Skip to content

Detach hidden states to avoid gradient carry #345

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

Merged
merged 1 commit into from
Nov 4, 2025
Merged

Detach hidden states to avoid gradient carry #345

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

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Detach hidden states to avoid gradient carry 
Detach old and reference hidden states to prevent gradient flow across micro-batches.
  • Loading branch information
@pluesclues
pluesclues authored Nov 4, 2025
commit d5ff9be234ef20bf982a57d9c41fb434122d0ee9
8 changes: 8 additions & 0 deletions unsloth_zoo/rl_replacements.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -626,6 +626,14 @@ def grpo_accumulated_loss(
image_sizes = image_sizes,
logits_to_keep = logits_to_keep + 1,
).logits

# Ensure old/ref states never carry grads across micro-batches
if old_hidden_states is not None and torch.is_tensor(old_hidden_states):
old_hidden_states = old_hidden_states.detach()

if ref_hidden_states is not None and torch.is_tensor(ref_hidden_states):
ref_hidden_states = ref_hidden_states.detach()

loss, completion_length, mean_kl, delta, flat_is_ratio = UnslothEfficientGRPO.apply(
new_hidden_states,
old_hidden_states,
Expand Down