Skip to content

Conversation

@sryap
Copy link
Contributor

@sryap sryap commented Dec 3, 2025

Summary:
Explicitly zero grad because simply setting grad to None does not
eliminate the grad accumulation step

Reviewed By: y-sq

Differential Revision: D85325066

@sryap sryap temporarily deployed to docker-s3-upload December 3, 2025 23:06 — with GitHub Actions Inactive
@sryap sryap temporarily deployed to docker-s3-upload December 3, 2025 23:06 — with GitHub Actions Inactive
@meta-cla meta-cla bot added the cla signed label Dec 3, 2025
@meta-codesync
Copy link

meta-codesync bot commented Dec 3, 2025

@sryap has exported this pull request. If you are a Meta employee, you can view the originating Diff in D85325066.

sryap added 2 commits December 3, 2025 16:34
Summary:

Add an option to generate inputs as large as the GPU L2 cache size to
avoid an explicit cache clearing in every iteration

Reviewed By: henrylhtsang

Differential Revision: D85318459
Summary:

Explicitly zero grad because simply setting grad to None does not
eliminate the grad accumulation step

Reviewed By: y-sq

Differential Revision: D85325066
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants