Skip to content

Fix MEGNetConv state propagation - #734

Open
DHAKED440 wants to merge 2 commits into
JuliaGraphs:masterfrom
DHAKED440:fix/megnetconv-state-propagation
Open

DHAKED440 wants to merge 2 commits into
JuliaGraphs:masterfrom
DHAKED440:fix/megnetconv-state-propagation

Conversation

@DHAKED440

@DHAKED440 DHAKED440 commented Nov 8, 2026 •

Copy link
Copy Markdown

Fixes issue #732.

Problem

MEGNetConv wrapped its edge and node update layers in StatefulLuxLayer, but returned the original state unchanged. Updates from stateful sublayers such as BatchNorm were therefore lost between calls.

Solution

Return the updated states from both the edge update layer (ϕe) and node update layer (ϕv) after the MEGNet forward pass.

Changes Made

  • Updated the MEGNetConv forward method to return both sublayers’ updated states.
  • Added a regression test using BatchNorm in ϕe, including reuse of the returned state.
  • Updated the GNNLux development version and changelog.

Testing

  • Reproduced the issue before the fix.
  • Ran the MEGNetConv convolution test item.
  • Ran the GNNLux CPU test suite.

Test results: The focused runner completed with 227 passes and 2 broken/skipped tests. The full GNNLux CPU runner completed with 183 passes and 2 broken/skipped tests; both exited with code 0. The runs logged sandbox-related Revise path permission errors. Pkg.test() was blocked by a registry archive permission error, so I ran the package test runner directly.

This branch has not been deployed

No deployments
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.

1 participant