Skip to content

Fix FSDP CPU offloading when shards are already on CPU#21805

Open
zhixiangli wants to merge 1 commit into
Lightning-AI:masterfrom
zhixiangli:fix-fsdp-offload-to-cpu
Open

Fix FSDP CPU offloading when shards are already on CPU#21805
zhixiangli wants to merge 1 commit into
Lightning-AI:masterfrom
zhixiangli:fix-fsdp-offload-to-cpu

Conversation

@zhixiangli

Copy link
Copy Markdown
Contributor

What does this PR do?

Fixes #21804

Offloading to CPU when FSDP is on CPU triggers a use-after-free in PyTorch's FlatParamHandle.to_cpu(). This change checks if the parameters are on CPU and disables offload_to_cpu if they are.

Before submitting
  • Was this discussed/agreed via a GitHub issue? (not for typos and docs)
  • Did you read the contributor guideline, Pull Request section?
  • Did you make sure your PR does only one thing, instead of bundling different changes together?
  • Did you make sure to update the documentation with your changes? (if necessary)
  • Did you write any new necessary tests? (not for typos and docs)
  • Did you verify new and existing tests pass locally with your changes?
  • Did you list all the breaking changes introduced by this pull request?
  • Did you update the CHANGELOG? (not for typos, docs, test updates, or minor internal changes/refactors)

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing, make sure you have read the review guidelines. In short, see the following bullet-list:

Reviewer checklist
  • Is this pull request ready for review? (if not, please submit in draft mode)
  • Check that all items from Before submitting are resolved
  • Make sure the title is self-explanatory and the description concisely explains the PR
  • Add labels and milestones (and optionally projects) to the PR so it can be classified

Offloading to CPU when FSDP is on CPU triggers a use-after-free in PyTorch's FlatParamHandle.to_cpu().
This change checks if the parameters are on CPU and disables `offload_to_cpu` if they are.

TAG=agy
CONV=7d4c3ada-86b4-475b-9e1a-ce254b554825
@zhixiangli zhixiangli marked this pull request as ready for review July 4, 2026 15:23
@codecov-commenter

codecov-commenter commented Jul 4, 2026

Copy link
Copy Markdown

⚠️ Please install the 'codecov app svg image' to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87%. Comparing base (4819088) to head (e00c8ea).
❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files
@@           Coverage Diff           @@
##           master   #21805   +/-   ##
=======================================
  Coverage      87%      87%           
=======================================
  Files         270      270           
  Lines       24010    24012    +2     
=======================================
+ Hits        20787    20792    +5     
+ Misses       3223     3220    -3     

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

SIGSEGV (Segmentation Fault) when saving checkpoints with FSDP on CPU

2 participants