Tong Li
c8b368c294
add overlength sample count ( #6332 )
...
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
2025-05-28 19:18:09 +08:00
Tong Li
de2ad3b206
fix default eval setting ( #6321 )
...
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
2025-05-22 11:52:41 +08:00
Tong Li
32afa7bf29
fix empty tensor ( #6319 )
...
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
2025-05-20 17:41:44 +08:00
YeAnbang
bcf2459db5
Merge pull request #6314 from hpcaitech/grpo-reward-dev
...
[feat] upgrade reward functions
2025-05-20 10:06:00 +08:00
YeAnbang
f8bd2db33f
add uuid to rollout log
2025-05-20 09:45:56 +08:00
YeAnbang
107470a360
fix logging rollouts
2025-05-17 21:12:58 +08:00
YeAnbang
03b41d6fb5
upgrade reward functions
2025-05-16 18:04:38 +08:00
YeAnbang
3c42c0ce82
Merge pull request #6309 from hpcaitech/grpo-eval-dev
...
[feat] Support evaluation during training
2025-05-16 16:11:23 +08:00
YeAnbang
021914c565
support logging rollouts to wandb
2025-05-16 15:56:03 +08:00
YeAnbang
203dfb1536
address conversation
2025-05-16 14:15:35 +08:00
Tong Li
ab95624915
handle empty index ( #6311 )
...
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
2025-05-16 10:00:10 +08:00
YeAnbang
6abffb9100
fix evaluation
2025-05-16 09:42:35 +08:00
Tong Li
1644adf684
handle empty index
2025-05-15 20:32:05 +08:00
YeAnbang
957e3a521a
disable wandb tb syncing
2025-05-15 16:52:31 +08:00
YeAnbang
4ec73298b2
use consumer global step
2025-05-15 14:15:40 +08:00
YeAnbang
094f119b3a
merge
2025-05-14 18:13:47 +08:00
YeAnbang
50070c1e84
move logging to producer
2025-05-14 18:10:57 +08:00
Tong Li
aca547623f
[feat] Support prompt level dynamic ( #6300 )
...
* adjust to dynamic prompt bs
* remove debug
* update pad seq (#6303 )
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
* adjust to dynamic prompt bs
* remove debug
* fix dp issue
* fix
* fix default settings
---------
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
2025-05-14 16:40:35 +08:00
YeAnbang
47a7dc7142
Support evaluation during training
2025-05-14 11:03:11 +08:00
Tong Li
b920af427b
update pad seq ( #6303 )
...
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
2025-05-13 16:51:27 +08:00
YeAnbang
eb6b5dd62e
[fix] revert reward update and evaluation ( #6295 )
...
* Revert "rewrite reward fn"
This reverts commit d06042b434a6085ff67601640ebd65bd61918dcf57a88395fe
2025-05-07 10:56:47 +08:00
YeAnbang
17928ad84f
Merge pull request #6292 from hpcaitech/grpo-latest-dev-reward-update
...
[feat] Update reward verification
2025-05-03 10:00:32 +08:00
YeAnbang
d06042b434
rewrite reward fn
2025-05-01 11:28:05 +08:00
YeAnbang
a6085ff676
upgrade reward math verification
2025-04-30 22:59:54 +08:00
YeAnbang
01640ebd65
fix bug
2025-04-30 22:53:12 +08:00
YeAnbang
bd61918dcf
reuse comm-group
2025-04-30 21:36:11 +08:00
YeAnbang
57a88395fe
Support evaluation during training
2025-04-30 18:31:49 +08:00
Tong Li
5fd4bcb9d8
[feat] Sync shard model ( #6289 )
...
* [feat] support hybrid parallel model sync
* update consumer and producer
* update files
* update producer
* remove print
* update
---------
Co-authored-by: duanjunwen <935724073@qq.com>
Co-authored-by: YeAnbang <44796419+YeAnbang@users.noreply.github.com>
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
2025-04-30 14:47:01 +08:00
YeAnbang
14f237ce7e
[feat] Support boxed math reward ( #6284 )
...
* fix pp+tp, fix dataloader
* fixed plugin micro-batch size
* support boxed reward
* add boxed reward
* fix pp state dict incomplete issue
* Revert "fix pp state dict incomplete issue"
This reverts commit 6c1b3b694f
2025-04-29 16:46:47 +08:00
YeAnbang
2ca1e3c630
fix pp+tp, fix dataloader ( #6280 )
2025-04-28 17:10:00 +08:00
Tong Li
28795f560c
fix save issue ( #6279 )
...
