Skip to content

fix(trainer): add torch.cuda.empty_cache() after FSDP update_actor #541

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
pan-x-c merged 1 commit into from
May 19, 2026
Merged

fix(trainer): add torch.cuda.empty_cache() after FSDP update_actor #541

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
6 changes: 6 additions & 0 deletions trinity/trainer/verl/fsdp_workers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -958,6 +958,12 @@ def update_actor(self, data: DataProto):
"After offload actor optimizer during update_actor", logger=self.logger
)

# Release reserved GPU memory held by PyTorch's caching allocator after
# backward passes. Without this, memory_reserved grows monotonically and
# eventually starves vLLM during weight sync in colocate mode.
# Matches the pattern in megatron_workers.py update_actor().
torch.cuda.empty_cache()

return output

@register(dispatch_mode=make_nd_compute_dataproto_dispatch_fn(mesh_name="actor"))
Expand Down
Loading