Wensun/apo #96
Open
Wensun/apo #96
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
wensun
requested review from
bcui-db,
dakinggg,
gupta-abhay,
abaheti95 and
jdchang1
as code owners
bcui-db
reviewed
Jun 23, 2025
bcui-db
reviewed
Jun 23, 2025
Comment on lines
958
to
1004
| mean_ift = masked_mean( | ||
| env_outs['ift_kl'], | ||
| env_outs['action_mask'], | ||
| ) | ||
| env_outs['advantages'] = advantages | ||
| self.kl_ift.append(mean_ift.cpu()) | ||
|
|
||
| iter_batch.update(env_outs) | ||
|
|
||
| iter_batch.update({ | ||
| 'max_gen_len': | ||
| torch.ones(self.iter_batch_size).to(torch.int32) * | ||
| self.max_gen_len, | ||
| 'adv_masked_mean': | ||
| torch.ones(self.iter_batch_size) * batch_adv_mean.cpu(), | ||
| 'adv_masked_var': | ||
| torch.ones(self.iter_batch_size) * batch_adv_var.cpu(), | ||
| 'ift_kl_scalar': | ||
| torch.ones(self.iter_batch_size) * self.kl_ctl.value, | ||
| 'reward_std': | ||
| torch.ones(self.iter_batch_size) * | ||
| env_outs['rewards'].std().to('cpu'), | ||
| }) | ||
| else: | ||
| raise ValueError( | ||
| f'Invalid loss type: {self.actor_critic.loss_type}. ' + | ||
| 'Valid options are: ppo, grpo.', | ||
| ) | ||
| # APO and REBEL | ||
|
|
||
| batch_adv_mean, batch_adv_var = dist_compute_masked_mean_and_var( | ||
| env_outs['advantages'], | ||
| env_outs['action_mask'], | ||
| ) | ||
| mean_ift = masked_mean( | ||
| env_outs['ift_kl'], | ||
| env_outs['action_mask'], | ||
| ) | ||
| self.kl_ift.append(mean_ift.cpu()) | ||
|
|
||
| iter_batch.update(env_outs) | ||
|
|
||
| iter_batch.update({ | ||
| 'max_gen_len': | ||
| torch.ones(self.iter_batch_size).to(torch.int32) * | ||
| self.max_gen_len, | ||
| 'adv_masked_mean': | ||
| torch.ones(self.iter_batch_size), | ||
| 'adv_masked_var': | ||
| torch.ones(self.iter_batch_size), | ||
| 'ift_kl_scalar': | ||
| torch.ones(self.iter_batch_size) * self.kl_ctl.value, | ||
| 'reward_std': | ||
| torch.ones(self.iter_batch_size) * | ||
| env_outs['rewards'].std().to('cpu'), | ||
| }) |
There was a problem hiding this comment.
Isn't this block of code for both algorithms very similar to each other, except for the adv_masked_mean bit? If so can we condense it?
Co-authored-by: bcui-db <141345999+bcui-db@users.noreply.github.com>
…into wensun/apo
jdchang1
reviewed
Jun 23, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Implementation of the new algorithm A*PO.
An example of successful run on GSM8K on mlflow: link