Skip to content

fix: weight PPO token-mean micro-batches by valid tokens #292

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
taking-lying-flat wants to merge 1 commit into
base: main
Choose a base branch
from
+10 −0
Open

fix: weight PPO token-mean micro-batches by valid tokens #292

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 10 additions & 0 deletions src/twinkle/loss/grpo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -384,6 +384,16 @@ def __init__(self, loss_agg_mode: Literal['token-mean', 'seq-mean-token-mean'] =
f'Expected one of {sorted(self._LOSS_AGG_MODES)}.')
self.loss_agg_mode = loss_agg_mode

def micro_batch_scale(self, inputs, indices):
if self.loss_agg_mode != 'token-mean':
return super().micro_batch_scale(inputs, indices)
import torch
token_counts = [
self._resolve_loss_mask(model_input, torch.atleast_2d(torch.as_tensor(model_input['labels']))).sum().item()
for model_input in inputs
]
return sum(token_counts[index] for index in indices) / max(sum(token_counts), 1)

def _aggregate_loss(
self,
per_token_loss: 'torch.Tensor',
Expand Down