Skip to content
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.

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

free orig parameters after sharding #988

Merged
merged 8 commits into from
Aug 19, 2024
Merged

free orig parameters after sharding #988

merged 8 commits into from
Aug 19, 2024

Conversation

t-vi
Copy link
Collaborator

@t-vi t-vi commented Aug 18, 2024

Fixes #478

This is an alternative to #932 also covering the original #478 that leverages the Thunder meta tensor handling for the MaterializationTransform.
It builds on #986 (I need to change the target branch if that is merged.)

The test has been taken (and parametrized to cover cuda and meta) from #932 (all credit to @crcrpar ).

Thank you @kshitij12345 for the reproducer.

@t-vi t-vi requested review from mruberry and lantiga as code owners August 18, 2024 16:26
@t-vi
Copy link
Collaborator Author

t-vi commented Aug 18, 2024

This is ready for review, but I'm marking as draft for the target branch.

@t-vi t-vi marked this pull request as draft August 18, 2024 16:28
@t-vi t-vi mentioned this pull request Aug 18, 2024
4 tasks
@t-vi t-vi requested a review from crcrpar August 18, 2024 16:33
@crcrpar crcrpar requested a review from IvanYashchuk August 18, 2024 16:47
Co-authored-by: Masaki Kozuki <mkozuki@nvidia.com>
@t-vi t-vi force-pushed the tom/fsdp-composable-step2 branch from b663385 to 2357aad Compare August 19, 2024 12:21
Base automatically changed from tom/fsdp-composable-step1 to main August 19, 2024 14:00
@t-vi t-vi marked this pull request as ready for review August 19, 2024 14:35
@t-vi t-vi merged commit e8dabdd into main Aug 19, 2024
40 checks passed
@t-vi t-vi deleted the tom/fsdp-composable-step2 branch August 19, 2024 14:39
@t-vi
Copy link
Collaborator Author

t-vi commented Aug 19, 2024

Thank you @crcrpar @IvanYashchuk

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.

fsdp(jit(...)) transform can use more memory compared to jit(fsdp(...))
3 participants