github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-10-29 21:03:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix bugs

Browse Source
This commit is contained in:
YeAnbang
2025-06-20 15:44:13 +08:00
parent ff6696a9bb
commit c2561f826a
6 changed files with 90 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
applications/ColossalChat/rl_example.py
View File

@@ -263,6 +263,7 @@ if __name__ == "__main__":
grpo_config = {
"lr": args.learning_rate,
"train_microbatch_size": args.train_microbatch_size,
"num_minibatch_during_rollout": 1, # number of mini batches to pop out from buffer and used for training during rollout of the producer after it syncs the model. Hint, set to a proper value close to the number of mini batches for training that takes roughly the same time as the rollout of the producer. A value that is too large or too small will cause bubble time on the trainer or the producer.
"beta": args.kl_coeff, # KL penalty coefficient
"loss_variation": "sample_level",
"reward_fn_type": args.reward_type,