Skip to content

fix: prevent 5 GB+ CUDA memory leak in activation offloading by syncing streams and clear stashes in OffloadActivations.__exit__ #5700

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
kashif merged 2 commits into from
May 7, 2026
+17 −0
Merged

fix: prevent 5 GB+ CUDA memory leak in activation offloading by syncing streams and clear stashes in OffloadActivations.__exit__ #5700

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
320bd85
fix: prevent CUDA stream leaks in OffloadActivations.__exit__
butterwecksolutions May 7, 2026
f1377a4
Apply suggestion from @kashif
kashif May 7, 2026
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
17 changes: 17 additions & 0 deletions trl/models/activation_offloading.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -568,6 +568,23 @@ def update_model_params(self, model: nn.Module):

self.param_storages = param_storages

def __exit__(self, *args, **kwargs):
"""Sync streams and clear stashes before parent cleanup.

try/finally ensures the saved_tensors_hooks parent cleanup runs even if stream sync raises — otherwise hooks
stay permanently installed, creating a silent memory leak.
"""
try:
if self.use_streams:
self.s0.synchronize()
self.s1.synchronize()
self.bwd_tensor_stash.clear()
self.bwd_ev_stash.clear()
self.fwd_stash.clear()
finally:
result = super().__exit__(*args, **kwargs)
return result

Comment on lines +571 to +587

@kashif kashif May 6, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
def __exit__(self, *args, **kwargs):
"""Sync streams and clear stashes before parent cleanup.
Prevents leaked CUDA blocks with garbage stream IDs when the
context manager exits with in-flight async copies or stashed
tensors still referencing the offload stream.
NOTE: tracker is NOT cleared — the backward pass unpacks tensors
via tracker AFTER __exit__ (see class docstring example).
"""
import torch
# 1. Sync both streams so async copies finish before cleanup
if self.use_streams and self.s0 is not None:
self.s0.synchronize()
if self.use_streams and self.s1 is not None:
self.s1.synchronize()
# 2. Clear stashed tensors (only exist when use_streams=True)
if self.use_streams:
self.bwd_tensor_stash.clear()
self.bwd_ev_stash.clear()
self.fwd_stash.clear()
return super().__exit__(*args, **kwargs)
def __exit__(self, *args, **kwargs):
"""Sync streams and clear stashes before parent cleanup."""
try:
if self.use_streams:
self.s0.synchronize()
self.s1.synchronize()
self.bwd_tensor_stash.clear()
self.bwd_ev_stash.clear()
self.fwd_stash.clear()
finally:
result = super().__exit__(*args, **kwargs)
return result

@kashif kashif May 6, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Lets make the parent __exit__ run in a finally so the saved tensor hooks are always popped. I’d also drop the is not None checks unless there is a real code path where self.s0 or self.s1 can be None. In this class, self.s0 is always initialized, and self.s1 exists whenever use_streams=True.


class NoOpManager(saved_tensors_hooks):
"""
Expand Down
Loading