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
2025-04-27 17:54:06 +08:00
YeAnbang
38008858e4
fix checkpoint naming; add num_epoch parameter ( #6277 )
2025-04-26 14:00:28 +08:00
YeAnbang
26d859f68e
[feat] Support DAPO ( #6263 )
...
* update help information
* update style
* fix
* minor fix
* support PP training
* add pp support
* remove unused code
* address conversation
* fix memory leakage support tp+pp
* move empty cache
* move empty cache
* add DAPO support
* remove format reward
* fix filtering, still buggy
* small fix
* add DAPO support
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* tested multi-node training; fix bind_batch bug
* fix conversation; support sleep mode
* support reusing excessive samples
* add dynamic batching control flag
* add dynamic batching control flag
* refactored
* fix logging
---------
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
2025-04-25 17:39:17 +08:00
Tong Li
b823c6eec7
[feat] Add final save at the end ( #6274 )
...
* add final save
* default 1 episode
2025-04-23 10:03:46 +08:00
Tong Li
03f4b1dde3
add prompt template ( #6273 )
...
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
2025-04-22 10:39:47 +08:00
YeAnbang
9467c10690
[hot-fix] Fix memory leakage bug, support TP+PP ( #6258 )
...
* update help information
* update style
* fix
* minor fix
* support PP training
* add pp support
* remove unused code
* address conversation
* fix memory leakage support tp+pp
* move empty cache
* move empty cache
---------
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
2025-04-10 10:52:18 +08:00
YeAnbang
ed43a4be04
[Distributed RLHF] Integration of PP ( #6257 )
...
* update help information
* update style
* fix
* minor fix
* support PP training
* add pp support
* remove unused code
* address conversation
---------
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
2025-04-09 13:23:24 +08:00
YeAnbang
50153005b4
[feat] add microbatch forwarding ( #6251 )
...
* add microbatch forwarding
* fix forward microbatch
* fix producer OOM
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* change project name
* fix temperature annealing
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* address conversation
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
2025-03-28 10:24:58 +08:00
YeAnbang
489f215ad9
Merge pull request #6250 from hpcaitech/grpo-latest-dev
...
[feat] Fix Vllm, logprob, add filtering, temperature annealing, lr descent
2025-03-21 16:25:35 +08:00
YeAnbang
2aa7385c88
update logging
2025-03-21 16:12:07 +08:00
YeAnbang
d8eaf0d473
simplify vllm preprocessing input ids
2025-03-21 15:03:10 +08:00
YeAnbang
0472f44163
fix logprob, add filtering, temperature annealing, lr descent
2025-03-21 10:24:24 +08:00
YeAnbang
7ee4452f8c
fix vllm
2025-03-19 17:11:10 +08:00
duanjunwen
7795d4c50d
[Feature] Support Distributed LogProb for GRPO Training ( #6247 )
...
* [fix] fix qwen VocabParallelLMHead1D and gather output
* fix tp bug
* fix consumer
* [feat] Support Distributed LogProb for GRPO Training
* [fix] fix loss func
* [fix] fix log prob plugin
* [fix] fix qwen modeling param
* [fix] rm comments
* [fix] rm hard-code;fix non-dist version
* [fix] fix test file param name and benchmark tp gather output=True/False
* [fix] rm non-dist version in dist log prob
* [fix] fix comments
* [fix] fix dis log prob plugin
* [fix] fix test case
* [fix] fix qwen VocabParallelLMHead1D and gather output
* [fix] fix DistLogProb comments
* [fix] restore tp size
* [fix] fix comments
* [fix] fix comment; fix LogSoftmax usage
---------
Co-authored-by: Tong Li <tong.li35271158@gmail.com>
2025-03-18 17:47:55 +08:00
YeAnbang
bc0171d392
fix transformers backend
2025-03-14 18:12:35 +08:00
Tong Li
57b49da5e4
setup update
2025-03-13 16:52:15 +08:00
Tong Li
45ac6c6cb2
print results
2025-03-13 16:51:22 +08:00
Tong Li
4702d57841
convert to 8 generation
2025-03-13 16:49:02 +08:00
Tong Li
afddfde2dd
fix consumer
2025-03-13 14:55:26 +08:00
Tong Li
131eeceb5d
fix tp bug
2025-03-13 14:52:09 +08:00
