f1688924d tiny fix
Former-commit-id: d75e377b0f6f3fd7c034676b81ddef3aab1d6901
2023-08-18 13:07:35 +08:00
hiyouga
b88f0b396c support ppo score norm (trl 0.5.1.dev required)
Former-commit-id: 53e33418d02ee0f34c783e30ae510b811308c598
2023-08-18 12:02:42 +08:00
hiyouga
03edfd07e7 fix PPO trainer #551 , update readme
Former-commit-id: 90205244186df558cd6b0000728d638348db3a10
2023-08-18 11:43:10 +08:00
hiyouga
fceca0bb6a update training resuming
Former-commit-id: 58f13e22da18babed0d2d4348474e07745da8fa5
2023-08-18 01:41:17 +08:00
hoshi-hiyouga
49d4ae3704 Merge branch 'main' into main
Former-commit-id: 725290324562e093565fae79a05341ebf64486d5
2023-08-18 01:37:23 +08:00
hiyouga
66771352bb support bf16 ppo #551
Former-commit-id: d125218cde893c7c8527ab27b4d2dfb2474c384d
2023-08-18 00:40:32 +08:00
hiyouga
caf4a61e21 fix ChatGLM2 ppo #527 #528
Former-commit-id: 9f4c2adc9a9ca8e458d3868805e077182e0d336a
2023-08-18 00:34:59 +08:00
hiyouga
623a34b16f fix generation bug #532
Former-commit-id: be21fc83f9aed0af1e5a2f83f5d5eeb36f1d283c
2023-08-17 22:21:34 +08:00
hiyouga
a46f277477 fix streaming in pt stage #548 #549
Former-commit-id: b0ed0dec5e6788a0344c09a6cc58d1116265fd68
2023-08-17 17:59:26 +08:00
hiyouga
048f99354f fix generation
Former-commit-id: d9e62711a3349d7c6fd3512fb25c709bdfbb311a
2023-08-16 22:39:54 +08:00
hiyouga
edc15c62fa fix system prompt
Former-commit-id: 7407d9daa16bf6b3cd5002e16b2c53e402d2bc39
2023-08-16 01:35:52 +08:00
hiyouga
a9ab8f71d7 fix ChatGLM RLHF
Former-commit-id: af6c011fcb8ea9e5cf2eb4699da33d8668df04b4
2023-08-15 11:19:20 +08:00
hiyouga
d15fe288df alert pad_token source
Former-commit-id: 80b4053602c02aec724ecf980f8a279ffdf9f975
2023-08-15 00:07:56 +08:00
hiyouga
2aa2c363ad fix ChatGLM lm_head #494
Former-commit-id: d0199568081f46e5d338ea511266eb420dd43594
2023-08-14 14:14:48 +08:00
hiyouga
6c9b035c0e web UI integrating RLHF
Former-commit-id: ec94274ca155300aee27621c018dd1bbaf78194b
2023-08-14 10:48:47 +08:00
hiyouga
e75024fde3 fix #480
Former-commit-id: 2f2fd55d8175eb3c6ce94bc821ab4e6331f79d8e
2023-08-14 00:23:56 +08:00
hiyouga
e1b43dfc7f tiny fix
Former-commit-id: 9dc6a296e327c5ff27cbd1697437d9d3145e3d9a
2023-08-12 22:02:43 +08:00
hiyouga
bd611e0090 fix rope scaling
Former-commit-id: 8545c11c45906b33c78e144c2338963eaf0406b8
2023-08-12 22:00:01 +08:00
hiyouga
ba65dcb15e update readme
Former-commit-id: 1836c020c514e7a94aaa48abdf19ea8accbc1a2a
2023-08-12 21:00:11 +08:00
hiyouga
3f0a2d6adc support rope scaling, fix #475 #476 #478
Former-commit-id: fa940c17b8d3e379af08804003f1a522c1cd6ac4
2023-08-12 20:46:27 +08:00
hiyouga
7bd4c59b7e fix unusual output of 8bit models #278 #391
Former-commit-id: dd51c242032ce3f878cb191dc144536db4a2bb45
2023-08-12 00:25:29 +08:00