Skip to content

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

Merged
kashif merged 2 commits into
huggingface:mainfrom
butterwecksolutions:stream-cleanup
May 7, 2026
Merged

fix: prevent 5 GB+ CUDA memory leak in activation offloading by syncing streams and clear stashes in OffloadActivations.__exit__#5700
kashif merged 2 commits into
huggingface:mainfrom
butterwecksolutions:stream-cleanup

Commits

Commits on May 7, 2026