Skip to content

feat: RGS for wanda++ #2537

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

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Conversation

namgyu-youn
Copy link
Contributor

@namgyu-youn namgyu-youn commented Jul 13, 2025

This update introduces one of the main concepts of Wanda++ (arXiv), Regional Gradient Score (RGS).

Algorithm : RGS

RGS uses block-level gradients instead of magnitude (Wanda):

$$ RGS \ : \ S_ij = (α * G_ij + ||X_j||_2) * |W_ij| $$

  • G_ij : Regional gradient
  • ||X_j||_2 : Activation norm (same in Wanda)
  • |W_ij| : Weight magnitude (same in Wanda)
  • α : Regional gradient weight (default: 100 from paper)

Algorithm : Wanda++

  • Collect Calibration : Save input samples in block-level
  • Compute Regional Gradient Score (RGS)
  • Apply sparsity (pruning)

Copy link

pytorch-bot bot commented Jul 13, 2025

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/2537

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@facebook-github-bot facebook-github-bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Jul 13, 2025
@namgyu-youn
Copy link
Contributor Author

@pytorchbot label "sparsity"

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. sparsity
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants